feat(C153): catálogo de falas do Solla — schema, access, import idempotente e busca

Collections speech/speechSegment no grupo Comunicação, papel communicator (leitura restrita) e índice trigram normalizado para a busca por segmento. Import idempotente da Câmara (discurso→evento→trecho→VOD→ASR) com facetas gazetteer+LLM, proveniência e preservação de curadoria manual; parser de quem presidia; reusa a lógica pura do piloto C152 via scripts/lib. Related #955
2026-11-13 15:13:07 +00:00 · 2026-09-12 21:39:11 -03:00 · 2026-09-12 21:39:11 -03:00 · 6c0261fd0f
commit 6c0261fd0f
parent fa6933c1cd
30 changed files with 18709 additions and 111 deletions
--- a/docs/plans/catalogo-falas-solla-impl.md
+++ b/docs/plans/catalogo-falas-solla-impl.md
@ -0,0 +1,190 @@
+# Impl: Catálogo de falas do Solla: import idempotente e busca textual
+
+Status: em execução
+Atualizado em: 2026-09-12
+Issue: #955
+Intenção: docs/plans/catalogo-falas-solla.md
+Appetite restante: herdado (~2–3 dias; a Fase 4 é wall-clock de máquina, sem trabalho humano)
+
+## Leitura da intenção
+
+- **Outcome:** o acervo da 57ª legislatura existe no banco local com transcrição oficial, segmentos ASR com minutagem, facetas (tema/alcance) com proveniência e menções; o import é reexecutável sem duplicar e imprime relatório por execução; a busca por palavra sobre os segmentos acha com e sem acento. C155 leva a produção.
+- **O que NÃO negociar:**
+  - Facetas com proveniência `gazetteer | llm | manual`; keyword oficial crua preservada (nunca substituída pela faceta).
+  - Leitura restrita a `communicator`, `coordinator` e `candidate` (+ Payload admin). `advisor` e `leader` não leem — nem via `/api`.
+  - Sem UI (C154), sem espelho de mídia/S3, sem normalizar pessoas citadas como `Contact`.
+  - Identidade do registro não é a URL do MP4 (regerável, hash não é contrato); import escreve só no DB local (C155 faz produção).
+- **O que reavaliar (hipóteses da intenção):**
+  - "LLM do Sollinha via `DEEPSEEK_API_KEY`" — essa chave não existe no ambiente do import; o mesmo modelo (`deepseek-ai/DeepSeek-V4-Flash`) está verificado ao vivo na Deep Infra, provedor/chave que a ASR já usa. Reavaliado na D3.
+  - "chave natural por áudio+trecho" (C152) sozinha não cobre discurso sem trecho (54ª/sessões sem âncora). Ela vira **identidade de vídeo** (decide pular VOD/ASR), não chave do registro. Ver D2.
+  - "Direção no codebase: script no padrão `recover-media.mjs`" — o padrão mais próximo é `seed-tse-results.mjs` (importa `src/` via tsx + `seed-loader`, guard de banco local, `assertLocalDatabase`). É o que este plano usa.
+
+## Abordagem recomendada
+
+```mermaid
+flowchart LR
+  subgraph Fonte Câmara
+    A[API discursos<br/>dadosabertos] --> B[selectSpeechEvents]
+    B --> C[event page cache<br/>data/camara/events]
+    C --> D[matchExcerpt]
+    C --> P[parsePresidingTransitions]
+  end
+  D --> E{VOD PRONTO?}
+  E -- sim --> F[MP4 cache<br/>data/camara]
+  F --> G[Deep Infra Whisper<br/>verbose_json segments]
+  E -- INDISPONIVEL / sem trecho --> R[registra sem segmentos]
+  A --> H[classifySpeech<br/>gazetteer + LLM]
+  G --> I[upsertSpeechBundle<br/>transação por sourceKey]
+  H --> I
+  P --> I
+  R --> I
+  I --> J[(speech + speechSegment)]
+  J --> K[searchText normalizado<br/>+ GIN pg_trgm]
+```
+
+**Opções consideradas:** A | B | C por decisão (D1–D6 abaixo).
+**Recomendação:** as seis decisões, na ordem em que bloqueiam as demais; D1 e D2 travam o schema, D3–D5 travam o import e o access, D6 é dado de saída.
+**Rejeitadas:** registradas dentro de cada decisão (formato obrigatório da decision-quality).
+
+### Decisões de engenharia
+
+#### D1 — Segmentos em collection própria, não array do discurso
+
+- **Opções:** A) collection `speechSegment` (FK `speech`, `order`, `startSeconds`, `endSeconds`, `text`, `searchText`) | B) array de segmentos dentro de `speech` | C) um blob JSON em `speech`.
+- **Recomendação:** A — a busca global (`contains` + GIN trgm) roda sobre uma tabela própria, paginável e `select`-enxuta; o C154 busca segmento e abre o discurso sem carregar 40–100 segmentos por doc. `speechSegment` fica admin-hidden (precedente `SupporterImportBatch.ts:26`) porque a superfície é o C154.
+- **Rejeitadas:** B porque a query textual sobre array relacional do Payload não expõe `contains` por subcampo com o shape que o C154 precisa e infla todo read de discurso; C porque mata filtro por segmento e a indexação trigram.
+
+#### D2 — `sourceKey` da identidade da API + identidade de vídeo armazenada
+
+- **Opções:** A) `sourceKey = dataHoraInicio|tipoDiscurso|fase` (unique) e `eventId+audioId+excerptTMs` armazenados só para decidir re-transcrição | B) unique em `eventId+audioId+excerptTMs` | C) URL do MP4 como identidade.
+- **Recomendação:** A — cobre discurso sem trecho (senão ele nunca teria chave) e sobrevive à regeneração do VOD; o import acha por `sourceKey`, e se `audioId+excerptTMs` baterem com o que já está no banco **e houver segmentos**, pula VOD+ASR e só atualiza metadados. `--skip-transcribe`/ASR falho passam `segments: undefined` (preserva os antigos, não zera).
+- **Rejeitadas:** B porque discurso sem trecho ficaria de fora ou duplicaria a cada execução; C porque o `<hash>.mp4` é regerável e a permanência só foi medida em janela curta (C152).
+
+#### D3 — Classificador LLM via Deep Infra raw fetch (mesmo provedor da ASR)
+
+- **Opções:** A) raw fetch OpenAI-compat na Deep Infra, `model: deepseek-ai/DeepSeek-V4-Flash`, `response_format: json_object`, zod + fallback gazetteer | B) `@ai-sdk/deepseek` contra `api.deepseek.com` | C) só gazetteer.
+- **Recomendação:** A — entrega a opção A da intenção (DeepSeek flash + validação + fallback C) sem segundo provedor nem segredo novo: a `DEEPINFRA_API_KEY` já existe no env e é a mesma da transcrição; custo medido US$ 0,00007/discurso (234 ≈ US$ 0,016). Saída validada por zod; HTTP erro/timeout/parse inválido → `classifiedBy='gazetteer'` com topics/municípios do lexicon (scopes vazios se o lexicon não achar termo). `classifiedBy='manual'` nunca é sobrescrito (D2/hook abaixo).
+- **Rejeitadas:** B porque exige `DEEPSEEK_API_KEY` (ausente), acopla um SDK ao CLI para uma chamada única e cria dois caminhos de LLM no repo; C porque perde `scopes` (Bahia/Brasil/Internacional) e pessoas/programas, que são aceite de produto.
+
+#### D4 — Busca: `searchText` normalizado + GIN `pg_trgm`
+
+- **Opções:** A) coluna `searchText` normalizada app-side (NFD sem acento + lowercase) + índice GIN `gin_trgm_ops` hand-written | B) extensão `unaccent` + `ILIKE unaccent(text)` | C) `tsvector` com dicionário português.
+- **Recomendação:** A — `contains` funciona igual em dev/test/prod sem extensão nova além do `pg_trgm` já instalado (precedente `20260719_020000_add_contact_trgm_index.ts`); a normalização é pura, unit-testada em `src/lib/speechSearch.ts`, e o C154 importa a MESMA função para montar a query. Hook `beforeValidate` em `speechSegment` deriva `searchText` de `text` (criação e edição manual).
+- **Rejeitadas:** B porque `unaccent(text)` não é IMMUTABLE para índice funcional sem wrapper, depende de extension adicional e collation vira pegadinha; C porque `tsvector` faz stemming (piora prefixo/substring), exige config de dicionário e a semântica pedida é substring com/sem acento, não ranking lexical.
+
+#### D5 — Papel `communicator` com predicado próprio, fora de `isStaffCampaignRole`
+
+- **Opções:** A) novo valor no enum `campaignUser.role` + `canReadSpeechCatalog(role)` em `src/lib/campaignRoles.ts` + módulo `src/utilities/access/speeches.ts` reexportado por `campaignAccess.ts` | B) reusar `advisor`/staff | C) criar o leitor no `users` (admin Payload).
+- **Recomendação:** A — o assessor de comunicação lê o catálogo sem herdar nenhuma área de staff: `isStaffCampaignRole` (`src/lib/campaignRoles.ts:13`) fica intocado. Leitura: Payload admin **ou** campaignUser `communicator|coordinator|candidate`. Update (correção manual de facetas): `coordinator|candidate|admin`. Create/delete: `payloadAdminOnly`. Import CLI escreve com `overrideAccess: true` (comentário "bypass", padrão seed-tse). Guardas mínimas de shell para o papel não vazar staff: entrada em `campaignRoleLabels` (`campaignUserProfile.ts:9` — o `Record` quebra o typecheck sem ela), branch explícito em `getCampaignNav` (`nav.ts:85`, hoje devolveria `staffNav` a qualquer não-leader; `communicator` → nav vazia até o C154), e nenhuma mudança em `homeActionsForRole` (já devolve `[]` para não-staff, verificado em `campaignHomeActions.ts:124`).
+- **Rejeitadas:** B porque abriria o /campanha inteiro (viola "leitura restrita"); C porque o perfil de comunicação não é editor do CMS e ganharia o `/admin` todo.
+
+#### D6 — "Quem presidia": heurística "Troca da mesa" na página do evento
+
+- **Opções:** A) parse dos headings `g-l-assista__categoria-outros-videos` "Troca da mesa Presidente X por Participante Y" (transição = maior `t` do grupo; presidente em T = Y da transição mais recente ≤ T; antes da 1ª = X da mais antiga; sem transições = null) | B) deixar `presidingOfficer` sempre null | C) tirar do Diário (`urlTexto`) ou de outro endpoint.
+- **Recomendação:** A — é a única fonte observada com o dado (API não tem campo); o parser entra no `scripts/lib/camaraSpeeches.mjs` (dono da leitura da página, reusado pelo import) como `parsePresidingOfficerTransitions`/`resolvePresidingOfficer`, com fixture de HTML e unit tests; layout mudou → parse devolve `[]`/null e o import nunca quebra.
+- **Rejeitadas:** B porque joga fora dado pedido no aceite e já disponível; C porque `urlTexto` é a taquigrafia (não registra a mesa) e não há endpoint com o campo.
+
+### Componentes / mudanças
+
+- **`src/collections/Speech.ts`** (novo): slug `speech`, labels "Discurso"/"Discursos", `admin.group: 'Comunicação'`, `useAsTitle: 'speechAt'`, `admin.description` com o crédito CC BY 4.0 da Câmara. Campos: `sourceKey` (text, unique, index), `speechAt` (date, index), `year` (number, index; derivado por hook), `legislature` (select 54|55|56|57, index), `type`, `phase`, `durationSeconds` (number), `summary` (textarea), `officialTranscript` (textarea), `officialTextUrl`, `keywords` (text hasMany, cru), `eventId` (number, index), `eventType`, `eventStartAt`/`eventEndAt` (date), `youtubeUrl`, `audioId` (number, index), `excerptTMs` (number, index), `vodPlaybackUrl`/`vodDownloadUrl` (último conhecido, não identidade), `presidingOfficer`, `topics` (select hasMany, index, 18 valores de `SPEECH_TOPICS`), `scopes` (select hasMany, index, 3), `classifiedBy` (select `gazetteer|llm|manual`, default `gazetteer`), `mentionedMunicipalities` (relationship `municipality` hasMany), `mentionedPeople`/`mentionedPrograms`/`mentionedProjects` (text hasMany). Hooks: `beforeValidate` deriva `year`; `beforeChange` preserva facetas quando `originalDoc.classifiedBy === 'manual'` e a escrita é sem `req.user` (import/CLI) — edição autenticada no admin continua podendo corrigir e manter `manual`; `beforeDelete` apaga os segmentos do discurso (relacionamento Payload não tem FK). Access: `create/delete: payloadAdminOnly`, `read: canReadSpeech`, `update: canUpdateSpeech`.
+- **`src/collections/SpeechSegment.ts`** (novo): slug `speechSegment`, labels "Segmento de fala", `admin.hidden: () => true` (precedente `SupporterImportBatch.ts:26`). Campos: `speech` (relationship, required, index), `order` (number, required), `startSeconds`/`endSeconds` (number, required), `text` (textarea, required), `searchText` (text, required, `admin.readOnly`, sem `index` — o GIN é hand-written). Hook `beforeValidate`: `searchText = normalizeForSearch(data.text ?? originalDoc?.text)`. Access espelha `speech` (read `canReadSpeech`; mutações `payloadAdminOnly`).
+- **`src/payload.config.ts`**: registrar `Speech` e `SpeechSegment` no array `collections` (imports + entradas; grupo novo "Comunicação" aparece no admin).
+- **`src/lib/speechFacets.ts`** (novo, puro/client-safe): contrato `SPEECH_TOPICS` (18 slugs pt — valores de dado, precedente `post.type`; nomes de código em inglês), `SPEECH_SCOPES`, `SPEECH_CLASSIFICATION_SOURCES`; lexicon gazetteer tema→termos; `matchTopicFacets(text)` com boundary de token (teste: "sus" NÃO casa "suspeito"); `matchMunicipalityMentions(text)` contra `municipalityCatalog` com consumo guloso do nome mais longo (testes: Conde vs Condeúba; São Félix vs São Félix do Coribe) — **menção a "Salvador" mapeia para as 19 zonas do catálogo** (a cidade não tem linha própria; a menção à cidade vale para qualquer zona, e o C154 filtra por município; documentar no campo/admin description). `parseLlmFacetResponse(json)` com zod (filtra para valores conhecidos, dedupe normalizado, caps de pessoas/programas/projetos); `mergeFacetClassification(gazetteer, llm)` (topics = união; scopes = do LLM validado ou do lexicon; `classifiedBy = llm ? 'llm' : 'gazetteer'`).
+- **`src/lib/speechSearch.ts`** (novo, puro/client-safe): `normalizeForSearch(value)` (NFD, remove combining marks, lowercase, colapsa espaços) e `uniqueByNormalizedForm(values)`; importado pelo hook do segmento, pelo gazetteer e pelo C154.
+- **`src/utilities/speech/speechImport.ts`** (novo, `server-only`): `findSpeechImportState(payload, sourceKey)` (leitura para o skip de VOD/ASR) e `upsertSpeechBundle(payload, bundle)` — reusa `withPayloadTransaction` (`src/utilities/payloadTransaction.ts:86`), acha por `sourceKey`, cria/atualiza, pula campos de faceta se existente é `manual`, substitui segmentos **somente quando `bundle.segments !== undefined`**, tudo com `req: { transactionID }` e `overrideAccess: true` com comentário "bypass". Retorna counts (created/updated/segmentsInserted/segmentsDeleted/manualPreserved) para o relatório.
+- **`src/utilities/speech/speechClassifier.ts`** (novo, `server-only`): `classifySpeech(input)` com `DEEPINFRA_API_KEY`; raw fetch OpenAI-compat (`https://api.deepinfra.com/v1/openai/chat/completions`, `deepseek-ai/DeepSeek-V4-Flash`, `response_format: json_object`, `temperature: 0`, timeout), transcript truncado (≈12k chars) + keywords cruas; valida com `parseLlmFacetResponse`; qualquer falha/chave ausente → gazetteer-only. Nunca lança.
+- **`src/utilities/access/speeches.ts`** (novo): `canReadSpeech` (admin ∪ campaignUser `communicator|coordinator|candidate`), `canUpdateSpeech` (admin ∪ coordinator/candidate); reexport no barrel `src/utilities/campaignAccess.ts` (prefixo `src/utilities/access` já mapeado no E2E manifest — sem mudança de manifesto).
+- **`src/lib/campaignRoles.ts`**: `canReadSpeechCatalog(role)` client-safe (o predicado próprio, fora de `isStaffCampaignRole`).
+- **`src/collections/CampaignUser.ts`**: opção `{ label: 'Assessor de Comunicação', value: 'communicator' }` no select de role (linhas 483–488).
+- **`src/utilities/campaignUserProfile.ts`**: `communicator: 'Assessor de Comunicação'` em `campaignRoleLabels` (obrigatório para o typecheck pós-`generate:types`).
+- **`src/components/campaign/shell/nav.ts`**: branch `if (role === 'communicator') return []` em `getCampaignNav` (nav vazia; C154 adiciona "Falas"). `getCampaignSecondaryNav`/`getCampaignBottomNav` já devolvem `[]` via `isStaffCampaignRole`.
+- **`scripts/lib/camaraSpeeches.mjs`** (editar o dono): adicionar `parsePresidingOfficerTransitions(html)` e `resolvePresidingOfficer(transitions, tMs)` + export; reusar `parseEventExcerpts`, `buildVodUrl`, `parseVodStatus`, `selectSpeechEvents`, `matchExcerpt`, `normalizeTranscription`, `LEGISLATURE_RANGES`, `SOLLA_DEPUTY_ID`, `CAMARA_USER_AGENT`.
+- **`scripts/import-camara-speeches.mjs`** (novo): `dieWithLabel('camara:import')`, `loadCliEnv`, `ensureCachedDownload`/`writeRepoFile` de `scripts/lib/cli.mjs`, `assertLocalDatabase`; flags `--legislature` (default 57) `--date` `--limit` `--skip-transcribe` `--reclassify` `--out` `--help`; cache de lista de eventos por dia e de página de evento em `data/camara/events/`, MP4 em `data/camara/` (já gitignored); prioriza `Sessão Deliberativa`/`Breves Comunicações` e capa a varredura de eventos por discurso (C152: ~50s/dia quando não há trecho); retry com backoff e delay educado; importa `speechImport`/`speechClassifier`/config via dynamic import literal (knip resolve); relatório no stdout + JSON em `data/camara/reports/<run>.json`; falha por discurso entra em `failures[]` e o run continua (resumível); exit 1 só em fatal (DB/guard).
+- **`package.json`**: `"camara:import": "cross-env NODE_OPTIONS=\"--no-deprecation --import=tsx/esm --import=./scripts/seed-loader.mjs\" node scripts/import-camara-speeches.mjs"` (padrão `db:seed:tse`; `seed-loader` cobre `server-only` e `next/cache`).
+- **Migrations:** `pnpm migrate:create add_speech_catalog` (rebase antes, conforme engineering-standards) → se o gerador não emitir, adicionar à mão `ALTER TYPE "public"."enum_campaign_user_role" ADD VALUE IF NOT EXISTS 'communicator'` (precedente `20260723_200000_remodel_municipalities.ts:105`; PG 12+ aceita dentro da transação e nada usa o valor nela; `down` irreversível documentado). Depois, hand-written `src/migrations/<ts>_add_speech_segment_trgm_index.ts` (`CREATE EXTENSION IF NOT EXISTS pg_trgm; CREATE INDEX IF NOT EXISTS "speech_segment_search_text_trgm_idx" ON "speech_segment" USING gin ("search_text" gin_trgm_ops);`, `down` dropa só o índice) registrado em `src/migrations/index.ts` como o `20260719_020000_add_contact_trgm_index`.
+- **`pnpm generate:types`**: `payload-types.ts` ganha `Speech`/`SpeechSegment` e `'communicator'` no union de role.
+- **Testes:** `tests/unit/speechFacets.unit.spec.ts`, `tests/unit/speechSearch.unit.spec.ts`, extensão de `tests/unit/camaraSpeeches.unit.spec.ts` (presiding + HTML fixture), `tests/int/speechCatalog.int.spec.ts` (access allow/deny + hook `searchText` + `contains`), `tests/int/speechImport.int.spec.ts` (idempotência, replace de segmentos, manual preservado, `segments: undefined` preserva, estado de skip).
+- **Changelog:** `docs/changelog/2026-09-12-c153.md`.
+
+### Dados → forma (se aplicável)
+
+- **Sem UI** (N/A por decisão de produto). A "forma" deste item é o **relatório de execução** (stdout + JSON): `runAt`, `range`, `totals { listed, processed, created, updated, skippedExisting, withVideo, withSegments, withoutExcerpt, failed }`, `asr { calls, audioSeconds, estimatedCostUsd }` (US$ 0,00045/min), `llm { calls, totalTokens, estimatedCostUsd }`, `elapsedMs`, `failures[] { sourceKey, speechAt, stage: 'event'|'excerpt'|'vod'|'asr'|'classify'|'upsert', message }`.
+- **Rejeitadas:** dashboard/admin como superfície (é o C154; aqui só dados), CSV de saída (o JSON do relatório já é legível por máquina), relatório por discurso no banco (ruído; o relatório é artefato de execução, não entidade).
+
+## Fases verificáveis
+
+1. **Tracer / schema+access+migration** (≈0,5 dia): collections registradas, papel `communicator` (option + label + predicado + nav), `access/speeches.ts` + reexport, migrations (catálogo + enum + trgm), `generate:types`. **Prova vertical (tracer cedo):** `tests/int/speechCatalog.int.spec.ts` cria 1 `speech` + 1 `speechSegment` via Local API, lê por `searchText contains 'saude'` com texto acentuado, e prova allow/deny por papel (communicator/coordinator/candidate/admin leem; advisor/leader/anônimo não; mutação só admin/coordinator/candidate). Aplicar migrations no DB de dev e no `_test` (skill `local-database`) antes do int.
+2. **Módulos puros + unit** (≈0,5 dia): `speechFacets.ts`, `speechSearch.ts`, `canReadSpeechCatalog`; extensão do `camaraSpeeches.mjs` (presiding). Unit verde: boundary de tema, Conde/Condeúba, Salvador → 19 zonas, merge gazetteer+llm, zod rejeitando saída fora da taxonomia, normalização de busca, heurística da mesa com HTML congelado.
+3. **Import + upsert + int** (≈1 dia): `speechImport.ts`, `speechClassifier.ts`, `scripts/import-camara-speeches.mjs`, `package.json`. Int verde: upsert duas vezes = 1 speech + N segmentos estáveis; `classifiedBy='manual'` preservado por upsert sem `req.user`; `segments: undefined` não apaga; busca `contains` normalizada acha acento; estado de skip bate por `audioId+excerptTMs`. Smoke real com `--legislature 57 --limit 3` (com e sem `--skip-transcribe`).
+4. **Prova viva 57ª** (≈0,5 dia de operação; ≈2–3h de máquina, sem trabalho humano): `pnpm camara:import --legislature 57` em `teqo_wt153`; depois **re-executar** e conferir `created=0`/sem duplicata; verificar contagens do relatório (234 listados, `%` com segmentos, falhas), busca por "saude", "SUS", "Petrobras" e filtro por município citado (ex.: Feira de Santana) via Local API; guardar o JSON do relatório.
+5. **Gates / changelog / PR** (≈0,25 dia): `pnpm gate:fast`, `pnpm migrate` + `migrate:status` limpos, `pnpm knip`, `pnpm generate:types` sem diff, `pnpm test:e2e:affected` (prefixo `src/utilities/access` → curated `campaignPermissionProfileHttp`/`campaignDemandVisibility`), changelog, `pnpm push`; PR ready (merge pelo fluxo OPS71). C155 fica fora.
+
+**Quota total:** ~2,75 dias, dentro do appetite; o caminho crítico é a Fase 4 (wall-clock), que roda desatendida.
+
+## Rabbit holes / Não escopo (engenharia)
+
+- **Knip e dynamic import:** o script importa `src/` por dynamic import literal (como `seed-tse-results.mjs:44-50`); não usar `import(variable)` nem `require` — quebra o grafo do knip (`exports: error`). Testes unit/int importam os módulos estaticamente e ancoram o resto.
+- **Top-level pin de `src/utilities`:** tudo em `src/utilities/speech/`; nada solto na raiz de `utilities/`.
+- **Editor de vídeo / clipes / espelhamento S3:** fora (C155/fase própria); só links do VOD e cache local gitignored.
+- **Comissões e audiências:** a API de discursos cobre Plenário; fora.
+- **Normalizar pessoas citadas como `Contact`:** fora; menções são texto pesquisável.
+- **54ª legislatura (2011–2015):** sem trecho por decisão; tratar como vazia, não como falha.
+- **UI/busca do assessor (C154):** aqui só o contrato (constantes de faceta, `normalizeForSearch`, shape da query).
+- **Reclassificação automática agendada / backfill 2011–2026 / produção:** C155.
+- **Busca semântica, `unaccent`, `tsvector`, query builder compartilhado:** fora (D4); se o C154 precisar de mais que `searchText contains normalizeForSearch(q)`, revisitar lá.
+- **Collection `person` paralela / Consent novo:** dado parlamentar público (CC BY 4.0), sem PII de cidadão; o guardrail é access, não Consent.
+- **Rate-limit scheduler / fila:** o run é sequencial e resumível; nada de job queue.
+
+## Riscos e mitigação
+
+- **Anti-crawl (400 a UA de curl):** UA de browser do `CAMARA_USER_AGENT`; nunca bater na URL de trecho com `crawl=no`; só página do evento + `video-sob-demanda`; cache por evento em `data/camara/events/`; delay educado e retry com backoff.
+- **VOD `INDISPONIVEL` ou preso em `GERANDO`:** `parseVodStatus` cobre; fallback YouTube (`urlRegistro` + offset `t/1000 − startTime`); sem `urlRegistro` → "sem vídeo" no relatório; nunca falha o import.
+- **54ª/sessão sem âncora:** "sem trecho" documentado (C152); metadados entram, ASR não; cobertura útil 2015+.
+- **LLM sem chave/falha:** fallback gazetteer-only (`classifiedBy='gazetteer'`, scopes vazios); `--reclassify` reprocessa depois sem re-transcrever.
+- **Custo/tempo do run longo:** ~15–16h de áudio ≈ US$ 0,42 de ASR + ~US$ 0,02 de LLM + 2–3h de parede; script resumível (`sourceKey` + cache de MP4/página), `--limit` para smoke, falha por discurso não aborta o run.
+- **Enum ADD VALUE em transação:** PG 12+ permite, mas o valor não pode ser usado na mesma transação; a migration só adiciona (`IF NOT EXISTS`) e nada insere `communicator` nela; `down` irreversível documentado (precedente do remodel).
+- **Layout da página do evento mudar:** parsers devolvem vazio/null, nunca lançam; unit com HTML congelado detecta a quebra; relatório expõe `withoutExcerpt`/`presidingOfficer=null`.
+- **Rate limit/instabilidade da API:** sequencial, `itens=100`, retry limitado com backoff (padrão do pilot), cache agressivo; falha vira `failures[]` e retoma na próxima execução.
+- **Colisão de `sourceKey`** (dois itens com mesmo `dataHoraInicio|tipo|fase`): detectada no run, segundo item vira falha reportada (nunca sobrescreve em silêncio); se ocorrer, revisitar a chave incluindo `urlTexto`.
+- **Faceta manual sobrescrita:** importer pula campos quando existente é `manual` + hook de preservação para escrita sem `req.user`; int test prova.
+
+## Aceite de engenharia
+
+- [ ] Aceite de produto da intenção coberto: 57ª no banco local com transcrição oficial + segmentos + facetas + menções; reexecução sem duplicar; relatório por execução (total, com vídeo, sem trecho, falhas, tempo, custo ASR); busca por palavra com e sem acento; keywords cruas preservadas.
+- [ ] Invariantes AGENTS/engineering-standards: `access` explícito nas duas collections; migration via `pnpm migrate:create` (+ índice hand-written registrado); `src/utilities/speech/`; `server-only` nos módulos acoplados; identificadores em inglês; sem `Contact` paralelo; `data/camara/` gitignored; `overrideAccess: true` com comentário "bypass"; knip sem órfão; crédito CC BY 4.0 no admin/relatório.
+- [ ] Testes de domínio: unit (facets, search, presiding, role predicate) e int (access allow/deny, upsert idempotente, replace de segmentos, manual preservado, `segments: undefined`, hook `searchText` + `contains`).
+- [ ] Gates: `pnpm gate:fast`; `pnpm migrate`/`migrate:status` limpos; `pnpm knip`; `pnpm generate:types` sem diff; `pnpm test:e2e:affected`; `pnpm push`; changelog `docs/changelog/2026-09-12-c153.md`.
+- [ ] Prova viva registrada (relatório JSON da 57ª + re-run idempotente) no PR.
+
+## Self-score decision-quality
+
+1. **Decisões caras têm rejeitadas?** 5/5 — as seis decisões caras (schema, unicidade/idempotência, provedor LLM, mecanismo de busca, RBAC, dado de "quem presidia") estão no formato Opções|Recomendação|Rejeitadas, com o porquê das rejeitadas; as baratas (nomes de campos, ordem de colunas, prompt) ficam como fill-in.
+2. **Cabe no appetite?** 5/5 — ~2,75 dias contra 2–3 herdados; o único trecho longo é wall-clock desatendido (Fase 4), sem UI e sem pipeline de mídia.
+3. **Rabbit holes nomeados?** 5/5 — os quatro de produto da intenção + os de engenharia (knip/dynamic import, top-level pin, unaccent/tsvector, query builder, job queue, collection de pessoa).
+4. **Depth check reusa?** 5/5 — `withPayloadTransaction`, `payloadAdminOnly`/padrão de `access/*` + barrel, `pg_trgm` hand-written, `seed-tse-results.mjs` como molde de CLI, `SupporterImportBatch` como precedente de admin-hidden, `municipalityCatalog` como gazetteer, e o `camaraSpeeches.mjs` **estendido** (não duplicado) para a mesa. Nenhum pass-through novo.
+5. **Intenção permanece satisfeita?** 5/5 — a engenharia não reescreveu o outcome: sem UI, sem espelho, sem `Contact`, leitura restrita com papel novo que não herda staff, e a única reavaliação (D3) entrega a mesma opção A da intenção por uma rota que já tem chave, registrada para o gate.
+
+## Débitos (triage pós-simplify)
+
+Nenhum achado dos dois revisores virou Issue nova (nenhum `expensive_lock` com score ≥4). O que ficou:
+
+**Deferido com gatilho:**
+
+- **Cache/paginação por script** (`scripts/import-camara-speeches.mjs` vs `pilot-camara-speeches.mjs`): a extração parou no HTTP (`scripts/lib/camaraFetch.mjs`); o fluxo de cache de evento/paginação vive só no import. Gatilho: o pilot virar recorrente ou um 3º consumidor do fluxo.
+- **Constantes Deep Infra duplicadas** (`scripts/lib/camaraSpeeches.mjs` vs `src/utilities/ai/deepInfraTranscribe.ts`): o pilot roda em node puro e não importa `src/`. Gatilho: mudança de endpoint/modelo ou 3º consumidor.
+- **`llm.calls` conta tentativas** mesmo sem `DEEPINFRA_API_KEY` (o import degrada para gazetteer por design; o pilot faz preflight e morre). Gatilho: próximo ajuste do relatório de import.
+- **Índice em `speechSegment.order`**: YAGNI (poucos segmentos por discurso; a FK `speech` já filtra). Gatilho: EXPLAIN/latência real no C154.
+- **Chips do Sollinha para `communicator`** (`src/lib/sollinhaOpeningQuestions.ts`): o fallback não-staff devolve chips de líder; inofensivo (access é checado no servidor), mas fora do vertical. Gatilho: C154, dono da vertical de comunicação.
+
+**Descartado:** redundância do `searchText` no create (exigida pelo `required` do tipo gerado); ordenação defensiva dupla do presiding officer; comentários "Intentional bypass" por site (exigidos pelo ratchet); wall-clock do run (resumível/idempotente).
+
+**Já resolvido no simplify (não reabrir):**
+
+- Perda de mídia quando o evento não resolve no run: `excerptChanged` só compara com excerpt presente e `preserveStoredMedia` mantém `audioId`/`excerptTMs`/VOD/duração (bug HIGH do revisor de qualidade).
+- `--reclassify` com LLM fora do ar não rebaixa mais uma classificação `llm` para gazetteer (preserva people/programs/projects).
+- Envelope do provider validado por zod (sem `as` no `speechClassifier`).
+- `parseDurationToSeconds` estrito (rejeita `1::30`, `1:99:99`, `0h99'00"`).
+- Testes unit dos helpers puros (`speechSourceKey`, `parseOfficialKeywords`, `legislatureForDate`, `parseDurationToSeconds`).
+- Léxico gazetteer separado em `src/lib/speechGazetteer.ts` (o contrato `speechFacets.ts` fica leve para o bundle do C154).
+- Aliases de label do `topicByKey` derivados de `SPEECH_TOPICS` (fim da lista manual).
+- Guard de `year` no import (data malformada não vira 0).
--- a/package.json
+++ b/package.json
@ -82,6 +82,7 @@
    "images:resize": "node scripts/resize-images.mjs",
    "images:fill": "node scripts/fill-image.mjs",
    "camara:pilot": "node scripts/pilot-camara-speeches.mjs",
+    "camara:import": "cross-env NODE_OPTIONS=\"--no-deprecation --import=tsx/esm --import=./scripts/seed-loader.mjs\" node scripts/import-camara-speeches.mjs",
    "typecheck": "tsc --noEmit",
    "typecheck:tsc6": "tsc6 --noEmit",
    "prepare": "husky",
--- a/scripts/import-camara-speeches.mjs
+++ b/scripts/import-camara-speeches.mjs
@ -0,0 +1,510 @@
+/**
+ * C153 import — populates the speech catalog from the Câmara sources:
+ * open-data speeches → event on that date → speaker excerpt on the event page
+ * → async VOD → Deep Infra Whisper segments (minutagem) + facet classification
+ * (gazetteer + LLM) → idempotent upsert into `speech`/`speechSegment`.
+ *
+ * Re-running is safe: records are matched by `sourceKey`, metadata is
+ * refreshed, VOD/ASR is skipped when the excerpt identity and segments are
+ * already stored, and a `manual` facet curation is never overwritten.
+ *
+ * Media and reports land under `data/camara/` (gitignored). The local-DB guard
+ * refuses a non-local DATABASE_URL; production is C155.
+ *
+ * Usage:
+ *   pnpm camara:import --legislature 57
+ *   pnpm camara:import --date 2023-02-07 --limit 3
+ *   pnpm camara:import --legislature 57 --skip-transcribe
+ *   pnpm camara:import --legislature 57 --reclassify
+ */
+import { mkdir, readFile, writeFile } from 'node:fs/promises'
+import { dirname, join } from 'node:path'
+
+import { getPayload } from 'payload'
+
+import { assertLocalDatabase } from './assert-local-database.mjs'
+import {
+  DEEPINFRA_COST_PER_MINUTE_USD,
+  downloadToBuffer,
+  eventPageUrl,
+  eventsUrl,
+  getJson,
+  getText,
+  resolveVod,
+  speechesUrl,
+  transcribeSpeechAudio,
+} from './lib/camaraFetch.mjs'
+import {
+  LEGISLATURE_RANGES,
+  SOLLA_DEPUTY_ID,
+  legislatureForDate,
+  matchExcerpt,
+  parseDurationToSeconds,
+  parseEventExcerpts,
+  parseOfficialKeywords,
+  parsePresidingOfficerTransitions,
+  resolvePresidingOfficer,
+  selectSpeechEvents,
+  speechDate,
+  speechSourceKey,
+} from './lib/camaraSpeeches.mjs'
+import { dieWithLabel, ensureCachedDownload, loadCliEnv } from './lib/cli.mjs'
+
+loadCliEnv()
+
+const die = dieWithLabel('camara:import')
+
+const config = (await import('../src/payload.config.ts')).default
+const { findSpeechImportState, upsertSpeechBundle } =
+  await import('../src/utilities/speech/speechImport.ts')
+const { classifySpeech } = await import('../src/utilities/speech/speechClassifier.ts')
+
+const DEFAULT_OUT_DIR = 'data/camara'
+const MAX_EVENT_CANDIDATES = 5
+const EVENT_TYPE_PRIORITY = [
+  'Sessão Deliberativa',
+  'Breves Comunicações',
+  'Sessão Não Deliberativa',
+  'Sessão Solene',
+  'Outro Evento',
+]
+
+// ---------------------------------------------------------------------------
+// CLI
+// ---------------------------------------------------------------------------
+
+const parseArgs = (argv) => {
+  const options = {
+    legislature: 57,
+    date: null,
+    limit: null,
+    speaker: 'Jorge Solla',
+    out: DEFAULT_OUT_DIR,
+    skipTranscribe: false,
+    reclassify: false,
+    help: false,
+  }
+  for (let index = 0; index < argv.length; index += 1) {
+    const arg = argv[index]
+    const value = () => {
+      const next = argv[++index]
+      if (next === undefined || next.startsWith('--')) die(`faltou valor para ${arg}.`)
+      return next
+    }
+    if (arg === '--help' || arg === '-h') options.help = true
+    else if (arg === '--skip-transcribe') options.skipTranscribe = true
+    else if (arg === '--reclassify') options.reclassify = true
+    else if (arg === '--legislature') options.legislature = Number(value())
+    else if (arg === '--date') options.date = value()
+    else if (arg === '--limit') options.limit = Number(value())
+    else if (arg === '--speaker') options.speaker = value()
+    else if (arg === '--out') options.out = value()
+    else die(`argumento desconhecido: ${arg}`)
+  }
+  if (!options.help && !options.date && !LEGISLATURE_RANGES[options.legislature]) {
+    die('informe --legislature <54|55|56|57> ou --date <YYYY-MM-DD> (veja --help).')
+  }
+  if (!options.help && options.date && !/^\d{4}-\d{2}-\d{2}$/.test(options.date)) {
+    die('--date deve ser YYYY-MM-DD.')
+  }
+  if (options.limit !== null && (!Number.isInteger(options.limit) || options.limit < 1)) {
+    die('--limit deve ser >= 1.')
+  }
+  if (!options.help && options.out.includes('..')) {
+    die('--out não pode escapar do diretório do repo (sem "..").')
+  }
+  return options
+}
+
+const HELP = `
+Uso: pnpm camara:import [--legislature 54|55|56|57] [--date YYYY-MM-DD] [opções]
+
+Opções:
+  --legislature <n>   54|55|56|57 (default 57; ignorado se --date)
+  --date <YYYY-MM-DD> discursos de um único dia
+  --limit <n>         quantos discursos processar (default: todos)
+  --speaker <nome>    nome do orador a casar no evento (default "Jorge Solla")
+  --out <dir>         diretório de cache/artefatos (default data/camara)
+  --skip-transcribe   não chama a Deep Infra (preserva os segmentos existentes)
+  --reclassify        refaz a classificação mesmo já tendo facetas da LLM
+  --help              esta ajuda
+
+A página do evento é cacheada em <out>/events; apague-a para reler a Câmara.
+`
+
+// ---------------------------------------------------------------------------
+// Caches
+// ---------------------------------------------------------------------------
+
+const eventsByDate = new Map()
+const eventPages = new Map()
+
+async function getEventsForDate(date) {
+  const cached = eventsByDate.get(date)
+  if (cached) return cached
+  const events = (await getJson(eventsUrl(date))).dados ?? []
+  eventsByDate.set(date, events)
+  return events
+}
+
+async function loadEventPage(eventId, options) {
+  const cached = eventPages.get(eventId)
+  if (cached) return cached
+  const path = join(options.out, 'events', `${eventId}.html`)
+  let html
+  try {
+    html = await readFile(path, 'utf8')
+  } catch (error) {
+    if (error?.code !== 'ENOENT') throw error
+    html = await getText(eventPageUrl(eventId), { timeoutMs: 60_000 })
+    await mkdir(dirname(path), { recursive: true })
+    await writeFile(path, html)
+  }
+  eventPages.set(eventId, html)
+  return html
+}
+
+async function fetchSpeeches(deputyId, from, to) {
+  const speeches = []
+  for (let page = 1; ; page += 1) {
+    const body = await getJson(`${speechesUrl(deputyId, from, to)}&pagina=${page}`)
+    const batch = body.dados ?? []
+    speeches.push(...batch)
+    const hasNext = (body.links ?? []).some((link) => link?.rel === 'next')
+    if (!hasNext || batch.length === 0) break
+  }
+  return speeches
+}
+
+// ---------------------------------------------------------------------------
+// Pipeline
+// ---------------------------------------------------------------------------
+
+const eventPriority = (event) => {
+  const index = EVENT_TYPE_PRIORITY.indexOf(event?.descricaoTipo)
+  return index === -1 ? EVENT_TYPE_PRIORITY.length : index
+}
+
+/** Nearest event of the speech date, preferred session types first. */
+async function resolveSpeechContext(speech, options) {
+  const candidates = selectSpeechEvents(
+    await getEventsForDate(speechDate(speech.dataHoraInicio)),
+    speech.dataHoraInicio,
+  )
+  const ordered = [...candidates].sort((left, right) => eventPriority(left) - eventPriority(right))
+  for (const event of ordered.slice(0, MAX_EVENT_CANDIDATES)) {
+    const html = await loadEventPage(event.id, options)
+    const excerpt = matchExcerpt(
+      parseEventExcerpts(html, event.id),
+      speech.dataHoraInicio,
+      options.speaker,
+    )
+    if (excerpt) return { event, html, excerpt }
+  }
+  return { event: ordered[0] ?? null, html: null, excerpt: null }
+}
+
+async function transcribeExcerpt(eventId, excerpt, options) {
+  const vod = await resolveVod(eventId, excerpt)
+  if (vod.status.state !== 'PRONTO' || !vod.status.video?.downloadUrl) {
+    return {
+      vodState: vod.status.state,
+      failure: `VOD não ficou PRONTO (estado=${vod.status.state})`,
+    }
+  }
+
+  const cached = await ensureCachedDownload({
+    label: 'camara:import',
+    key: `${eventId}-${excerpt.audioId}-${excerpt.tMs}`,
+    url: vod.status.video.downloadUrl,
+    ext: 'mp4',
+    cacheDir: options.out,
+    download: downloadToBuffer,
+  })
+  const { transcription, elapsedMs } = await transcribeSpeechAudio(cached.buffer)
+  return {
+    vodState: vod.status.state,
+    segments: transcription.segments.map((segment) => ({
+      startSeconds: segment.start,
+      endSeconds: segment.end,
+      text: segment.text,
+    })),
+    durationSeconds: transcription.duration,
+    playbackUrl: vod.status.video.playbackUrl,
+    downloadUrl: vod.status.video.downloadUrl,
+    audioSeconds: transcription.duration ?? 0,
+    elapsedMs,
+  }
+}
+
+async function processSpeech(payload, speech, options, run) {
+  const sourceKey = speechSourceKey(speech)
+  const report = {
+    sourceKey,
+    speechAt: speech.dataHoraInicio,
+    eventId: null,
+    excerpt: null,
+    vodState: null,
+    presidingOfficer: null,
+    classifiedBy: null,
+    segmentCount: 0,
+    created: false,
+    manualFacetsPreserved: false,
+    failures: [],
+  }
+
+  const existing = await findSpeechImportState(payload, sourceKey)
+  const keywords = parseOfficialKeywords(speech.keywords)
+  const date = speechDate(speech.dataHoraInicio)
+  const year = /^\d{4}-\d{2}-\d{2}$/.test(date) ? Number(date.slice(0, 4)) : null
+
+  let context = { event: null, html: null, excerpt: null }
+  try {
+    context = await resolveSpeechContext(speech, options)
+  } catch (error) {
+    report.failures.push({ stage: 'event', message: error?.message ?? String(error) })
+  }
+  const { event, excerpt } = context
+  if (event) report.eventId = event.id
+  if (excerpt) {
+    report.excerpt = { audioId: excerpt.audioId, tMs: excerpt.tMs, startTime: excerpt.startTime }
+  }
+  const transitions = context.html ? parsePresidingOfficerTransitions(context.html) : []
+  report.presidingOfficer = excerpt ? resolvePresidingOfficer(transitions, excerpt.tMs) : null
+
+  // Identity/media carry-over: a run that failed to resolve the event page must
+  // not wipe the stored excerpt or force a re-transcription next time.
+  const excerptChanged =
+    Boolean(excerpt) &&
+    Boolean(existing) &&
+    (existing.audioId !== excerpt.audioId || existing.excerptTMs !== excerpt.tMs)
+  const preserveStoredMedia = !excerpt || !excerptChanged
+  const needsTranscription =
+    Boolean(excerpt) &&
+    !options.skipTranscribe &&
+    (!existing || excerptChanged || existing.segmentCount === 0)
+
+  let media = {}
+  if (needsTranscription) {
+    try {
+      media = await transcribeExcerpt(event.id, excerpt, options)
+      if (media.failure) report.failures.push({ stage: 'vod', message: media.failure })
+      if (media.segments) {
+        run.asr.calls += 1
+        run.asr.audioSeconds += media.audioSeconds
+        run.asr.elapsedMs += media.elapsedMs ?? 0
+      }
+    } catch (error) {
+      run.asr.failures += 1
+      report.failures.push({ stage: 'asr', message: error?.message ?? String(error) })
+    }
+  }
+
+  const skipClassification =
+    existing?.classifiedBy === 'manual' || (existing?.classifiedBy === 'llm' && !options.reclassify)
+  let facets
+  if (skipClassification) {
+    report.classifiedBy = existing.classifiedBy
+  } else {
+    const classification = await classifySpeech({
+      transcript: speech.transcricao ?? '',
+      summary: speech.sumario ?? null,
+      keywords,
+    })
+    run.llm.calls += 1
+    if (classification.llm.used) {
+      facets = classification.facets
+      report.classifiedBy = classification.facets.classifiedBy
+      run.llm.used += 1
+      run.llm.totalTokens += classification.llm.totalTokens ?? 0
+      run.llm.estimatedCostUsd += classification.llm.estimatedCostUsd ?? 0
+    } else if (existing?.classifiedBy === 'llm') {
+      // Never downgrade a stored LLM classification to the gazetteer fallback
+      // (that would drop the extracted people/programs/projects).
+      report.classifiedBy = existing.classifiedBy
+      report.failures.push({
+        stage: 'classify',
+        message: `LLM indisponível (${classification.llm.error ?? 'erro'}); facetas anteriores preservadas`,
+      })
+      run.llm.failed += 1
+    } else {
+      facets = classification.facets
+      report.classifiedBy = classification.facets.classifiedBy
+      run.llm.failed += 1
+      if (classification.llm.error && run.llm.sampleError === null) {
+        run.llm.sampleError = classification.llm.error
+      }
+    }
+  }
+
+  const durationSeconds =
+    media.durationSeconds ??
+    (preserveStoredMedia ? (existing?.durationSeconds ?? null) : null) ??
+    parseDurationToSeconds(excerpt?.duration)
+  const bundle = {
+    sourceKey,
+    speechAt: String(speech.dataHoraInicio ?? '').trim(),
+    year,
+    legislature: legislatureForDate(speech.dataHoraInicio),
+    type: speech.tipoDiscurso ?? null,
+    phase: speech.faseEvento?.titulo ?? null,
+    durationSeconds,
+    summary: speech.sumario ?? null,
+    officialTranscript: speech.transcricao ?? null,
+    officialTextUrl: speech.urlTexto ?? null,
+    keywords,
+    eventId: event?.id ?? null,
+    eventType: event?.descricaoTipo ?? null,
+    eventStartAt: event?.dataHoraInicio ?? null,
+    eventEndAt: event?.dataHoraFim ?? null,
+    youtubeUrl: event?.urlRegistro ?? null,
+    presidingOfficer: report.presidingOfficer,
+    audioId: excerpt?.audioId ?? (preserveStoredMedia ? (existing?.audioId ?? null) : null),
+    excerptTMs: excerpt?.tMs ?? (preserveStoredMedia ? (existing?.excerptTMs ?? null) : null),
+    vodPlaybackUrl:
+      media.playbackUrl ?? (preserveStoredMedia ? (existing?.vodPlaybackUrl ?? null) : null),
+    vodDownloadUrl:
+      media.downloadUrl ?? (preserveStoredMedia ? (existing?.vodDownloadUrl ?? null) : null),
+    facets,
+    segments: media.segments,
+  }
+
+  try {
+    const counts = await upsertSpeechBundle(payload, bundle)
+    report.created = counts.created
+    report.manualFacetsPreserved = counts.manualFacetsPreserved
+    report.segmentCount = media.segments ? media.segments.length : (existing?.segmentCount ?? 0)
+    run.totals.created += counts.created ? 1 : 0
+    run.totals.updated += counts.created ? 0 : 1
+  } catch (error) {
+    report.failures.push({ stage: 'upsert', message: error?.message ?? String(error) })
+  }
+
+  run.totals.withExcerpt += excerpt ? 1 : 0
+  run.totals.withoutExcerpt += excerpt ? 0 : 1
+  run.totals.withSegments += report.segmentCount > 0 ? 1 : 0
+  run.totals.failed += report.failures.length > 0 ? 1 : 0
+  return report
+}
+
+// ---------------------------------------------------------------------------
+// Report
+// ---------------------------------------------------------------------------
+
+const printReport = (run) => {
+  const { totals, asr, llm, failures } = run
+  console.log('\n=== Relatório do import C153 ===')
+  console.log(
+    `discursos: ${totals.listed} listados, ${totals.processed} processados ` +
+      `(${totals.created} criados, ${totals.updated} atualizados)`,
+  )
+  console.log(
+    `vídeo: ${totals.withExcerpt} com trecho, ${totals.withoutExcerpt} sem trecho; ` +
+      `${totals.withSegments} com segmentos`,
+  )
+  console.log(
+    `ASR: ${asr.calls} chamadas, ${(asr.audioSeconds / 60).toFixed(1)} min de áudio, ` +
+      `~US$ ${((asr.audioSeconds * DEEPINFRA_COST_PER_MINUTE_USD) / 60).toFixed(4)}, ` +
+      `${(asr.elapsedMs / 1000).toFixed(0)}s de processamento, ${asr.failures} falhas`,
+  )
+  console.log(
+    `LLM: ${llm.used}/${llm.calls} usadas, ${llm.totalTokens} tokens, ` +
+      `~US$ ${llm.estimatedCostUsd.toFixed(4)}${llm.failed > 0 ? `, ${llm.failed} falhas` : ''}`,
+  )
+  console.log(`tempo total: ${(run.elapsedMs / 1000).toFixed(1)}s; falhas: ${failures.length}`)
+  for (const failure of failures) {
+    console.log(`  ✗ ${failure.speechAt} [${failure.stage}] ${failure.message}`)
+  }
+}
+
+const speechLine = (report) => {
+  const parts = [
+    report.created ? 'criado' : 'atualizado',
+    report.excerpt ? `trecho a=${report.excerpt.audioId} t=${report.excerpt.tMs}` : 'sem trecho',
+  ]
+  if (report.vodState) parts.push(`VOD ${report.vodState}`)
+  if (report.segmentCount > 0) parts.push(`${report.segmentCount} segmentos`)
+  if (report.presidingOfficer) parts.push(`presidiu ${report.presidingOfficer}`)
+  if (report.classifiedBy) parts.push(`facetas ${report.classifiedBy}`)
+  if (report.failures.length > 0) parts.push(`${report.failures.length} falha(s)`)
+  return parts.join(' · ')
+}
+
+// ---------------------------------------------------------------------------
+// Main
+// ---------------------------------------------------------------------------
+
+async function main() {
+  const options = parseArgs(process.argv.slice(2))
+  if (options.help) {
+    console.log(HELP)
+    process.exit(0)
+  }
+
+  assertLocalDatabase('camara:import', 'O import C153 escreve no banco local; produção é C155.')
+  const payload = await getPayload({ config })
+
+  const from = options.date ?? LEGISLATURE_RANGES[options.legislature][0]
+  const to = options.date ?? LEGISLATURE_RANGES[options.legislature][1]
+  const startedAt = Date.now()
+  console.log(
+    `[camara:import] deputado ${SOLLA_DEPUTY_ID} | ${from}..${to} | speaker "${options.speaker}"`,
+  )
+
+  const speeches = await fetchSpeeches(SOLLA_DEPUTY_ID, from, to)
+  const targets = options.limit === null ? speeches : speeches.slice(0, options.limit)
+  console.log(`[camara:import] ${speeches.length} discursos na API; processando ${targets.length}`)
+  if (targets.length === 0) die(`nenhum discurso em ${from}..${to}.`)
+
+  const run = {
+    runAt: new Date().toISOString(),
+    range: { from, to },
+    options: {
+      legislature: options.date ? null : options.legislature,
+      date: options.date,
+      limit: options.limit,
+      speaker: options.speaker,
+      skipTranscribe: options.skipTranscribe,
+      reclassify: options.reclassify,
+    },
+    totals: {
+      listed: speeches.length,
+      processed: 0,
+      created: 0,
+      updated: 0,
+      withExcerpt: 0,
+      withoutExcerpt: 0,
+      withSegments: 0,
+      failed: 0,
+    },
+    asr: { calls: 0, audioSeconds: 0, elapsedMs: 0, failures: 0 },
+    llm: { calls: 0, used: 0, failed: 0, totalTokens: 0, estimatedCostUsd: 0, sampleError: null },
+    elapsedMs: 0,
+    speeches: [],
+    failures: [],
+  }
+
+  for (const [index, speech] of targets.entries()) {
+    const report = await processSpeech(payload, speech, options, run)
+    run.totals.processed += 1
+    run.speeches.push(report)
+    for (const failure of report.failures) {
+      run.failures.push({ sourceKey: report.sourceKey, speechAt: report.speechAt, ...failure })
+    }
+    console.log(`[${index + 1}/${targets.length}] ${report.speechAt} — ${speechLine(report)}`)
+  }
+
+  run.elapsedMs = Date.now() - startedAt
+  printReport(run)
+
+  const reportPath = join(options.out, 'reports', `import-${run.runAt.replace(/[:.]/g, '-')}.json`)
+  await mkdir(dirname(reportPath), { recursive: true })
+  await writeFile(reportPath, JSON.stringify(run, null, 2))
+  console.log(`\n[camara:import] relatório JSON: ${reportPath}`)
+  process.exit(0)
+}
+
+main().catch((error) => {
+  console.error(error)
+  die(error?.message || String(error))
+})
--- a/scripts/lib/camaraFetch.mjs
+++ b/scripts/lib/camaraFetch.mjs
@ -0,0 +1,119 @@
+/**
+ * HTTP surface shared by the Câmara speech scripts (C152 pilot + C153 import):
+ * the open-data endpoints, the anti-crawl browser UA, bounded retry, the async
+ * VOD poll and the Deep Infra transcription call. Kept apart from
+ * `camaraSpeeches.mjs` (pure) so the fragile rules stay unit-testable without
+ * network.
+ */
+import {
+  buildVodUrl,
+  CAMARA_USER_AGENT,
+  DEEPINFRA_TRANSCRIBE_URL,
+  DEEPINFRA_WHISPER_MODEL,
+  normalizeTranscription,
+  parseVodStatus,
+} from './camaraSpeeches.mjs'
+
+const camaraHeaders = { 'User-Agent': CAMARA_USER_AGENT, Accept: 'application/json' }
+
+const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms))
+
+async function fetchOnce(url, { timeoutMs = 30_000, headers = camaraHeaders } = {}) {
+  const response = await fetch(url, { headers, signal: AbortSignal.timeout(timeoutMs) })
+  if (!response.ok) throw new Error(`HTTP ${response.status} em ${url}`)
+  return response
+}
+
+/** The Câmara API drops connections now and then — bounded retry for GETs. */
+export async function getText(url, { attempts = 3, ...options } = {}) {
+  let lastError
+  for (let attempt = 1; attempt <= attempts; attempt += 1) {
+    try {
+      return await (await fetchOnce(url, options)).text()
+    } catch (error) {
+      lastError = error
+      if (attempt < attempts) await sleep(2_000)
+    }
+  }
+  throw lastError
+}
+
+export const getJson = async (url, options) => JSON.parse(await getText(url, options))
+
+export const downloadToBuffer = async (url, { attempts = 3, timeoutMs = 120_000 } = {}) => {
+  let lastError
+  for (let attempt = 1; attempt <= attempts; attempt += 1) {
+    try {
+      const response = await fetchOnce(url, {
+        timeoutMs,
+        headers: { 'User-Agent': CAMARA_USER_AGENT },
+      })
+      return Buffer.from(await response.arrayBuffer())
+    } catch (error) {
+      lastError = error
+      if (attempt < attempts) await sleep(3_000)
+    }
+  }
+  throw lastError
+}
+
+export const speechesUrl = (deputyId, from, to) =>
+  `https://dadosabertos.camara.leg.br/api/v2/deputados/${deputyId}/discursos` +
+  `?dataInicio=${from}&dataFim=${to}&itens=100&ordenarPor=dataHoraInicio&ordem=ASC`
+
+export const eventsUrl = (date) =>
+  `https://dadosabertos.camara.leg.br/api/v2/eventos?dataInicio=${date}&dataFim=${date}&itens=100`
+
+export const eventPageUrl = (eventId) => `https://www.camara.leg.br/evento-legislativo/${eventId}`
+
+const VOD_POLL_INTERVAL_MS = 5_000
+const VOD_POLL_MAX_ATTEMPTS = 40
+
+/** Polls the async VOD generation until PRONTO (or gives up). */
+export async function resolveVod(eventId, excerpt) {
+  const url = buildVodUrl(eventId, excerpt.audioId, excerpt.tMs)
+  let lastStatus = null
+  for (let attempt = 1; attempt <= VOD_POLL_MAX_ATTEMPTS; attempt += 1) {
+    const status = parseVodStatus(await getJson(url, { timeoutMs: 60_000 }))
+    lastStatus = status
+    if (status.state === 'PRONTO' || status.state === 'INDISPONIVEL') {
+      return { url, status, attempts: attempt }
+    }
+    await sleep(VOD_POLL_INTERVAL_MS)
+  }
+  return { url, status: lastStatus, attempts: VOD_POLL_MAX_ATTEMPTS }
+}
+
+export const DEEPINFRA_COST_PER_MINUTE_USD = 0.00045
+
+/**
+ * Sends the excerpt MP4 (video/mp4 is accepted; no ffmpeg) to Deep Infra's
+ * OpenAI-compatible endpoint and returns the normalized transcription.
+ * Throws on missing key or provider error — the caller owns the failure policy.
+ */
+export async function transcribeSpeechAudio(buffer) {
+  const apiKey = process.env.DEEPINFRA_API_KEY
+  if (!apiKey) {
+    throw new Error('DEEPINFRA_API_KEY ausente — configure no .env.local ou no env do processo.')
+  }
+  const form = new FormData()
+  form.append('file', new Blob([buffer], { type: 'video/mp4' }), 'trecho.mp4')
+  form.append('model', DEEPINFRA_WHISPER_MODEL)
+  form.append('language', 'pt')
+  form.append('response_format', 'verbose_json')
+  form.append('timestamp_granularities[]', 'segment')
+  const startedAt = Date.now()
+  const response = await fetch(DEEPINFRA_TRANSCRIBE_URL, {
+    method: 'POST',
+    headers: { Authorization: `Bearer ${apiKey}` },
+    body: form,
+    signal: AbortSignal.timeout(600_000),
+  })
+  if (!response.ok) {
+    throw new Error(`Deep Infra HTTP ${response.status}: ${(await response.text()).slice(0, 300)}`)
+  }
+  return {
+    transcription: normalizeTranscription(await response.json()),
+    elapsedMs: Date.now() - startedAt,
+  }
+}
--- a/scripts/lib/camaraSpeeches.mjs
+++ b/scripts/lib/camaraSpeeches.mjs
@ -1,7 +1,9 @@
 /**
- * Pure derivation for the C152 pilot (`scripts/pilot-camara-speeches.mjs`):
- * matching a Câmara speech to its legislative event and to the speaker's video
- * excerpt, plus normalizing the VOD status and the Deep Infra transcription.
+ * Pure derivation for the Câmara speech scripts — the C152 pilot
+ * (`scripts/pilot-camara-speeches.mjs`) and the C153 catalog import
+ * (`scripts/import-camara-speeches.mjs`): matching a Câmara speech to its
+ * legislative event and to the speaker's video excerpt, the presiding-officer
+ * transitions, plus normalizing the VOD status and the Deep Infra transcription.
 *
 * No I/O here — the CLI threads fetch/download/transcribe through these
 * decisions, so the fragile rules (HTML anchors, time-of-day matching, epoch
@ -73,6 +75,83 @@ export const speechTimeOfDaySeconds = (iso) => clockToSeconds(String(iso ?? '').
 /** `YYYY-MM-DD` prefix of a naive datetime. */
 export const speechDate = (iso) => String(iso ?? '').slice(0, 10)

+/**
+ * Stable natural key of an open-data speech. The API has no id, so identity is
+ * the naive datetime plus the speech type and event phase — the record is
+ * updated in place across imports (C153 idempotency).
+ *
+ * @param {{ dataHoraInicio?: string, tipoDiscurso?: string, faseEvento?: { titulo?: string } } | null | undefined} speech
+ * @returns {string}
+ */
+export const speechSourceKey = (speech) =>
+  [
+    String(speech?.dataHoraInicio ?? '').trim(),
+    String(speech?.tipoDiscurso ?? '').trim(),
+    String(speech?.faseEvento?.titulo ?? '').trim(),
+  ].join('|')
+
+/**
+ * Official keywords come as a newline-separated string; keep each keyword raw.
+ *
+ * @param {unknown} raw
+ * @returns {string[]}
+ */
+export const parseOfficialKeywords = (raw) =>
+  String(raw ?? '')
+    .split(/[\r\n]+/)
+    .map((keyword) => keyword.trim())
+    .filter(Boolean)
+
+/**
+ * Legislature of a `YYYY-MM-DD` date per `LEGISLATURE_RANGES`, or null when it
+ * falls outside the supported range.
+ *
+ * @param {string} iso
+ * @returns {string | null}
+ */
+export const legislatureForDate = (iso) => {
+  const date = speechDate(iso)
+  if (!/^\d{4}-\d{2}-\d{2}$/.test(date)) return null
+  for (const [legislature, [from, to]] of Object.entries(LEGISLATURE_RANGES)) {
+    if (date >= from && date <= to) return legislature
+  }
+  return null
+}
+
+/**
+ * Parses the Câmara duration spellings into seconds: `0h04'03"` (excerpt card)
+ * and `0:04:07` / `4:07` (VOD status). Null when unparseable.
+ *
+ * @param {unknown} value
+ * @returns {number | null}
+ */
+export const parseDurationToSeconds = (value) => {
+  const raw = String(value ?? '').trim()
+  if (raw === '') return null
+
+  const card = /^(\d+)h(\d{1,2})'(\d{1,2})"?$/.exec(raw)
+  if (card) {
+    const minutes = Number(card[2])
+    const seconds = Number(card[3])
+    if (minutes > 59 || seconds > 59) return null
+    return Number(card[1]) * 3600 + minutes * 60 + seconds
+  }
+
+  const parts = raw.split(':')
+  if (!parts.every((part) => /^\d{1,2}$/.test(part))) return null
+  if (parts.length === 3) {
+    const [hours, minutes, seconds] = parts.map(Number)
+    if (minutes > 59 || seconds > 59) return null
+    return hours * 3600 + minutes * 60 + seconds
+  }
+  if (parts.length === 2) {
+    const [minutes, seconds] = parts.map(Number)
+    if (seconds > 59) return null
+    return minutes * 60 + seconds
+  }
+  return null
+}
+
 const integerOrNull = (value) => {
  const n = Number(value)
  return Number.isInteger(n) && n > 0 ? n : null
@ -200,6 +279,67 @@ export const matchExcerpt = (excerpts, speechIso, speaker) => {
  return bySpeaker.reduce((best, current) => (delta(current) < delta(best) ? current : best))
 }

+/**
+ * Parses the "Troca da mesa" transition groups of an event page. Each heading
+ * `<h4 class="g-l-assista__categoria-outros-videos">Troca da mesa Presidente X
+ * por Participante Y</h4>` labels the excerpt cards that follow it; the newest
+ * card of the group (`t` epoch ms) marks the observed moment of the change.
+ * Returns transitions ascending by time; `[]` when the layout changed or the
+ * page recorded none.
+ *
+ * @param {string} html
+ * @returns {Array<{ from: string, to: string, tMs: number }>}
+ */
+export const parsePresidingOfficerTransitions = (html) => {
+  const page = String(html ?? '')
+  const headingPattern =
+    /<h4 class="g-l-assista__categoria-outros-videos">Troca da mesa Presidente (.+?) por Participante (.+?)<\/h4>/g
+  const transitions = []
+  let heading
+  while ((heading = headingPattern.exec(page)) !== null) {
+    const nextHeading = page.indexOf(
+      'g-l-assista__categoria-outros-videos',
+      headingPattern.lastIndex,
+    )
+    const group = page.slice(heading.index, nextHeading === -1 ? page.length : nextHeading)
+    const times = [...group.matchAll(/<a id="link-trecho-video" href="[^"]*?t&#x3D;(\d+)[^"]*"/g)]
+      .map((match) => Number(match[1]))
+      .filter((value) => Number.isInteger(value) && value > 0)
+    if (times.length === 0) continue
+    transitions.push({
+      from: heading[1].trim(),
+      to: heading[2].trim(),
+      tMs: Math.max(...times),
+    })
+  }
+  return transitions.sort((left, right) => left.tMs - right.tMs)
+}
+
+/**
+ * Who presided at `tMs` according to the transition groups: the incoming
+ * officer (`to`) of the latest transition at or before the moment; before the
+ * first transition, the outgoing officer (`from`) of the earliest one. Returns
+ * null when the page recorded no transitions (unknown, never invented).
+ *
+ * @param {ReturnType<typeof parsePresidingOfficerTransitions>} transitions
+ * @param {number} tMs
+ * @returns {string | null}
+ */
+export const resolvePresidingOfficer = (transitions, tMs) => {
+  const ordered = (Array.isArray(transitions) ? transitions : [])
+    .filter((transition) => Number.isFinite(transition?.tMs))
+    .sort((left, right) => left.tMs - right.tMs)
+  const target = Number(tMs)
+  if (ordered.length === 0 || !Number.isFinite(target)) return null
+
+  let current = ordered[0].from
+  for (const transition of ordered) {
+    if (transition.tMs > target) break
+    current = transition.to
+  }
+  return current
+}
+
 /**
 * Normalizes a Deep Infra transcription response. Accepts the OpenAI-compat
 * shape (`segments[{start,end,text}]`) and the native whisper shape
--- a/scripts/pilot-camara-speeches.mjs
+++ b/scripts/pilot-camara-speeches.mjs
@ -21,16 +21,21 @@ import { access, mkdir, readFile, writeFile } from 'node:fs/promises'
 import { join } from 'node:path'

 import {
-  CAMARA_USER_AGENT,
-  DEEPINFRA_TRANSCRIBE_URL,
-  DEEPINFRA_WHISPER_MODEL,
+  DEEPINFRA_COST_PER_MINUTE_USD,
+  downloadToBuffer,
+  eventPageUrl,
+  eventsUrl,
+  getJson,
+  getText,
+  resolveVod,
+  speechesUrl,
+  transcribeSpeechAudio,
+} from './lib/camaraFetch.mjs'
+import {
  LEGISLATURE_RANGES,
  SOLLA_DEPUTY_ID,
-  buildVodUrl,
  matchExcerpt,
-  normalizeTranscription,
  parseEventExcerpts,
-  parseVodStatus,
  selectSpeechEvents,
 } from './lib/camaraSpeeches.mjs'
 import { dieWithLabel, ensureCachedDownload, loadCliEnv } from './lib/cli.mjs'
@ -39,9 +44,6 @@ loadCliEnv()

 const die = dieWithLabel('camara:pilot')

-const DEEPINFRA_COST_PER_MINUTE_USD = 0.00045
-const VOD_POLL_INTERVAL_MS = 5_000
-const VOD_POLL_MAX_ATTEMPTS = 40
 const DEFAULT_OUT_DIR = 'data/camara'

 // ---------------------------------------------------------------------------
@ -103,49 +105,10 @@ Opções:
  --help              esta ajuda
 `

-// ---------------------------------------------------------------------------
-// HTTP helpers
-// ---------------------------------------------------------------------------
-
-const camaraHeaders = { 'User-Agent': CAMARA_USER_AGENT, Accept: 'application/json' }
-
-const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms))
-
-async function fetchOnce(url, { timeoutMs = 30_000, headers = camaraHeaders } = {}) {
-  const response = await fetch(url, { headers, signal: AbortSignal.timeout(timeoutMs) })
-  if (!response.ok) throw new Error(`HTTP ${response.status} em ${url}`)
-  return response
-}
-
-/** The Câmara API drops connections now and then — bounded retry for GETs. */
-async function getText(url, { attempts = 3, ...options } = {}) {
-  let lastError
-  for (let attempt = 1; attempt <= attempts; attempt += 1) {
-    try {
-      return await (await fetchOnce(url, options)).text()
-    } catch (error) {
-      lastError = error
-      if (attempt < attempts) await sleep(2_000)
-    }
-  }
-  throw lastError
-}
-
-const getJson = async (url, options) => JSON.parse(await getText(url, options))
-
 // ---------------------------------------------------------------------------
 // Pipeline steps
 // ---------------------------------------------------------------------------

-const speechesUrl = (deputyId, from, to) =>
-  `https://dadosabertos.camara.leg.br/api/v2/deputados/${deputyId}/discursos` +
-  `?dataInicio=${from}&dataFim=${to}&itens=100&ordenarPor=dataHoraInicio&ordem=ASC`
-
-const eventsUrl = (date) =>
-  `https://dadosabertos.camara.leg.br/api/v2/eventos?dataInicio=${date}&dataFim=${date}&itens=100`
-
-const eventPageUrl = (eventId) => `https://www.camara.leg.br/evento-legislativo/${eventId}`
-
 const fileExists = async (path) => {
  try {
    await access(path)
@ -161,62 +124,6 @@ const resolveRange = (options) => {
  return { from, to }
 }

-/** Polls the async VOD generation until PRONTO (or gives up). */
-async function resolveVod(eventId, excerpt) {
-  const url = buildVodUrl(eventId, excerpt.audioId, excerpt.tMs)
-  let lastStatus = null
-  for (let attempt = 1; attempt <= VOD_POLL_MAX_ATTEMPTS; attempt += 1) {
-    const status = parseVodStatus(await getJson(url, { timeoutMs: 60_000 }))
-    lastStatus = status
-    if (status.state === 'PRONTO') return { url, status, attempts: attempt }
-    if (status.state === 'INDISPONIVEL') return { url, status, attempts: attempt }
-    await sleep(VOD_POLL_INTERVAL_MS)
-  }
-  return { url, status: lastStatus, attempts: VOD_POLL_MAX_ATTEMPTS }
-}
-
-const downloadToBuffer = async (url) => {
-  let lastError
-  for (let attempt = 1; attempt <= 3; attempt += 1) {
-    try {
-      const response = await fetchOnce(url, {
-        timeoutMs: 120_000,
-        headers: { 'User-Agent': CAMARA_USER_AGENT },
-      })
-      return Buffer.from(await response.arrayBuffer())
-    } catch (error) {
-      lastError = error
-      if (attempt < 3) await sleep(3_000)
-    }
-  }
-  throw lastError
-}
-
-async function transcribe(mp4Buffer) {
-  const apiKey = process.env.DEEPINFRA_API_KEY
-  if (!apiKey) die('DEEPINFRA_API_KEY ausente — configure no .env.local ou no env do processo.')
-  const form = new FormData()
-  form.append('file', new Blob([mp4Buffer], { type: 'video/mp4' }), 'trecho.mp4')
-  form.append('model', DEEPINFRA_WHISPER_MODEL)
-  form.append('language', 'pt')
-  form.append('response_format', 'verbose_json')
-  form.append('timestamp_granularities[]', 'segment')
-  const startedAt = Date.now()
-  const response = await fetch(DEEPINFRA_TRANSCRIBE_URL, {
-    method: 'POST',
-    headers: { Authorization: `Bearer ${apiKey}` },
-    body: form,
-    signal: AbortSignal.timeout(600_000),
-  })
-  if (!response.ok) {
-    throw new Error(`Deep Infra HTTP ${response.status}: ${(await response.text()).slice(0, 300)}`)
-  }
-  return {
-    transcription: normalizeTranscription(await response.json()),
-    elapsedMs: Date.now() - startedAt,
-  }
-}
-
 // ---------------------------------------------------------------------------
 // Per-speech orchestration
 // ---------------------------------------------------------------------------
@ -311,7 +218,7 @@ async function processSpeech(speech, options) {

    if (!options.skipTranscribe && report.mp4) {
      const buffer = await readFile(report.mp4.path)
-      const { transcription, elapsedMs } = await transcribe(buffer)
+      const { transcription, elapsedMs } = await transcribeSpeechAudio(buffer)
      const estimatedCostUsd = transcription.duration
        ? (transcription.duration / 60) * DEEPINFRA_COST_PER_MINUTE_USD
        : null
@ -396,6 +303,9 @@ async function main() {
    console.log(HELP)
    process.exit(0)
  }
+  if (!options.skipTranscribe && !process.env.DEEPINFRA_API_KEY) {
+    die('DEEPINFRA_API_KEY ausente — configure no .env.local ou no env do processo.')
+  }

  const { from, to } = resolveRange(options)
  console.log(
--- a/src/collections/CampaignUser.ts
+++ b/src/collections/CampaignUser.ts
@ -484,6 +484,7 @@ export const CampaignUser: CollectionConfig = {
        { label: 'Coordenador Geral', value: 'coordinator' },
        { label: 'Assessor', value: 'advisor' },
        { label: 'Candidato', value: 'candidate' },
+        { label: 'Assessor de Comunicação', value: 'communicator' },
        { label: 'Liderança', value: 'leader' },
      ],
      access: {
--- a/src/collections/Speech.ts
+++ b/src/collections/Speech.ts
@ -0,0 +1,310 @@
+import type {
+  CollectionBeforeChangeHook,
+  CollectionBeforeDeleteHook,
+  CollectionBeforeValidateHook,
+  CollectionConfig,
+} from 'payload'
+
+import { SPEECH_CLASSIFICATION_SOURCES, SPEECH_SCOPES, SPEECH_TOPICS } from '@/lib/speechFacets'
+import { canReadSpeech, canUpdateSpeech, payloadAdminOnly } from '@/utilities/campaignAccess'
+
+/**
+ * Speech catalog (C153) — one record per Jorge Solla speech in the Câmara,
+ * with the official transcript (taquigrafia), the ASR segments with timestamps
+ * (speechSegment), curated facets and extracted mentions.
+ *
+ * Data provenance: Câmara open data + VOD, CC BY 4.0 — keep the credit.
+ * Identity is the API speech (`sourceKey`), never the regenerable VOD URL.
+ */
+
+const FACET_FIELDS = [
+  'topics',
+  'scopes',
+  'classifiedBy',
+  'mentionedMunicipalities',
+  'mentionedPeople',
+  'mentionedPrograms',
+  'mentionedProjects',
+] as const
+
+const deriveSpeechYear: CollectionBeforeValidateHook = ({ data, originalDoc }) => {
+  if (!data) return data
+  const speechAt = typeof data.speechAt === 'string' ? data.speechAt : originalDoc?.speechAt
+  const year = Number(String(speechAt ?? '').slice(0, 4))
+  if (Number.isInteger(year) && year > 1900) data.year = year
+  return data
+}
+
+/**
+ * A manual facet correction is curation: unauthenticated writes (the import
+ * CLI) never overwrite it. Authenticated admin edits stay free to change both
+ * the facets and the provenance.
+ */
+const preserveManualFacets: CollectionBeforeChangeHook = ({
+  data,
+  operation,
+  originalDoc,
+  req,
+}) => {
+  if (!data || operation !== 'update' || originalDoc?.classifiedBy !== 'manual') return data
+  if (req.user) return data
+
+  const next = data as Record<string, unknown>
+  for (const field of FACET_FIELDS) {
+    if (field in next) next[field] = originalDoc[field]
+  }
+  return data
+}
+
+/** Segments are owned by the speech; Payload relationships do not cascade. */
+const deleteSpeechSegments: CollectionBeforeDeleteHook = async ({ id, req }) => {
+  await req.payload.delete({
+    collection: 'speechSegment',
+    where: { speech: { equals: id } },
+    req,
+    // Intentional bypass: the cascade is owned by this hook; segment access is
+    // admin-only and this runs inside the authorized speech delete request.
+    overrideAccess: true,
+  })
+}
+
+export const Speech: CollectionConfig = {
+  slug: 'speech',
+  labels: {
+    singular: 'Discurso',
+    plural: 'Discursos',
+  },
+  admin: {
+    group: 'Comunicação',
+    useAsTitle: 'speechAt',
+    defaultColumns: ['speechAt', 'type', 'phase', 'topics', 'classifiedBy'],
+    description:
+      'Acervo de falas do deputado na Câmara. Dados e vídeos da Câmara dos Deputados (CC BY 4.0).',
+  },
+  access: {
+    create: payloadAdminOnly,
+    read: canReadSpeech,
+    update: canUpdateSpeech,
+    delete: payloadAdminOnly,
+  },
+  hooks: {
+    beforeValidate: [deriveSpeechYear],
+    beforeChange: [preserveManualFacets],
+    beforeDelete: [deleteSpeechSegments],
+  },
+  fields: [
+    {
+      name: 'sourceKey',
+      type: 'text',
+      label: 'Chave da fonte',
+      required: true,
+      unique: true,
+      index: true,
+      admin: {
+        readOnly: true,
+        description: 'Identidade da fala na API da Câmara (data/hora + tipo + fase).',
+      },
+    },
+    {
+      // Wall-clock time of the source (Brasília local, no timezone). Kept as
+      // text on purpose: the Câmara API has no timezone and pretending it is an
+      // instant would shift old DST dates.
+      name: 'speechAt',
+      type: 'text',
+      label: 'Data e hora',
+      required: true,
+      index: true,
+      admin: {
+        readOnly: true,
+        description: 'Horário local de Brasília, como publicado pela Câmara.',
+      },
+    },
+    {
+      name: 'year',
+      type: 'number',
+      label: 'Ano',
+      index: true,
+      admin: { readOnly: true },
+    },
+    {
+      name: 'legislature',
+      type: 'select',
+      label: 'Legislatura',
+      index: true,
+      options: [
+        { label: '54ª', value: '54' },
+        { label: '55ª', value: '55' },
+        { label: '56ª', value: '56' },
+        { label: '57ª', value: '57' },
+      ],
+    },
+    {
+      name: 'type',
+      type: 'text',
+      label: 'Tipo de discurso',
+      admin: { readOnly: true },
+    },
+    {
+      name: 'phase',
+      type: 'text',
+      label: 'Fase da sessão',
+      index: true,
+      admin: { readOnly: true },
+    },
+    {
+      name: 'durationSeconds',
+      type: 'number',
+      label: 'Duração (s)',
+      admin: { readOnly: true },
+    },
+    {
+      name: 'summary',
+      type: 'textarea',
+      label: 'Sumário oficial',
+      admin: { readOnly: true },
+    },
+    {
+      name: 'officialTranscript',
+      type: 'textarea',
+      label: 'Transcrição oficial',
+      admin: { readOnly: true },
+    },
+    {
+      name: 'officialTextUrl',
+      type: 'text',
+      label: 'Texto oficial (Diário)',
+      admin: { readOnly: true },
+    },
+    {
+      name: 'keywords',
+      type: 'text',
+      label: 'Palavras-chave oficiais',
+      hasMany: true,
+      admin: {
+        readOnly: true,
+        description: 'Preservadas cruas; as facetas não as substituem.',
+      },
+    },
+    {
+      name: 'eventId',
+      type: 'number',
+      label: 'Evento',
+      index: true,
+      admin: { readOnly: true },
+    },
+    {
+      name: 'eventType',
+      type: 'text',
+      label: 'Tipo de sessão',
+      admin: { readOnly: true },
+    },
+    {
+      name: 'eventStartAt',
+      type: 'text',
+      label: 'Início da sessão',
+      admin: { readOnly: true },
+    },
+    {
+      name: 'eventEndAt',
+      type: 'text',
+      label: 'Fim da sessão',
+      admin: { readOnly: true },
+    },
+    {
+      name: 'youtubeUrl',
+      type: 'text',
+      label: 'YouTube da sessão',
+      admin: { readOnly: true },
+    },
+    {
+      name: 'presidingOfficer',
+      type: 'text',
+      label: 'Quem presidia',
+      admin: {
+        readOnly: true,
+        description: 'Derivado das trocas de mesa da sessão; vazio quando não há registro.',
+      },
+    },
+    {
+      name: 'audioId',
+      type: 'number',
+      label: 'Áudio',
+      index: true,
+      admin: { readOnly: true },
+    },
+    {
+      name: 'excerptTMs',
+      type: 'number',
+      label: 'Início do trecho (ms)',
+      index: true,
+      admin: { readOnly: true },
+    },
+    {
+      name: 'vodPlaybackUrl',
+      type: 'text',
+      label: 'VOD (reproduzir)',
+      admin: { readOnly: true, description: 'Último link conhecido; regerável pelo VOD.' },
+    },
+    {
+      name: 'vodDownloadUrl',
+      type: 'text',
+      label: 'VOD (baixar)',
+      admin: { readOnly: true, description: 'Último link conhecido; regerável pelo VOD.' },
+    },
+    {
+      name: 'topics',
+      type: 'select',
+      label: 'Temas',
+      hasMany: true,
+      index: true,
+      options: SPEECH_TOPICS.map(({ value, label }) => ({ value, label })),
+    },
+    {
+      name: 'scopes',
+      type: 'select',
+      label: 'Alcance',
+      hasMany: true,
+      index: true,
+      options: SPEECH_SCOPES.map(({ value, label }) => ({ value, label })),
+    },
+    {
+      name: 'classifiedBy',
+      type: 'select',
+      label: 'Proveniência da classificação',
+      required: true,
+      defaultValue: 'gazetteer',
+      index: true,
+      options: SPEECH_CLASSIFICATION_SOURCES.map((value) => ({
+        value,
+        label: value === 'gazetteer' ? 'Gazetteer' : value === 'llm' ? 'LLM' : 'Manual',
+      })),
+    },
+    {
+      name: 'mentionedMunicipalities',
+      type: 'relationship',
+      relationTo: 'municipality',
+      label: 'Municípios citados',
+      hasMany: true,
+      index: true,
+    },
+    {
+      name: 'mentionedPeople',
+      type: 'text',
+      label: 'Pessoas citadas',
+      hasMany: true,
+      admin: { description: 'Menções textuais — não é cadastro de contatos.' },
+    },
+    {
+      name: 'mentionedPrograms',
+      type: 'text',
+      label: 'Programas citados',
+      hasMany: true,
+    },
+    {
+      name: 'mentionedProjects',
+      type: 'text',
+      label: 'Projetos citados',
+      hasMany: true,
+    },
+  ],
+}
--- a/src/collections/SpeechSegment.ts
+++ b/src/collections/SpeechSegment.ts
@ -0,0 +1,84 @@
+import type { CollectionBeforeValidateHook, CollectionConfig } from 'payload'
+
+import { normalizeForSearch } from '@/lib/speechSearch'
+import { canReadSpeech, payloadAdminOnly } from '@/utilities/campaignAccess'
+
+/**
+ * Searchable ASR segment of a speech (C153). The official transcript is the
+ * reading surface; these segments carry the timestamps and the trigram-search
+ * index (`speech_segment_search_text_trgm_idx`, hand-written migration) used
+ * by the C154 search. Admin-hidden: the product surface is the catalog UI.
+ */
+
+const deriveSearchText: CollectionBeforeValidateHook = ({ data, originalDoc }) => {
+  if (!data) return data
+  const text = typeof data.text === 'string' ? data.text : originalDoc?.text
+  data.searchText = normalizeForSearch(text ?? '')
+  return data
+}
+
+export const SpeechSegment: CollectionConfig = {
+  slug: 'speechSegment',
+  labels: {
+    singular: 'Segmento de fala',
+    plural: 'Segmentos de fala',
+  },
+  admin: {
+    group: 'Comunicação',
+    hidden: () => true,
+    useAsTitle: 'text',
+  },
+  access: {
+    create: payloadAdminOnly,
+    read: canReadSpeech,
+    update: payloadAdminOnly,
+    delete: payloadAdminOnly,
+  },
+  hooks: {
+    beforeValidate: [deriveSearchText],
+  },
+  fields: [
+    {
+      name: 'speech',
+      type: 'relationship',
+      relationTo: 'speech',
+      label: 'Discurso',
+      required: true,
+      index: true,
+    },
+    {
+      name: 'order',
+      type: 'number',
+      label: 'Ordem',
+      required: true,
+    },
+    {
+      name: 'startSeconds',
+      type: 'number',
+      label: 'Início (s)',
+      required: true,
+    },
+    {
+      name: 'endSeconds',
+      type: 'number',
+      label: 'Fim (s)',
+      required: true,
+    },
+    {
+      name: 'text',
+      type: 'textarea',
+      label: 'Texto',
+      required: true,
+    },
+    {
+      name: 'searchText',
+      type: 'text',
+      label: 'Texto normalizado (busca)',
+      required: true,
+      admin: {
+        readOnly: true,
+        description: 'Derivado do texto (sem acentos, minúsculas) para a busca por palavra.',
+      },
+    },
+  ],
+}
--- a/src/components/campaign/shell/nav.ts
+++ b/src/components/campaign/shell/nav.ts
@ -84,6 +84,9 @@ const leaderNav: CampaignNavItem[] = [

 export const getCampaignNav = (role: CampaignUser['role']): CampaignNavItem[] => {
  if (role === 'leader') return leaderNav
+  // C153 — the communication assessor is not staff: no staff destinations.
+  // The catalog vertical (C154) adds its own nav items here.
+  if (role === 'communicator') return []

  return staffNav.filter((item) => {
    if (item.href === '/campanha/apoiadores') return canAccessSupporterArea(role)
--- a/src/lib/campaignRoles.ts
+++ b/src/lib/campaignRoles.ts
@ -16,3 +16,11 @@ export const isStaffCampaignRole = (role: CampaignRole): boolean =>
 /** Coordinator or candidate — unrestricted scope (all municipalities, decisions). */
 export const isUnrestrictedCampaignRole = (role: CampaignRole): boolean =>
  role === 'coordinator' || role === 'candidate'
+
+/**
+ * Speech catalog readers (C153): the communication assessor plus the
+ * unrestricted roles. Deliberately NOT part of `isStaffCampaignRole` — the
+ * communicator does not inherit any staff area, only the catalog vertical.
+ */
+export const canReadSpeechCatalog = (role: CampaignRole): boolean =>
+  role === 'communicator' || isUnrestrictedCampaignRole(role)
--- a/src/lib/speechFacets.ts
+++ b/src/lib/speechFacets.ts
@ -0,0 +1,166 @@
+import { z } from 'zod'
+
+import type { MunicipalityCatalogEntry } from '@/lib/municipalityCatalog'
+import { normalizeForSearch, uniqueByNormalizedForm } from '@/lib/speechSearch'
+
+/**
+ * Facet contract for the speech catalog (C153). The topic and scope values are
+ * persisted data (facet filters in C154); labels are the pt-BR product names.
+ * The offline gazetteer lives in `speechGazetteer.ts` and the LLM refinement in
+ * `src/utilities/speech/speechClassifier.ts`; this module validates the LLM
+ * response and merges both provenances. `manual` edits win over both.
+ */
+
+export const SPEECH_TOPICS = [
+  { value: 'saude', label: 'Saúde' },
+  { value: 'educacao', label: 'Educação' },
+  { value: 'cultura', label: 'Cultura' },
+  { value: 'esporte', label: 'Esporte' },
+  { value: 'seguranca-publica', label: 'Segurança Pública' },
+  { value: 'meio-ambiente', label: 'Meio Ambiente' },
+  { value: 'economia-trabalho', label: 'Economia e Trabalho' },
+  { value: 'direitos-humanos', label: 'Direitos Humanos e Assistência Social' },
+  { value: 'infraestrutura', label: 'Infraestrutura e Transporte' },
+  { value: 'ciencia-tecnologia', label: 'Ciência e Tecnologia' },
+  { value: 'politica-instituicoes', label: 'Política e Instituições' },
+  { value: 'agricultura', label: 'Agricultura e Agropecuária' },
+  { value: 'habitacao-cidades', label: 'Habitação e Cidades' },
+  { value: 'comunicacao-midia', label: 'Comunicação e Mídia' },
+  { value: 'igualdade-racial', label: 'Igualdade Racial' },
+  { value: 'mulheres-genero', label: 'Mulheres e Gênero' },
+  { value: 'juventude', label: 'Juventude' },
+  { value: 'pessoa-deficiencia', label: 'Pessoa com Deficiência' },
+] as const
+
+export type SpeechTopic = (typeof SPEECH_TOPICS)[number]['value']
+
+export const SPEECH_SCOPES = [
+  { value: 'bahia', label: 'Bahia' },
+  { value: 'brasil', label: 'Brasil' },
+  { value: 'internacional', label: 'Internacional' },
+] as const
+
+export type SpeechScope = (typeof SPEECH_SCOPES)[number]['value']
+
+export const SPEECH_CLASSIFICATION_SOURCES = ['gazetteer', 'llm', 'manual'] as const
+
+type SpeechClassificationSource = (typeof SPEECH_CLASSIFICATION_SOURCES)[number]
+
+export type SpeechFacetClassification = {
+  topics: SpeechTopic[]
+  scopes: SpeechScope[]
+  municipalities: MunicipalityCatalogEntry[]
+  people: string[]
+  programs: string[]
+  projects: string[]
+}
+
+export type MergedSpeechFacets = SpeechFacetClassification & {
+  classifiedBy: SpeechClassificationSource
+}
+
+const facetKey = (value: string): string => normalizeForSearch(value).replace(/-/g, ' ')
+
+const topicByKey = new Map(
+  SPEECH_TOPICS.flatMap(({ value, label }) => [
+    [facetKey(value), value],
+    [facetKey(label), value],
+  ]),
+)
+
+const scopeByKey = new Map(
+  SPEECH_SCOPES.flatMap(({ value, label }) => [
+    [facetKey(value), value],
+    [facetKey(label), value],
+  ]),
+)
+
+// ---------------------------------------------------------------------------
+// LLM response parsing and merge
+// ---------------------------------------------------------------------------
+
+const MAX_MENTION_ITEMS = 12
+const MAX_MENTION_LENGTH = 120
+
+const llmFacetResponseSchema = z.object({
+  topics: z.array(z.unknown()).optional(),
+  scopes: z.array(z.unknown()).optional(),
+  people: z.array(z.unknown()).optional(),
+  programs: z.array(z.unknown()).optional(),
+  projects: z.array(z.unknown()).optional(),
+})
+
+export type SpeechLlmFacets = {
+  topics: SpeechTopic[]
+  scopes: SpeechScope[]
+  people: string[]
+  programs: string[]
+  projects: string[]
+}
+
+const stringMentions = (value: unknown): string[] => {
+  if (!Array.isArray(value)) return []
+  return uniqueByNormalizedForm(
+    value
+      .filter((item): item is string => typeof item === 'string')
+      .map((item) => item.trim().slice(0, MAX_MENTION_LENGTH)),
+  ).slice(0, MAX_MENTION_ITEMS)
+}
+
+/**
+ * Validates the LLM JSON response against the known taxonomy. Unknown topics
+ * and scopes are dropped (never persisted as free text); an object without any
+ * known key returns null so the caller falls back to the gazetteer provenance.
+ */
+export const parseLlmFacetResponse = (json: unknown): SpeechLlmFacets | null => {
+  const parsed = llmFacetResponseSchema.safeParse(json)
+  if (!parsed.success || Object.keys(parsed.data).length === 0) return null
+  const data = parsed.data
+
+  const topics = uniqueByNormalizedForm(
+    (Array.isArray(data.topics) ? data.topics : []).filter(
+      (item): item is string => typeof item === 'string',
+    ),
+  )
+    .map((item) => topicByKey.get(facetKey(item)))
+    .filter((value): value is SpeechTopic => value !== undefined)
+
+  const scopes = uniqueByNormalizedForm(
+    (Array.isArray(data.scopes) ? data.scopes : []).filter(
+      (item): item is string => typeof item === 'string',
+    ),
+  )
+    .map((item) => scopeByKey.get(facetKey(item)))
+    .filter((value): value is SpeechScope => value !== undefined)
+
+  return {
+    topics: [...new Set(topics)],
+    scopes: [...new Set(scopes)],
+    people: stringMentions(data.people),
+    programs: stringMentions(data.programs),
+    projects: stringMentions(data.projects),
+  }
+}
+
+/**
+ * Gazetteer ∪ LLM. Topics are unioned (the lexicon is evidence too); scopes
+ * prefer the validated LLM output when it produced any; people/programs/
+ * projects come from the LLM only. Provenance is `llm` whenever a valid
+ * response refined the result, `gazetteer` otherwise.
+ */
+export const mergeFacetClassification = (
+  gazetteer: SpeechFacetClassification,
+  llm: SpeechLlmFacets | null,
+): MergedSpeechFacets => {
+  if (!llm) return { ...gazetteer, classifiedBy: 'gazetteer' }
+
+  return {
+    topics: [...new Set([...gazetteer.topics, ...llm.topics])],
+    scopes: llm.scopes.length > 0 ? llm.scopes : gazetteer.scopes,
+    municipalities: gazetteer.municipalities,
+    people: llm.people,
+    programs: llm.programs,
+    projects: llm.projects,
+    classifiedBy: 'llm',
+  }
+}
--- a/src/lib/speechGazetteer.ts
+++ b/src/lib/speechGazetteer.ts
@ -0,0 +1,443 @@
+import { municipalityCatalog, type MunicipalityCatalogEntry } from '@/lib/municipalityCatalog'
+import {
+  SPEECH_SCOPES,
+  SPEECH_TOPICS,
+  type SpeechFacetClassification,
+  type SpeechScope,
+  type SpeechTopic,
+} from '@/lib/speechFacets'
+import { normalizeForSearch } from '@/lib/speechSearch'
+
+/**
+ * Offline gazetteer pass of the speech facet classifier (C153): topic/scope
+ * lexicons plus municipality mentions. Deliberately a separate module from the
+ * `speechFacets` contract so the C154 client bundle can import the topic
+ * labels without pulling the lexicon and the municipality catalog.
+ */
+
+/** Topic keyword lexicon — terms are compared against normalized text. */
+const SPEECH_TOPIC_TERMS: Record<SpeechTopic, readonly string[]> = {
+  saude: [
+    'saude',
+    'sus',
+    'hospital',
+    'posto de saude',
+    'medicamento',
+    'farmacia popular',
+    'vacina',
+    'vacinacao',
+    'medico',
+    'enfermagem',
+    'atencao basica',
+    'ministerio da saude',
+    'agentes comunitarios',
+  ],
+  educacao: [
+    'educacao',
+    'escola',
+    'professor',
+    'professora',
+    'estudante',
+    'aluno',
+    'universidade',
+    'ensino',
+    'alfabetizacao',
+    'creche',
+    'merenda',
+    'pne',
+    'ifba',
+    'ufba',
+    'enem',
+  ],
+  cultura: [
+    'cultura',
+    'cultural',
+    'artista',
+    'teatro',
+    'cinema',
+    'musica',
+    'patrimonio',
+    'lei rouanet',
+    'carnaval',
+    'museu',
+    'livro',
+    'leitura',
+    'audiovisual',
+  ],
+  esporte: ['esporte', 'esportivo', 'futebol', 'atleta', 'olimpiada', 'campeonato'],
+  'seguranca-publica': [
+    'seguranca publica',
+    'seguranca',
+    'policia',
+    'policial',
+    'violencia',
+    'crime',
+    'criminalidade',
+    'trafico',
+    'armas',
+    'milicia',
+  ],
+  'meio-ambiente': [
+    'meio ambiente',
+    'ambiental',
+    'desmatamento',
+    'garimpo',
+    'queimada',
+    'clima',
+    'climatico',
+    'sustentabilidade',
+    'reciclagem',
+    'poluicao',
+    'floresta',
+    'amazonia',
+  ],
+  'economia-trabalho': [
+    'economia',
+    'emprego',
+    'desemprego',
+    'trabalhador',
+    'trabalhadora',
+    'salario',
+    'salario minimo',
+    'renda',
+    'industria',
+    'comercio',
+    'inflacao',
+    'juros',
+    'tributacao',
+    'reforma tributaria',
+  ],
+  'direitos-humanos': [
+    'direitos humanos',
+    'assistencia social',
+    'bolsa familia',
+    'cras',
+    'creas',
+    'pobreza',
+    'miseria',
+    'desigualdade',
+    'populacao de rua',
+    'idoso',
+    'seguridade',
+  ],
+  infraestrutura: [
+    'infraestrutura',
+    'transporte',
+    'rodovia',
+    'estrada',
+    'saneamento',
+    'mobilidade',
+    'metro',
+    'ferrovia',
+    'porto',
+    'aeroporto',
+    'energia eletrica',
+    'luz para todos',
+    'pavimentacao',
+  ],
+  'ciencia-tecnologia': [
+    'ciencia',
+    'cientifico',
+    'tecnologia',
+    'pesquisa',
+    'inovacao',
+    'embrapa',
+    'internet',
+    'banda larga',
+    'inteligencia artificial',
+  ],
+  'politica-instituicoes': [
+    'politica',
+    'politico',
+    'governo',
+    'congresso',
+    'camara dos deputados',
+    'senado',
+    'eleicao',
+    'eleitor',
+    'democracia',
+    'instituicoes',
+    'constituicao',
+    'impeachment',
+    'mandato',
+    'deputado',
+    'deputada',
+    'presidente da republica',
+    'ministro',
+    'golpe',
+    'ditadura',
+  ],
+  agricultura: [
+    'agricultura',
+    'agropecuaria',
+    'agricultor',
+    'agricultora',
+    'agricultura familiar',
+    'pecuaria',
+    'lavoura',
+    'safra',
+    'reforma agraria',
+    'rural',
+    'irrigacao',
+    'seca',
+    'semiarido',
+  ],
+  'habitacao-cidades': [
+    'habitacao',
+    'moradia',
+    'minha casa minha vida',
+    'cidade',
+    'urbano',
+    'urbana',
+    'favela',
+    'periferia',
+    'regularizacao fundiaria',
+    'aluguel',
+  ],
+  'comunicacao-midia': [
+    'comunicacao',
+    'midia',
+    'imprensa',
+    'jornal',
+    'jornalismo',
+    'radio',
+    'televisao',
+    'redes sociais',
+    'fake news',
+    'desinformacao',
+  ],
+  'igualdade-racial': [
+    'igualdade racial',
+    'racial',
+    'racismo',
+    'racista',
+    'negro',
+    'negra',
+    'quilombola',
+    'quilombo',
+    'consciencia negra',
+    'cotas',
+  ],
+  'mulheres-genero': [
+    'mulher',
+    'mulheres',
+    'genero',
+    'feminicidio',
+    'violencia domestica',
+    'violencia contra a mulher',
+    'igualdade de genero',
+    'maternidade',
+    'machismo',
+  ],
+  juventude: [
+    'juventude',
+    'jovem',
+    'jovens',
+    'adolescente',
+    'crianca',
+    'primeira infancia',
+    'juventude rural',
+  ],
+  'pessoa-deficiencia': [
+    'pessoa com deficiencia',
+    'deficiencia',
+    'deficiente',
+    'acessibilidade',
+    'autismo',
+    'autista',
+    'cadeirante',
+    'pcd',
+    'libras',
+  ],
+}
+
+const SPEECH_SCOPE_TERMS: Record<SpeechScope, readonly string[]> = {
+  bahia: ['bahia', 'baiano', 'baiana', 'baianos', 'baianas'],
+  brasil: [
+    'brasil',
+    'brasileiro',
+    'brasileira',
+    'brasileiros',
+    'brasileiras',
+    'nacional',
+    'pais',
+    'governo federal',
+    'uniao',
+    'republica',
+  ],
+  internacional: [
+    'internacional',
+    'mundo',
+    'mundial',
+    'exterior',
+    'onu',
+    'mercosul',
+    'otan',
+    'uniao europeia',
+    'china',
+    'estados unidos',
+    'eua',
+    'europa',
+    'africa',
+    'america latina',
+    'guerra',
+    'refugiados',
+    'embaixada',
+    'tratado',
+  ],
+}
+
+const escapeRegExp = (value: string): string => value.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')
+
+const containsTerm = (normalizedText: string, term: string): boolean =>
+  new RegExp(`(^|[^a-z0-9])${escapeRegExp(term)}($|[^a-z0-9])`).test(normalizedText)
+
+// ---------------------------------------------------------------------------
+// Municipality mentions
+// ---------------------------------------------------------------------------
+
+/**
+ * NFD-stripped, whitespace-collapsed but CASE-PRESERVING copy, index-aligned
+ * with `normalizeForSearch` (lowercasing does not change length in pt-BR).
+ * Used to reject lowercase common words that are also municipality names
+ * ("saúde" the noun vs. Saúde (BA), "central" the adjective vs. Central).
+ */
+const stripToComparable = (value: string): string =>
+  value
+    .normalize('NFD')
+    .replace(/[\u0300-\u036f]/g, '')
+    .replace(/\s+/g, ' ')
+    .trim()
+
+type MunicipalityMatcher = { normalized: string; entries: MunicipalityCatalogEntry[] }
+
+let municipalityMatchers: readonly MunicipalityMatcher[] | null = null
+
+const getMunicipalityMatchers = (): readonly MunicipalityMatcher[] => {
+  if (municipalityMatchers) return municipalityMatchers
+  const byCity = new Map<string, MunicipalityCatalogEntry[]>()
+  for (const entry of municipalityCatalog) {
+    const list = byCity.get(entry.city)
+    if (list) list.push(entry)
+    else byCity.set(entry.city, [entry])
+  }
+  municipalityMatchers = [...byCity.entries()]
+    .map(([city, entries]) => ({ normalized: normalizeForSearch(city), entries }))
+    .sort((left, right) => right.normalized.length - left.normalized.length)
+  return municipalityMatchers
+}
+
+/**
+ * Institutional phrases that contain a municipality name but are not a
+ * mention of the place (Saúde (BA) inside "Sistema Único de Saúde", Central
+ * (BA) inside "Banco Central", Planalto (BA) inside "Palácio do Planalto").
+ * Masked (same length) before matching so the capitalization heuristic does
+ * not turn them into false positives.
+ */
+const MUNICIPALITY_FALSE_POSITIVE_PHRASES: readonly string[] = [
+  'sistema unico de saude',
+  'ministerio da saude',
+  'secretaria de saude',
+  'secretaria municipal de saude',
+  'conselho de saude',
+  'conselho municipal de saude',
+  'saude publica',
+  'saude bucal',
+  'atencao basica',
+  'profissionais de saude',
+  'plano de saude',
+  'palacio do planalto',
+  'planalto central',
+  'banco central',
+  'central unica',
+  'central sindical',
+  'central de',
+]
+
+const maskFalsePositivePhrases = (value: string): string => {
+  let masked = value
+  for (const phrase of MUNICIPALITY_FALSE_POSITIVE_PHRASES) {
+    masked = masked.replace(new RegExp(escapeRegExp(phrase), 'gi'), (match) =>
+      ' '.repeat(match.length),
+    )
+  }
+  return masked
+}
+
+/**
+ * Matches catalog municipality names in the text. Longest names are consumed
+ * first so "São Félix do Coribe" never also yields "Coribe"; a match must
+ * start with an uppercase letter in the original text, so prose "saúde" does
+ * not become the municipality Saúde. Salvador has no city row — a mention of
+ * the city maps to its 19 zone entries (documented product decision).
+ */
+export const matchMunicipalityMentions = (text: string): MunicipalityCatalogEntry[] => {
+  const comparable = maskFalsePositivePhrases(stripToComparable(String(text ?? '')))
+  if (comparable === '') return []
+  const masked = comparable.toLowerCase().split('')
+  const matched: MunicipalityCatalogEntry[] = []
+  const seen = new Set<string>()
+
+  for (const matcher of getMunicipalityMatchers()) {
+    const pattern = new RegExp(`(^|[^a-z0-9])${escapeRegExp(matcher.normalized)}($|[^a-z0-9])`, 'g')
+    const haystack = masked.join('')
+    let found = false
+    let match: RegExpExecArray | null
+    while ((match = pattern.exec(haystack)) !== null) {
+      const start = match.index + match[1].length
+      if (!/[A-Z]/.test(comparable[start] ?? '')) continue
+      found = true
+      for (let index = start; index < start + matcher.normalized.length; index += 1) {
+        masked[index] = ' '
+      }
+    }
+    if (!found) continue
+    for (const entry of matcher.entries) {
+      if (seen.has(entry.slug)) continue
+      seen.add(entry.slug)
+      matched.push(entry)
+    }
+  }
+
+  return matched
+}
+
+// ---------------------------------------------------------------------------
+// Classification
+// ---------------------------------------------------------------------------
+
+export type SpeechFacetInput = {
+  transcript: string
+  summary?: string | null
+  keywords?: readonly string[]
+}
+
+/**
+ * Offline facet pass: topic lexicon + scope lexicon + municipality gazetteer.
+ * People/programs/projects require the LLM pass; this pass never invents them.
+ */
+export const classifySpeechByGazetteer = (input: SpeechFacetInput): SpeechFacetClassification => {
+  const text = [input.summary ?? '', input.keywords?.join(' ') ?? '', input.transcript].join(' ')
+  const normalized = normalizeForSearch(text)
+
+  const topics = SPEECH_TOPICS.filter(({ value }) =>
+    SPEECH_TOPIC_TERMS[value].some((term) => containsTerm(normalized, term)),
+  ).map(({ value }) => value)
+
+  const municipalities = matchMunicipalityMentions(text)
+  const matchedScopes = new Set<SpeechScope>(
+    SPEECH_SCOPES.filter(({ value }) =>
+      SPEECH_SCOPE_TERMS[value].some((term) => containsTerm(normalized, term)),
+    ).map(({ value }) => value),
+  )
+  if (municipalities.length > 0) matchedScopes.add('bahia')
+
+  return {
+    topics,
+    scopes: SPEECH_SCOPES.filter(({ value }) => matchedScopes.has(value)).map(({ value }) => value),
+    municipalities,
+    people: [],
+    programs: [],
+    projects: [],
+  }
+}
--- a/src/lib/speechSearch.ts
+++ b/src/lib/speechSearch.ts
@ -0,0 +1,31 @@
+/**
+ * Pure text normalization for the speech catalog (C153). The segment hook
+ * stores `normalizeForSearch(text)` in `speechSegment.searchText` and the
+ * search UI (C154) normalizes the query with the SAME function before the
+ * `contains` query, so accent/case variants match the trigram index.
+ */
+
+/** NFD-strips accents, lowercases and collapses whitespace. */
+export const normalizeForSearch = (value: string): string =>
+  value
+    .normalize('NFD')
+    .replace(/[\u0300-\u036f]/g, '')
+    .toLowerCase()
+    .replace(/\s+/g, ' ')
+    .trim()
+
+/**
+ * Dedupes values by their normalized form, preserving the first spelling.
+ * Used by the facet classifier before persisting people/program mentions.
+ */
+export const uniqueByNormalizedForm = (values: readonly string[]): string[] => {
+  const seen = new Set<string>()
+  const result: string[] = []
+  for (const value of values) {
+    const key = normalizeForSearch(value)
+    if (key === '' || seen.has(key)) continue
+    seen.add(key)
+    result.push(value.trim())
+  }
+  return result
+}
--- a/src/migrations/20260913_001112_add_speech_catalog.json
+++ b/src/migrations/20260913_001112_add_speech_catalog.json
--- a/src/migrations/20260913_001112_add_speech_catalog.ts
+++ b/src/migrations/20260913_001112_add_speech_catalog.ts
@ -0,0 +1,158 @@
+import { MigrateUpArgs, MigrateDownArgs, sql } from '@payloadcms/db-postgres'
+
+/**
+ * C153 — speech catalog tables + the `communicator` campaign role.
+ *
+ * The generated migration also re-emitted
+ * `DROP INDEX google_calendar_sync_oauth_connected_at_idx`, a snapshot drift
+ * from C149 (the index appears in that migration's JSON but its SQL never
+ * created it, so the statement would fail on every environment); it was
+ * removed by hand.
+ */
+export async function up({ db, payload, req }: MigrateUpArgs): Promise<void> {
+  await db.execute(sql`
+   CREATE TYPE "public"."enum_speech_topics" AS ENUM('saude', 'educacao', 'cultura', 'esporte', 'seguranca-publica', 'meio-ambiente', 'economia-trabalho', 'direitos-humanos', 'infraestrutura', 'ciencia-tecnologia', 'politica-instituicoes', 'agricultura', 'habitacao-cidades', 'comunicacao-midia', 'igualdade-racial', 'mulheres-genero', 'juventude', 'pessoa-deficiencia');
+  CREATE TYPE "public"."enum_speech_scopes" AS ENUM('bahia', 'brasil', 'internacional');
+  CREATE TYPE "public"."enum_speech_legislature" AS ENUM('54', '55', '56', '57');
+  CREATE TYPE "public"."enum_speech_classified_by" AS ENUM('gazetteer', 'llm', 'manual');
+  ALTER TYPE "public"."enum_campaign_user_role" ADD VALUE 'communicator' BEFORE 'leader';
+  CREATE TABLE "speech_topics" (
+  	"order" integer NOT NULL,
+  	"parent_id" integer NOT NULL,
+  	"value" "enum_speech_topics",
+  	"id" serial PRIMARY KEY NOT NULL
+  );
+  
+  CREATE TABLE "speech_scopes" (
+  	"order" integer NOT NULL,
+  	"parent_id" integer NOT NULL,
+  	"value" "enum_speech_scopes",
+  	"id" serial PRIMARY KEY NOT NULL
+  );
+  
+  CREATE TABLE "speech" (
+  	"id" serial PRIMARY KEY NOT NULL,
+  	"source_key" varchar NOT NULL,
+  	"speech_at" varchar NOT NULL,
+  	"year" numeric,
+  	"legislature" "enum_speech_legislature",
+  	"type" varchar,
+  	"phase" varchar,
+  	"duration_seconds" numeric,
+  	"summary" varchar,
+  	"official_transcript" varchar,
+  	"official_text_url" varchar,
+  	"event_id" numeric,
+  	"event_type" varchar,
+  	"event_start_at" varchar,
+  	"event_end_at" varchar,
+  	"youtube_url" varchar,
+  	"presiding_officer" varchar,
+  	"audio_id" numeric,
+  	"excerpt_t_ms" numeric,
+  	"vod_playback_url" varchar,
+  	"vod_download_url" varchar,
+  	"classified_by" "enum_speech_classified_by" DEFAULT 'gazetteer' NOT NULL,
+  	"updated_at" timestamp(3) with time zone DEFAULT now() NOT NULL,
+  	"created_at" timestamp(3) with time zone DEFAULT now() NOT NULL
+  );
+  
+  CREATE TABLE "speech_texts" (
+  	"id" serial PRIMARY KEY NOT NULL,
+  	"order" integer NOT NULL,
+  	"parent_id" integer NOT NULL,
+  	"path" varchar NOT NULL,
+  	"text" varchar
+  );
+  
+  CREATE TABLE "speech_rels" (
+  	"id" serial PRIMARY KEY NOT NULL,
+  	"order" integer,
+  	"parent_id" integer NOT NULL,
+  	"path" varchar NOT NULL,
+  	"municipality_id" integer
+  );
+  
+  CREATE TABLE "speech_segment" (
+  	"id" serial PRIMARY KEY NOT NULL,
+  	"speech_id" integer NOT NULL,
+  	"order" numeric NOT NULL,
+  	"start_seconds" numeric NOT NULL,
+  	"end_seconds" numeric NOT NULL,
+  	"text" varchar NOT NULL,
+  	"search_text" varchar NOT NULL,
+  	"updated_at" timestamp(3) with time zone DEFAULT now() NOT NULL,
+  	"created_at" timestamp(3) with time zone DEFAULT now() NOT NULL
+  );
+  
+  ALTER TABLE "payload_locked_documents_rels" ADD COLUMN "speech_id" integer;  ALTER TABLE "payload_locked_documents_rels" ADD COLUMN "speech_segment_id" integer;
+  ALTER TABLE "speech_topics" ADD CONSTRAINT "speech_topics_parent_fk" FOREIGN KEY ("parent_id") REFERENCES "public"."speech"("id") ON DELETE cascade ON UPDATE no action;
+  ALTER TABLE "speech_scopes" ADD CONSTRAINT "speech_scopes_parent_fk" FOREIGN KEY ("parent_id") REFERENCES "public"."speech"("id") ON DELETE cascade ON UPDATE no action;
+  ALTER TABLE "speech_texts" ADD CONSTRAINT "speech_texts_parent_fk" FOREIGN KEY ("parent_id") REFERENCES "public"."speech"("id") ON DELETE cascade ON UPDATE no action;
+  ALTER TABLE "speech_rels" ADD CONSTRAINT "speech_rels_parent_fk" FOREIGN KEY ("parent_id") REFERENCES "public"."speech"("id") ON DELETE cascade ON UPDATE no action;
+  ALTER TABLE "speech_rels" ADD CONSTRAINT "speech_rels_municipality_fk" FOREIGN KEY ("municipality_id") REFERENCES "public"."municipality"("id") ON DELETE cascade ON UPDATE no action;
+  ALTER TABLE "speech_segment" ADD CONSTRAINT "speech_segment_speech_id_speech_id_fk" FOREIGN KEY ("speech_id") REFERENCES "public"."speech"("id") ON DELETE set null ON UPDATE no action;
+  CREATE INDEX "speech_topics_order_idx" ON "speech_topics" USING btree ("order");
+  CREATE INDEX "speech_topics_parent_idx" ON "speech_topics" USING btree ("parent_id");
+  CREATE INDEX "speech_topics_value_idx" ON "speech_topics" USING btree ("value");
+  CREATE INDEX "speech_scopes_order_idx" ON "speech_scopes" USING btree ("order");
+  CREATE INDEX "speech_scopes_parent_idx" ON "speech_scopes" USING btree ("parent_id");
+  CREATE INDEX "speech_scopes_value_idx" ON "speech_scopes" USING btree ("value");
+  CREATE UNIQUE INDEX "speech_source_key_idx" ON "speech" USING btree ("source_key");
+  CREATE INDEX "speech_speech_at_idx" ON "speech" USING btree ("speech_at");
+  CREATE INDEX "speech_year_idx" ON "speech" USING btree ("year");
+  CREATE INDEX "speech_legislature_idx" ON "speech" USING btree ("legislature");
+  CREATE INDEX "speech_phase_idx" ON "speech" USING btree ("phase");
+  CREATE INDEX "speech_event_id_idx" ON "speech" USING btree ("event_id");
+  CREATE INDEX "speech_audio_id_idx" ON "speech" USING btree ("audio_id");
+  CREATE INDEX "speech_excerpt_t_ms_idx" ON "speech" USING btree ("excerpt_t_ms");
+  CREATE INDEX "speech_classified_by_idx" ON "speech" USING btree ("classified_by");
+  CREATE INDEX "speech_updated_at_idx" ON "speech" USING btree ("updated_at");
+  CREATE INDEX "speech_created_at_idx" ON "speech" USING btree ("created_at");
+  CREATE INDEX "speech_texts_order_parent" ON "speech_texts" USING btree ("order","parent_id");
+  CREATE INDEX "speech_rels_order_idx" ON "speech_rels" USING btree ("order");
+  CREATE INDEX "speech_rels_parent_idx" ON "speech_rels" USING btree ("parent_id");
+  CREATE INDEX "speech_rels_path_idx" ON "speech_rels" USING btree ("path");
+  CREATE INDEX "speech_rels_municipality_id_idx" ON "speech_rels" USING btree ("municipality_id");
+  CREATE INDEX "speech_segment_speech_idx" ON "speech_segment" USING btree ("speech_id");
+  CREATE INDEX "speech_segment_updated_at_idx" ON "speech_segment" USING btree ("updated_at");
+  CREATE INDEX "speech_segment_created_at_idx" ON "speech_segment" USING btree ("created_at");
+  ALTER TABLE "payload_locked_documents_rels" ADD CONSTRAINT "payload_locked_documents_rels_speech_fk" FOREIGN KEY ("speech_id") REFERENCES "public"."speech"("id") ON DELETE cascade ON UPDATE no action;
+  ALTER TABLE "payload_locked_documents_rels" ADD CONSTRAINT "payload_locked_documents_rels_speech_segment_fk" FOREIGN KEY ("speech_segment_id") REFERENCES "public"."speech_segment"("id") ON DELETE cascade ON UPDATE no action;
+  CREATE INDEX "payload_locked_documents_rels_speech_id_idx" ON "payload_locked_documents_rels" USING btree ("speech_id");
+  CREATE INDEX "payload_locked_documents_rels_speech_segment_id_idx" ON "payload_locked_documents_rels" USING btree ("speech_segment_id");`)
+}
+
+export async function down({ db, payload, req }: MigrateDownArgs): Promise<void> {
+  await db.execute(sql`
+   ALTER TABLE "speech_topics" DISABLE ROW LEVEL SECURITY;
+  ALTER TABLE "speech_scopes" DISABLE ROW LEVEL SECURITY;
+  ALTER TABLE "speech" DISABLE ROW LEVEL SECURITY;
+  ALTER TABLE "speech_texts" DISABLE ROW LEVEL SECURITY;
+  ALTER TABLE "speech_rels" DISABLE ROW LEVEL SECURITY;
+  ALTER TABLE "speech_segment" DISABLE ROW LEVEL SECURITY;
+  DROP TABLE "speech_topics" CASCADE;
+  DROP TABLE "speech_scopes" CASCADE;
+  DROP TABLE "speech" CASCADE;
+  DROP TABLE "speech_texts" CASCADE;
+  DROP TABLE "speech_rels" CASCADE;
+  DROP TABLE "speech_segment" CASCADE;
+  ALTER TABLE "payload_locked_documents_rels" DROP CONSTRAINT "payload_locked_documents_rels_speech_fk";
+  
+  ALTER TABLE "payload_locked_documents_rels" DROP CONSTRAINT "payload_locked_documents_rels_speech_segment_fk";
+  
+  ALTER TABLE "campaign_user" ALTER COLUMN "role" SET DATA TYPE text;
+  ALTER TABLE "campaign_user" ALTER COLUMN "role" SET DEFAULT 'leader'::text;
+  DROP TYPE "public"."enum_campaign_user_role";
+  CREATE TYPE "public"."enum_campaign_user_role" AS ENUM('coordinator', 'advisor', 'candidate', 'leader');
+  ALTER TABLE "campaign_user" ALTER COLUMN "role" SET DEFAULT 'leader'::"public"."enum_campaign_user_role";
+  ALTER TABLE "campaign_user" ALTER COLUMN "role" SET DATA TYPE "public"."enum_campaign_user_role" USING "role"::"public"."enum_campaign_user_role";
+  DROP INDEX "payload_locked_documents_rels_speech_id_idx";
+  DROP INDEX "payload_locked_documents_rels_speech_segment_id_idx";
+  ALTER TABLE "payload_locked_documents_rels" DROP COLUMN "speech_id";
+  ALTER TABLE "payload_locked_documents_rels" DROP COLUMN "speech_segment_id";
+  DROP TYPE "public"."enum_speech_topics";
+  DROP TYPE "public"."enum_speech_scopes";
+  DROP TYPE "public"."enum_speech_legislature";
+  DROP TYPE "public"."enum_speech_classified_by";`)
+}
--- a/src/migrations/20260913_001200_add_speech_segment_trgm_index.ts
+++ b/src/migrations/20260913_001200_add_speech_segment_trgm_index.ts
@ -0,0 +1,21 @@
+import { MigrateDownArgs, MigrateUpArgs, sql } from '@payloadcms/db-postgres'
+
+/**
+ * C153 — the C154 catalog search runs `ILIKE '%term%'` on the normalized
+ * `speech_segment.search_text`; the GIN trigram index makes it sargable
+ * instead of a sequential scan per query. Hand-written (not a Payload
+ * field/schema change) and idempotent — safe to re-run.
+ */
+export async function up({ db }: MigrateUpArgs): Promise<void> {
+  await db.execute(sql`CREATE EXTENSION IF NOT EXISTS pg_trgm;`)
+  await db.execute(
+    sql`CREATE INDEX IF NOT EXISTS "speech_segment_search_text_trgm_idx" ON "speech_segment" USING gin ("search_text" gin_trgm_ops);`,
+  )
+}
+
+export async function down({ db }: MigrateDownArgs): Promise<void> {
+  await db.execute(sql`DROP INDEX IF EXISTS "speech_segment_search_text_trgm_idx";`)
+  // `pg_trgm` is left installed: dropping a cluster-wide extension in a single
+  // feature's `down` is unnecessarily invasive (same rationale as the contact
+  // trigram index migration).
+}
--- a/src/migrations/index.ts
+++ b/src/migrations/index.ts
@ -66,6 +66,8 @@ import * as migration_20260824_000000_reconcile_allocation_decision_rationale_nu
 import * as migration_20260824_010000_make_supporter_import_batch_actor_nullable from './20260824_010000_make_supporter_import_batch_actor_nullable';
 import * as migration_20260912_052226_add_google_calendar_oauth_connection from './20260912_052226_add_google_calendar_oauth_connection';
 import * as migration_20260912_224719_add_activity_public_event from './20260912_224719_add_activity_public_event';
+import * as migration_20260913_001112_add_speech_catalog from './20260913_001112_add_speech_catalog';
+import * as migration_20260913_001200_add_speech_segment_trgm_index from './20260913_001200_add_speech_segment_trgm_index';

 export const migrations = [
  {
@ -401,11 +403,21 @@ export const migrations = [
  {
    up: migration_20260912_052226_add_google_calendar_oauth_connection.up,
    down: migration_20260912_052226_add_google_calendar_oauth_connection.down,
-    name: '20260912_052226_add_google_calendar_oauth_connection'
+    name: '20260912_052226_add_google_calendar_oauth_connection',
  },
  {
    up: migration_20260912_224719_add_activity_public_event.up,
    down: migration_20260912_224719_add_activity_public_event.down,
-    name: '20260912_224719_add_activity_public_event'
+    name: '20260912_224719_add_activity_public_event',
+  },
+  {
+    up: migration_20260913_001112_add_speech_catalog.up,
+    down: migration_20260913_001112_add_speech_catalog.down,
+    name: '20260913_001112_add_speech_catalog'
+  },
+  {
+    up: migration_20260913_001200_add_speech_segment_trgm_index.up,
+    down: migration_20260913_001200_add_speech_segment_trgm_index.down,
+    name: '20260913_001200_add_speech_segment_trgm_index'
  },
 ];
--- a/src/payload-types.ts
+++ b/src/payload-types.ts
@ -86,6 +86,8 @@ export interface Config {
    supporterImportBatch: SupporterImportBatch;
    municipalityUpdate: MunicipalityUpdate;
    activity: Activity;
+    speech: Speech;
+    speechSegment: SpeechSegment;
    calendarFeed: CalendarFeed;
    googleCalendarSync: GoogleCalendarSync;
    electionTally: ElectionTally;
@ -127,6 +129,8 @@ export interface Config {
    supporterImportBatch: SupporterImportBatchSelect<false> | SupporterImportBatchSelect<true>;
    municipalityUpdate: MunicipalityUpdateSelect<false> | MunicipalityUpdateSelect<true>;
    activity: ActivitySelect<false> | ActivitySelect<true>;
+    speech: SpeechSelect<false> | SpeechSelect<true>;
+    speechSegment: SpeechSegmentSelect<false> | SpeechSegmentSelect<true>;
    calendarFeed: CalendarFeedSelect<false> | CalendarFeedSelect<true>;
    googleCalendarSync: GoogleCalendarSyncSelect<false> | GoogleCalendarSyncSelect<true>;
    electionTally: ElectionTallySelect<false> | ElectionTallySelect<true>;
@ -272,7 +276,7 @@ export interface CampaignUser {
  contact?: (number | null) | Contact;
  name: string;
  avatar?: (number | null) | Media;
-  role: 'coordinator' | 'advisor' | 'candidate' | 'leader';
+  role: 'coordinator' | 'advisor' | 'candidate' | 'communicator' | 'leader';
  /**
   * O que o assessor enxerga no /campanha.
   */
@ -909,6 +913,105 @@ export interface MunicipalityUpdate {
  updatedAt: string;
  createdAt: string;
 }
+/**
+ * Acervo de falas do deputado na Câmara. Dados e vídeos da Câmara dos Deputados (CC BY 4.0).
+ *
+ * This interface was referenced by `Config`'s JSON-Schema
+ * via the `definition` "speech".
+ */
+export interface Speech {
+  id: number;
+  /**
+   * Identidade da fala na API da Câmara (data/hora + tipo + fase).
+   */
+  sourceKey: string;
+  /**
+   * Horário local de Brasília, como publicado pela Câmara.
+   */
+  speechAt: string;
+  year?: number | null;
+  legislature?: ('54' | '55' | '56' | '57') | null;
+  type?: string | null;
+  phase?: string | null;
+  durationSeconds?: number | null;
+  summary?: string | null;
+  officialTranscript?: string | null;
+  officialTextUrl?: string | null;
+  /**
+   * Preservadas cruas; as facetas não as substituem.
+   */
+  keywords?: string[] | null;
+  eventId?: number | null;
+  eventType?: string | null;
+  eventStartAt?: string | null;
+  eventEndAt?: string | null;
+  youtubeUrl?: string | null;
+  /**
+   * Derivado das trocas de mesa da sessão; vazio quando não há registro.
+   */
+  presidingOfficer?: string | null;
+  audioId?: number | null;
+  excerptTMs?: number | null;
+  /**
+   * Último link conhecido; regerável pelo VOD.
+   */
+  vodPlaybackUrl?: string | null;
+  /**
+   * Último link conhecido; regerável pelo VOD.
+   */
+  vodDownloadUrl?: string | null;
+  topics?:
+    | (
+        | 'saude'
+        | 'educacao'
+        | 'cultura'
+        | 'esporte'
+        | 'seguranca-publica'
+        | 'meio-ambiente'
+        | 'economia-trabalho'
+        | 'direitos-humanos'
+        | 'infraestrutura'
+        | 'ciencia-tecnologia'
+        | 'politica-instituicoes'
+        | 'agricultura'
+        | 'habitacao-cidades'
+        | 'comunicacao-midia'
+        | 'igualdade-racial'
+        | 'mulheres-genero'
+        | 'juventude'
+        | 'pessoa-deficiencia'
+      )[]
+    | null;
+  scopes?: ('bahia' | 'brasil' | 'internacional')[] | null;
+  classifiedBy: 'gazetteer' | 'llm' | 'manual';
+  mentionedMunicipalities?: (number | Municipality)[] | null;
+  /**
+   * Menções textuais — não é cadastro de contatos.
+   */
+  mentionedPeople?: string[] | null;
+  mentionedPrograms?: string[] | null;
+  mentionedProjects?: string[] | null;
+  updatedAt: string;
+  createdAt: string;
+}
+/**
+ * This interface was referenced by `Config`'s JSON-Schema
+ * via the `definition` "speechSegment".
+ */
+export interface SpeechSegment {
+  id: number;
+  speech: number | Speech;
+  order: number;
+  startSeconds: number;
+  endSeconds: number;
+  text: string;
+  /**
+   * Derivado do texto (sem acentos, minúsculas) para a busca por palavra.
+   */
+  searchText: string;
+  updatedAt: string;
+  createdAt: string;
+}
 /**
 * This interface was referenced by `Config`'s JSON-Schema
 * via the `definition` "calendarFeed".
@ -1446,6 +1549,14 @@ export interface PayloadLockedDocument {
        relationTo: 'activity';
        value: number | Activity;
      } | null)
+    | ({
+        relationTo: 'speech';
+        value: number | Speech;
+      } | null)
+    | ({
+        relationTo: 'speechSegment';
+        value: number | SpeechSegment;
+      } | null)
    | ({
        relationTo: 'calendarFeed';
        value: number | CalendarFeed;
@ -1954,6 +2065,56 @@ export interface ActivitySelect<T extends boolean = true> {
  updatedAt?: T;
  createdAt?: T;
 }
+/**
+ * This interface was referenced by `Config`'s JSON-Schema
+ * via the `definition` "speech_select".
+ */
+export interface SpeechSelect<T extends boolean = true> {
+  sourceKey?: T;
+  speechAt?: T;
+  year?: T;
+  legislature?: T;
+  type?: T;
+  phase?: T;
+  durationSeconds?: T;
+  summary?: T;
+  officialTranscript?: T;
+  officialTextUrl?: T;
+  keywords?: T;
+  eventId?: T;
+  eventType?: T;
+  eventStartAt?: T;
+  eventEndAt?: T;
+  youtubeUrl?: T;
+  presidingOfficer?: T;
+  audioId?: T;
+  excerptTMs?: T;
+  vodPlaybackUrl?: T;
+  vodDownloadUrl?: T;
+  topics?: T;
+  scopes?: T;
+  classifiedBy?: T;
+  mentionedMunicipalities?: T;
+  mentionedPeople?: T;
+  mentionedPrograms?: T;
+  mentionedProjects?: T;
+  updatedAt?: T;
+  createdAt?: T;
+}
+/**
+ * This interface was referenced by `Config`'s JSON-Schema
+ * via the `definition` "speechSegment_select".
+ */
+export interface SpeechSegmentSelect<T extends boolean = true> {
+  speech?: T;
+  order?: T;
+  startSeconds?: T;
+  endSeconds?: T;
+  text?: T;
+  searchText?: T;
+  updatedAt?: T;
+  createdAt?: T;
+}
 /**
 * This interface was referenced by `Config`'s JSON-Schema
 * via the `definition` "calendarFeed_select".
@ -2684,6 +2845,8 @@ export interface TaskCreateCollectionExport {
      | 'supporterImportBatch'
      | 'municipalityUpdate'
      | 'activity'
+      | 'speech'
+      | 'speechSegment'
      | 'calendarFeed'
      | 'googleCalendarSync'
      | 'electionTally'
--- a/src/payload.config.ts
+++ b/src/payload.config.ts
@ -33,6 +33,8 @@ import { Petition } from './collections/Petition'
 import { Post } from './collections/Post'
 import { PushSubscription } from './collections/PushSubscription'
 import { Signature } from './collections/Signature'
+import { Speech } from './collections/Speech'
+import { SpeechSegment } from './collections/SpeechSegment'
 import { StateDeputy } from './collections/StateDeputy'
 import { Subscription } from './collections/Subscription'
 import { Supporter } from './collections/Supporter'
@ -111,6 +113,8 @@ export default buildConfig({
    SupporterImportBatch,
    MunicipalityUpdate,
    Activity,
+    Speech,
+    SpeechSegment,
    CalendarFeed,
    GoogleCalendarSync,
    ElectionTally,
--- a/src/utilities/access/speeches.ts
+++ b/src/utilities/access/speeches.ts
@ -0,0 +1,31 @@
+// ---------------------------------------------------------------------------
+// Speech catalog (C153) — the "acervo de falas" read surface
+// ---------------------------------------------------------------------------
+
+import type { Access } from 'payload'
+
+import { canReadSpeechCatalog } from '@/lib/campaignRoles'
+import {
+  getFreshCampaignUser,
+  isCampaignUnrestricted,
+  isPayloadAdmin,
+} from '@/utilities/access/shared'
+
+/**
+ * Readers of the speech catalog: Payload admin or campaign users with the
+ * `communicator`, `coordinator` or `candidate` role. Advisors and leaders are
+ * denied — the catalog is not municipality-scoped and not a staff area.
+ */
+export const canReadSpeech: Access = async ({ req }) => {
+  if (isPayloadAdmin(req.user)) return true
+
+  const currentUser = await getFreshCampaignUser(req)
+  return currentUser ? canReadSpeechCatalog(currentUser.role) : false
+}
+
+/** Manual facet correction — unrestricted roles (and Payload admin) only. */
+export const canUpdateSpeech: Access = async ({ req }) => {
+  if (isPayloadAdmin(req.user)) return true
+
+  return isCampaignUnrestricted(await getFreshCampaignUser(req))
+}
--- a/src/utilities/campaignAccess.ts
+++ b/src/utilities/campaignAccess.ts
@ -168,3 +168,5 @@ export {
  canSetCalendarFeedSystemField,
  canUpdateCalendarFeed,
 } from '@/utilities/access/calendarFeeds'
+
+export { canReadSpeech, canUpdateSpeech } from '@/utilities/access/speeches'
--- a/src/utilities/campaignUserProfile.ts
+++ b/src/utilities/campaignUserProfile.ts
@ -10,6 +10,7 @@ export const campaignRoleLabels: Record<CampaignUser['role'], string> = {
  coordinator: 'Coordenador Geral',
  advisor: 'Assessor',
  candidate: 'Candidato',
+  communicator: 'Assessor de Comunicação',
  leader: 'Liderança',
 }

--- a/src/utilities/speech/speechClassifier.ts
+++ b/src/utilities/speech/speechClassifier.ts
@ -0,0 +1,136 @@
+import 'server-only'
+
+import { z } from 'zod'
+
+import {
+  mergeFacetClassification,
+  parseLlmFacetResponse,
+  SPEECH_SCOPES,
+  SPEECH_TOPICS,
+  type MergedSpeechFacets,
+} from '@/lib/speechFacets'
+import { classifySpeechByGazetteer, type SpeechFacetInput } from '@/lib/speechGazetteer'
+
+/**
+ * Facet classification for one speech (C153): the offline gazetteer pass plus
+ * one LLM refinement on Deep Infra (same key/provider as the ASR). The LLM
+ * output is validated against the taxonomy; any failure — missing key, HTTP
+ * error, timeout, invalid JSON — degrades to the gazetteer provenance and
+ * never throws, so the import continues.
+ */
+
+const DEEPINFRA_CHAT_URL = 'https://api.deepinfra.com/v1/openai/chat/completions'
+const DEEPINFRA_FACET_MODEL = 'deepseek-ai/DeepSeek-V4-Flash'
+const FACET_TIMEOUT_MS = 60_000
+const TRANSCRIPT_PROMPT_LIMIT = 12_000
+
+export type SpeechClassificationResult = {
+  facets: MergedSpeechFacets
+  llm: {
+    used: boolean
+    totalTokens: number | null
+    estimatedCostUsd: number | null
+    error: string | null
+  }
+}
+
+const TOPIC_PROMPT_VALUES = SPEECH_TOPICS.map(({ value, label }) => `${value} (${label})`).join(
+  '; ',
+)
+const SCOPE_PROMPT_VALUES = SPEECH_SCOPES.map(({ value, label }) => `${value} (${label})`).join(
+  '; ',
+)
+
+const FACET_SYSTEM_PROMPT = [
+  'Você classifica discursos do deputado federal Jorge Solla (PT-BA) para um catálogo de falas.',
+  'Responda SOMENTE com um JSON válido, sem markdown, no formato:',
+  '{"topics": [], "scopes": [], "people": [], "programs": [], "projects": []}',
+  `topics: subconjunto de [${TOPIC_PROMPT_VALUES}] — use exatamente esses valores.`,
+  `scopes: subconjunto de [${SCOPE_PROMPT_VALUES}] — use exatamente esses valores.`,
+  'people: pessoas citadas, nome completo quando houver (máx. 12).',
+  'programs: programas de governo citados (ex.: "Minha Casa, Minha Vida").',
+  'projects: proposições citadas (PL, PEC, medida provisória), com o número quando houver.',
+  'Não invente: liste apenas o que aparece no sumário, nas palavras-chave ou na transcrição.',
+  'Máximo 4 topics. Sem conteúdo em uma lista, devolva [].',
+].join('\n')
+
+const llmEnvelopeSchema = z.object({
+  choices: z.array(z.object({ message: z.object({ content: z.string() }).passthrough() })).min(1),
+  usage: z
+    .object({
+      total_tokens: z.number().optional(),
+      estimated_cost: z.number().optional(),
+    })
+    .optional(),
+})
+
+const buildUserPrompt = (input: SpeechFacetInput): string => {
+  const transcript = input.transcript.slice(0, TRANSCRIPT_PROMPT_LIMIT)
+  return [
+    `SUMÁRIO: ${input.summary ?? '—'}`,
+    `PALAVRAS-CHAVE OFICIAIS: ${input.keywords?.join('; ') ?? '—'}`,
+    `TRANSCRIÇÃO (trecho): ${transcript}`,
+  ].join('\n')
+}
+
+export const classifySpeech = async (
+  input: SpeechFacetInput,
+): Promise<SpeechClassificationResult> => {
+  const gazetteer = classifySpeechByGazetteer(input)
+  const apiKey = process.env.DEEPINFRA_API_KEY
+  if (!apiKey) {
+    return {
+      facets: mergeFacetClassification(gazetteer, null),
+      llm: { used: false, totalTokens: null, estimatedCostUsd: null, error: 'chave ausente' },
+    }
+  }
+
+  try {
+    const response = await fetch(DEEPINFRA_CHAT_URL, {
+      method: 'POST',
+      headers: {
+        Authorization: `Bearer ${apiKey}`,
+        'Content-Type': 'application/json',
+      },
+      body: JSON.stringify({
+        model: DEEPINFRA_FACET_MODEL,
+        messages: [
+          { role: 'system', content: FACET_SYSTEM_PROMPT },
+          { role: 'user', content: buildUserPrompt(input) },
+        ],
+        response_format: { type: 'json_object' },
+        temperature: 0,
+        max_tokens: 500,
+      }),
+      signal: AbortSignal.timeout(FACET_TIMEOUT_MS),
+    })
+    if (!response.ok) {
+      throw new Error(`HTTP ${response.status}`)
+    }
+
+    const envelope = llmEnvelopeSchema.safeParse(await response.json())
+    if (!envelope.success) throw new Error('resposta do provedor fora do formato')
+    const parsed = parseLlmFacetResponse(JSON.parse(envelope.data.choices[0].message.content))
+    if (!parsed) throw new Error('resposta fora da taxonomia')
+
+    return {
+      facets: mergeFacetClassification(gazetteer, parsed),
+      llm: {
+        used: true,
+        totalTokens: envelope.data.usage?.total_tokens ?? null,
+        estimatedCostUsd: envelope.data.usage?.estimated_cost ?? null,
+        error: null,
+      },
+    }
+  } catch (error) {
+    return {
+      facets: mergeFacetClassification(gazetteer, null),
+      llm: {
+        used: false,
+        totalTokens: null,
+        estimatedCostUsd: null,
+        error: error instanceof Error ? error.message : String(error),
+      },
+    }
+  }
+}
--- a/src/utilities/speech/speechImport.ts
+++ b/src/utilities/speech/speechImport.ts
@ -0,0 +1,248 @@
+import 'server-only'
+
+import type { Payload, RequiredDataFromCollectionSlug } from 'payload'
+
+import type { MergedSpeechFacets } from '@/lib/speechFacets'
+import { normalizeForSearch } from '@/lib/speechSearch'
+import type { Speech } from '@/payload-types'
+import { withPayloadTransaction } from '@/utilities/payloadTransaction'
+
+/**
+ * Idempotent persistence of one imported speech (C153). Identity is
+ * `sourceKey`; metadata is refreshed on every run, segments are replaced only
+ * when the caller supplies them, and a `manual` facet curation is never
+ * overwritten. Everything runs in one transaction on the Payload session.
+ */
+
+type SpeechSegmentInput = {
+  startSeconds: number
+  endSeconds: number
+  text: string
+}
+
+export type SpeechImportBundle = {
+  sourceKey: string
+  speechAt: string
+  year: number | null
+  legislature: Speech['legislature']
+  type: string | null
+  phase: string | null
+  durationSeconds: number | null
+  summary: string | null
+  officialTranscript: string | null
+  officialTextUrl: string | null
+  keywords: string[]
+  eventId: number | null
+  eventType: string | null
+  eventStartAt: string | null
+  eventEndAt: string | null
+  youtubeUrl: string | null
+  presidingOfficer: string | null
+  audioId: number | null
+  excerptTMs: number | null
+  vodPlaybackUrl: string | null
+  vodDownloadUrl: string | null
+  /** Undefined preserves the stored facets (e.g. a run that skipped classification). */
+  facets?: MergedSpeechFacets
+  /** Undefined preserves the stored segments (e.g. `--skip-transcribe`). */
+  segments?: SpeechSegmentInput[]
+}
+
+export type SpeechImportState = {
+  id: number
+  audioId: number | null
+  excerptTMs: number | null
+  segmentCount: number
+  classifiedBy: Speech['classifiedBy']
+  durationSeconds: number | null
+  vodPlaybackUrl: string | null
+  vodDownloadUrl: string | null
+}
+
+export type SpeechUpsertCounts = {
+  created: boolean
+  segmentsInserted: number
+  segmentsDeleted: number
+  manualFacetsPreserved: boolean
+}
+
+/** Current stored state of one speech, for the importer's skip decisions. */
+export const findSpeechImportState = async (
+  payload: Payload,
+  sourceKey: string,
+): Promise<SpeechImportState | null> => {
+  const found = await payload.find({
+    collection: 'speech',
+    where: { sourceKey: { equals: sourceKey } },
+    limit: 1,
+    depth: 0,
+    select: {
+      audioId: true,
+      excerptTMs: true,
+      classifiedBy: true,
+      durationSeconds: true,
+      vodPlaybackUrl: true,
+      vodDownloadUrl: true,
+    },
+    // Intentional bypass: the import CLI is a trusted actor with no session.
+    overrideAccess: true,
+  })
+  const speech = found.docs[0]
+  if (!speech) return null
+
+  const segments = await payload.count({
+    collection: 'speechSegment',
+    where: { speech: { equals: speech.id } },
+    // Intentional bypass: the import CLI is a trusted actor with no session.
+    overrideAccess: true,
+  })
+
+  return {
+    id: speech.id,
+    audioId: speech.audioId ?? null,
+    excerptTMs: speech.excerptTMs ?? null,
+    segmentCount: segments.totalDocs,
+    classifiedBy: speech.classifiedBy,
+    durationSeconds: speech.durationSeconds ?? null,
+    vodPlaybackUrl: speech.vodPlaybackUrl ?? null,
+    vodDownloadUrl: speech.vodDownloadUrl ?? null,
+  }
+}
+
+type SpeechWriteData = RequiredDataFromCollectionSlug<'speech'>
+
+const metadataData = (
+  bundle: SpeechImportBundle,
+): Omit<SpeechWriteData, 'sourceKey' | 'classifiedBy'> => ({
+  speechAt: bundle.speechAt,
+  year: bundle.year,
+  legislature: bundle.legislature,
+  type: bundle.type,
+  phase: bundle.phase,
+  durationSeconds: bundle.durationSeconds,
+  summary: bundle.summary,
+  officialTranscript: bundle.officialTranscript,
+  officialTextUrl: bundle.officialTextUrl,
+  keywords: bundle.keywords,
+  eventId: bundle.eventId,
+  eventType: bundle.eventType,
+  eventStartAt: bundle.eventStartAt,
+  eventEndAt: bundle.eventEndAt,
+  youtubeUrl: bundle.youtubeUrl,
+  presidingOfficer: bundle.presidingOfficer,
+  audioId: bundle.audioId,
+  excerptTMs: bundle.excerptTMs,
+  vodPlaybackUrl: bundle.vodPlaybackUrl,
+  vodDownloadUrl: bundle.vodDownloadUrl,
+})
+
+export const upsertSpeechBundle = async (
+  payload: Payload,
+  bundle: SpeechImportBundle,
+): Promise<SpeechUpsertCounts> =>
+  withPayloadTransaction(payload, async ({ req }) => {
+    const found = await payload.find({
+      collection: 'speech',
+      where: { sourceKey: { equals: bundle.sourceKey } },
+      limit: 1,
+      depth: 0,
+      select: { id: true, classifiedBy: true },
+      req,
+      // Intentional bypass: the import CLI is a trusted actor with no session.
+      overrideAccess: true,
+    })
+    const current = found.docs[0]
+    const manualFacetsPreserved = current?.classifiedBy === 'manual'
+
+    const metadata = metadataData(bundle)
+    const facets: Partial<SpeechWriteData> = {}
+    if (bundle.facets && !manualFacetsPreserved) {
+      const municipalitySlugs = bundle.facets.municipalities.map((entry) => entry.slug)
+      const municipalities =
+        municipalitySlugs.length === 0
+          ? []
+          : (
+              await payload.find({
+                collection: 'municipality',
+                where: { slug: { in: municipalitySlugs } },
+                limit: municipalitySlugs.length,
+                depth: 0,
+                select: { slug: true },
+                req,
+                // Intentional bypass: resolving gazetteer slugs inside the import transaction.
+                overrideAccess: true,
+              })
+            ).docs
+      facets.topics = bundle.facets.topics
+      facets.scopes = bundle.facets.scopes
+      facets.classifiedBy = bundle.facets.classifiedBy
+      facets.mentionedMunicipalities = municipalities.map((doc) => doc.id)
+      facets.mentionedPeople = bundle.facets.people
+      facets.mentionedPrograms = bundle.facets.programs
+      facets.mentionedProjects = bundle.facets.projects
+    }
+    const data = { ...metadata, ...facets }
+
+    let speechId: number
+    let created = false
+    if (current) {
+      await payload.update({
+        collection: 'speech',
+        id: current.id,
+        data,
+        depth: 0,
+        req,
+        // Intentional bypass: the import CLI is a trusted actor with no session.
+        overrideAccess: true,
+      })
+      speechId = current.id
+    } else {
+      const speech = await payload.create({
+        collection: 'speech',
+        data: {
+          sourceKey: bundle.sourceKey,
+          classifiedBy: bundle.facets?.classifiedBy ?? 'gazetteer',
+          ...data,
+        },
+        depth: 0,
+        req,
+        // Intentional bypass: the import CLI is a trusted actor with no session.
+        overrideAccess: true,
+      })
+      speechId = speech.id
+      created = true
+    }
+
+    let segmentsInserted = 0
+    let segmentsDeleted = 0
+    if (bundle.segments !== undefined) {
+      const removed = await payload.delete({
+        collection: 'speechSegment',
+        where: { speech: { equals: speechId } },
+        req,
+        // Intentional bypass: replacing segments owned by the speech being imported.
+        overrideAccess: true,
+      })
+      segmentsDeleted = removed.docs.length
+
+      for (const [index, segment] of bundle.segments.entries()) {
+        await payload.create({
+          collection: 'speechSegment',
+          data: {
+            speech: speechId,
+            order: index + 1,
+            startSeconds: segment.startSeconds,
+            endSeconds: segment.endSeconds,
+            text: segment.text,
+            searchText: normalizeForSearch(segment.text),
+          },
+          req,
+          // Intentional bypass: the import CLI is a trusted actor with no session.
+          overrideAccess: true,
+        })
+        segmentsInserted += 1
+      }
+    }
+
+    return { created, segmentsInserted, segmentsDeleted, manualFacetsPreserved }
+  })
--- a/tests/int/speechCatalog.int.spec.ts
+++ b/tests/int/speechCatalog.int.spec.ts
@ -0,0 +1,277 @@
+// @vitest-environment node
+
+import { randomUUID } from 'node:crypto'
+
+import type { Payload, RequiredDataFromCollectionSlug } from 'payload'
+import { getPayload } from 'payload'
+import { afterAll, beforeAll, describe, expect, it } from 'vitest'
+
+import type { Speech } from '@/payload-types'
+import config from '@/payload.config'
+
+import { installCampaignFixtures } from '../helpers/campaignFixtures'
+
+let payload: Payload
+const createdSpeechIds: number[] = []
+const campaignFixtures = installCampaignFixtures({
+  getPayload: () => payload,
+  setPayload: (nextPayload) => {
+    payload = nextPayload
+  },
+})
+
+const createSpeech = async (
+  overrides: Partial<RequiredDataFromCollectionSlug<'speech'>> = {},
+): Promise<Speech> => {
+  const speech = await payload.create({
+    collection: 'speech',
+    data: {
+      sourceKey: `test-${randomUUID()}`,
+      speechAt: '2023-02-07T17:28',
+      type: 'BREVES COMUNICAÇÕES',
+      phase: 'Breves Comunicações',
+      summary: 'A defesa do SUS e da saúde pública baiana.',
+      officialTranscript: 'O SR. JORGE SOLLA (Bloco/PT - BA) - Sr. Presidente...',
+      keywords: ['Saúde', 'SUS'],
+      topics: ['saude'],
+      classifiedBy: 'gazetteer',
+      ...overrides,
+    },
+    depth: 0,
+    overrideAccess: true,
+  })
+  createdSpeechIds.push(speech.id)
+  return speech
+}
+
+describe('speech catalog (C153)', () => {
+  beforeAll(async () => {
+    payload = await getPayload({ config: await config })
+  })
+
+  afterAll(async () => {
+    for (const id of createdSpeechIds) {
+      try {
+        await payload.delete({ collection: 'speech', id, depth: 0, overrideAccess: true })
+      } catch {
+        // already deleted by a cascade assertion
+      }
+    }
+  })
+
+  it('derives the year from the wall-clock speechAt', async () => {
+    const speech = await createSpeech()
+
+    expect(speech.year).toBe(2023)
+  })
+
+  it('normalizes the segment text into searchText and matches accent-insensitively', async () => {
+    const speech = await createSpeech()
+    const segment = await payload.create({
+      collection: 'speechSegment',
+      data: {
+        speech: speech.id,
+        order: 1,
+        startSeconds: 0,
+        endSeconds: 4.2,
+        text: 'A defesa do SUS e da saúde pública',
+        // Deliberately wrong: the beforeValidate hook owns this field.
+        searchText: 'errado',
+      },
+      depth: 0,
+      overrideAccess: true,
+    })
+
+    expect(segment.searchText).toBe('a defesa do sus e da saude publica')
+
+    const hits = await payload.find({
+      collection: 'speechSegment',
+      where: { searchText: { contains: 'saude' } },
+      depth: 0,
+      overrideAccess: true,
+    })
+    expect(hits.docs.map((doc) => doc.id)).toContain(segment.id)
+  })
+
+  it('deletes the speech segments with the speech', async () => {
+    const speech = await createSpeech()
+    const segment = await payload.create({
+      collection: 'speechSegment',
+      data: {
+        speech: speech.id,
+        order: 1,
+        startSeconds: 0,
+        endSeconds: 2,
+        text: 'Trecho qualquer',
+        searchText: '',
+      },
+      depth: 0,
+      overrideAccess: true,
+    })
+
+    await payload.delete({ collection: 'speech', id: speech.id, depth: 0, overrideAccess: true })
+
+    const remaining = await payload.find({
+      collection: 'speechSegment',
+      where: { id: { equals: segment.id } },
+      depth: 0,
+      overrideAccess: true,
+    })
+    expect(remaining.docs).toHaveLength(0)
+  })
+
+  it('reads for communicator, coordinator, candidate and admin; denies advisor, leader and anonymous', async () => {
+    const speech = await createSpeech()
+    const fixtures = campaignFixtures()
+    const communicator = await fixtures.createCampaignUser('communicator')
+    const coordinator = await fixtures.createCampaignUser('coordinator')
+    const candidate = await fixtures.createCampaignUser('candidate')
+    const advisor = await fixtures.createCampaignUser('advisor')
+    const leader = await fixtures.createCampaignUser('leader')
+    const admin = await fixtures.createAdminUser()
+
+    for (const reader of [communicator, coordinator, candidate, admin]) {
+      const found = await payload.find({
+        collection: 'speech',
+        where: { id: { equals: speech.id } },
+        depth: 0,
+        user: reader,
+        overrideAccess: false,
+      })
+      expect(found.docs).toHaveLength(1)
+    }
+
+    for (const denied of [advisor, leader, undefined]) {
+      await expect(
+        payload.find({
+          collection: 'speech',
+          where: { id: { equals: speech.id } },
+          depth: 0,
+          user: denied,
+          overrideAccess: false,
+        }),
+      ).rejects.toThrow(/permissão/i)
+    }
+  })
+
+  it('updates facets for coordinator/candidate/admin but not for communicator/advisor', async () => {
+    const speech = await createSpeech()
+    const fixtures = campaignFixtures()
+    const coordinator = await fixtures.createCampaignUser('coordinator')
+    const candidate = await fixtures.createCampaignUser('candidate')
+    const communicator = await fixtures.createCampaignUser('communicator')
+    const advisor = await fixtures.createCampaignUser('advisor')
+    const admin = await fixtures.createAdminUser()
+
+    for (const editor of [coordinator, candidate, admin]) {
+      const updated = await payload.update({
+        collection: 'speech',
+        id: speech.id,
+        data: { topics: ['educacao'] },
+        depth: 0,
+        user: editor,
+        overrideAccess: false,
+      })
+      expect(updated.topics).toEqual(['educacao'])
+    }
+
+    for (const denied of [communicator, advisor]) {
+      await expect(
+        payload.update({
+          collection: 'speech',
+          id: speech.id,
+          data: { topics: ['cultura'] },
+          depth: 0,
+          user: denied,
+          overrideAccess: false,
+        }),
+      ).rejects.toThrow(/permissão/i)
+    }
+  })
+
+  it('creates/deletes only for Payload admin', async () => {
+    const fixtures = campaignFixtures()
+    const coordinator = await fixtures.createCampaignUser('coordinator')
+    const admin = await fixtures.createAdminUser()
+
+    await expect(
+      payload.create({
+        collection: 'speech',
+        data: {
+          sourceKey: `test-${randomUUID()}`,
+          speechAt: '2023-02-08T10:00',
+          classifiedBy: 'gazetteer',
+        },
+        depth: 0,
+        user: coordinator,
+        overrideAccess: false,
+      }),
+    ).rejects.toThrow(/permissão/i)
+
+    const speech = await createSpeech()
+    await expect(
+      payload.delete({
+        collection: 'speech',
+        id: speech.id,
+        depth: 0,
+        user: coordinator,
+        overrideAccess: false,
+      }),
+    ).rejects.toThrow(/permissão/i)
+
+    const adminCreated = await payload.create({
+      collection: 'speech',
+      data: {
+        sourceKey: `test-${randomUUID()}`,
+        speechAt: '2023-02-08T10:00',
+        classifiedBy: 'gazetteer',
+      },
+      depth: 0,
+      user: admin,
+      overrideAccess: false,
+    })
+    createdSpeechIds.push(adminCreated.id)
+    expect(adminCreated.id).toBeGreaterThan(0)
+  })
+
+  it('preserves manual facets on unauthenticated writes and lets admin curation through', async () => {
+    const speech = await createSpeech({ classifiedBy: 'manual', topics: ['saude'] })
+
+    await payload.update({
+      collection: 'speech',
+      id: speech.id,
+      data: { topics: ['cultura'], scopes: ['brasil'], classifiedBy: 'llm' },
+      depth: 0,
+      overrideAccess: true,
+    })
+
+    const afterImport = await payload.findByID({
+      collection: 'speech',
+      id: speech.id,
+      depth: 0,
+      overrideAccess: true,
+    })
+    expect(afterImport.topics).toEqual(['saude'])
+    expect(afterImport.scopes).toEqual([])
+    expect(afterImport.classifiedBy).toBe('manual')
+
+    const admin = await campaignFixtures().createAdminUser()
+    await payload.update({
+      collection: 'speech',
+      id: speech.id,
+      data: { topics: ['cultura'] },
+      depth: 0,
+      user: admin,
+      overrideAccess: false,
+    })
+
+    const curated = await payload.findByID({
+      collection: 'speech',
+      id: speech.id,
+      depth: 0,
+      overrideAccess: true,
+    })
+    expect(curated.topics).toEqual(['cultura'])
+    expect(curated.classifiedBy).toBe('manual')
+  })
+})
--- a/tests/int/speechImport.int.spec.ts
+++ b/tests/int/speechImport.int.spec.ts
@ -0,0 +1,206 @@
+// @vitest-environment node
+
+import { randomUUID } from 'node:crypto'
+
+import type { Payload } from 'payload'
+import { getPayload } from 'payload'
+import { afterAll, beforeAll, describe, expect, it } from 'vitest'
+
+import { matchMunicipalityMentions } from '@/lib/speechGazetteer'
+import config from '@/payload.config'
+import {
+  findSpeechImportState,
+  upsertSpeechBundle,
+  type SpeechImportBundle,
+} from '@/utilities/speech/speechImport'
+
+let payload: Payload
+const createdSourceKeys = new Set<string>()
+
+const baseBundle = (overrides: Partial<SpeechImportBundle> = {}): SpeechImportBundle => {
+  const sourceKey = `test-import-${randomUUID()}`
+  createdSourceKeys.add(sourceKey)
+  return {
+    sourceKey,
+    speechAt: '2023-02-07T17:28',
+    year: 2023,
+    legislature: '57',
+    type: 'BREVES COMUNICAÇÕES',
+    phase: 'Breves Comunicações',
+    durationSeconds: 252,
+    summary: 'Saúde e educação em Feira de Santana.',
+    officialTranscript: 'O SR. JORGE SOLLA (Bloco/PT - BA) - Sr. Presidente...',
+    officialTextUrl: null,
+    keywords: ['SUS'],
+    eventId: 67091,
+    eventType: 'Sessão Deliberativa',
+    eventStartAt: '2023-02-07T15:00',
+    eventEndAt: '2023-02-07T21:23',
+    youtubeUrl: null,
+    presidingOfficer: 'Pompeo de Mattos',
+    audioId: 558641,
+    excerptTMs: 1675801808560,
+    vodPlaybackUrl: null,
+    vodDownloadUrl: null,
+    facets: {
+      topics: ['saude'],
+      scopes: ['bahia'],
+      municipalities: matchMunicipalityMentions('Feira de Santana'),
+      people: ['Lula'],
+      programs: [],
+      projects: [],
+      classifiedBy: 'gazetteer',
+    },
+    segments: [{ startSeconds: 0, endSeconds: 2.7, text: 'A saúde pública baiana' }],
+    ...overrides,
+  }
+}
+
+const findBySourceKey = async (sourceKey: string) => {
+  const found = await payload.find({
+    collection: 'speech',
+    where: { sourceKey: { equals: sourceKey } },
+    depth: 1,
+    overrideAccess: true,
+  })
+  return found.docs[0] ?? null
+}
+
+const countSegments = async (speechId: number) => {
+  const counted = await payload.count({
+    collection: 'speechSegment',
+    where: { speech: { equals: speechId } },
+    overrideAccess: true,
+  })
+  return counted.totalDocs
+}
+
+describe('upsertSpeechBundle (C153)', () => {
+  beforeAll(async () => {
+    payload = await getPayload({ config: await config })
+  })
+
+  afterAll(async () => {
+    for (const sourceKey of createdSourceKeys) {
+      const speech = await findBySourceKey(sourceKey)
+      if (!speech) continue
+      await payload.delete({ collection: 'speech', id: speech.id, depth: 0, overrideAccess: true })
+    }
+  })
+
+  it('creates once and replaces segments on re-run without duplicating', async () => {
+    const bundle = baseBundle()
+
+    const first = await upsertSpeechBundle(payload, bundle)
+    expect(first.created).toBe(true)
+    expect(first.segmentsInserted).toBe(1)
+
+    const speech = await findBySourceKey(bundle.sourceKey)
+    expect(speech).not.toBeNull()
+    expect(speech?.topics).toEqual(['saude'])
+    expect(speech?.mentionedMunicipalities?.[0]).toMatchObject({ slug: 'feira-de-santana' })
+
+    const firstSegments = await payload.find({
+      collection: 'speechSegment',
+      where: { speech: { equals: speech?.id } },
+      depth: 0,
+      overrideAccess: true,
+    })
+    expect(firstSegments.docs[0]?.searchText).toBe('a saude publica baiana')
+
+    const second = await upsertSpeechBundle(payload, bundle)
+    expect(second.created).toBe(false)
+    expect(second.segmentsDeleted).toBe(1)
+    expect(second.segmentsInserted).toBe(1)
+
+    const again = await findBySourceKey(bundle.sourceKey)
+    expect(again?.id).toBe(speech?.id)
+    expect(await countSegments(again?.id ?? 0)).toBe(1)
+  })
+
+  it('preserves manual facets and skips facet writes', async () => {
+    const bundle = baseBundle()
+    await upsertSpeechBundle(payload, bundle)
+    const speech = await findBySourceKey(bundle.sourceKey)
+    await payload.update({
+      collection: 'speech',
+      id: speech?.id,
+      data: { classifiedBy: 'manual', topics: ['cultura'] },
+      depth: 0,
+      overrideAccess: true,
+    })
+
+    const counts = await upsertSpeechBundle(
+      payload,
+      baseBundle({
+        sourceKey: bundle.sourceKey,
+        facets: {
+          topics: ['saude'],
+          scopes: ['bahia'],
+          municipalities: matchMunicipalityMentions('Feira de Santana'),
+          people: [],
+          programs: [],
+          projects: [],
+          classifiedBy: 'gazetteer',
+        },
+      }),
+    )
+
+    expect(counts.manualFacetsPreserved).toBe(true)
+    const after = await findBySourceKey(bundle.sourceKey)
+    expect(after?.topics).toEqual(['cultura'])
+    expect(after?.classifiedBy).toBe('manual')
+  })
+
+  it('preserves the stored facets when the bundle carries none', async () => {
+    const bundle = baseBundle()
+    await upsertSpeechBundle(payload, bundle)
+    const speech = await findBySourceKey(bundle.sourceKey)
+
+    await upsertSpeechBundle(
+      payload,
+      baseBundle({ sourceKey: bundle.sourceKey, facets: undefined }),
+    )
+
+    const after = await findBySourceKey(bundle.sourceKey)
+    expect(after?.topics).toEqual(speech?.topics)
+    expect(after?.classifiedBy).toBe('gazetteer')
+  })
+
+  it('preserves the stored segments when the bundle carries none', async () => {
+    const bundle = baseBundle()
+    await upsertSpeechBundle(payload, bundle)
+    const speech = await findBySourceKey(bundle.sourceKey)
+
+    await upsertSpeechBundle(
+      payload,
+      baseBundle({ sourceKey: bundle.sourceKey, segments: undefined }),
+    )
+
+    expect(await countSegments(speech?.id ?? 0)).toBe(1)
+  })
+
+  it('reports the stored import state for skip decisions', async () => {
+    const bundle = baseBundle()
+    await upsertSpeechBundle(payload, bundle)
+
+    const state = await findSpeechImportState(payload, bundle.sourceKey)
+    expect(state).toMatchObject({
+      audioId: 558641,
+      excerptTMs: 1675801808560,
+      segmentCount: 1,
+      classifiedBy: 'gazetteer',
+    })
+    expect(await findSpeechImportState(payload, 'test-import-missing')).toBeNull()
+  })
+
+  it('keeps a facet-less new record with the default provenance', async () => {
+    const bundle = baseBundle({ facets: undefined, segments: undefined })
+    const counts = await upsertSpeechBundle(payload, bundle)
+
+    expect(counts.created).toBe(true)
+    const speech = await findBySourceKey(bundle.sourceKey)
+    expect(speech?.classifiedBy).toBe('gazetteer')
+    expect(speech?.topics).toEqual([])
+  })
+})
--- a/tests/unit/camaraSpeeches.unit.spec.ts
+++ b/tests/unit/camaraSpeeches.unit.spec.ts
@ -5,13 +5,19 @@ import { describe, expect, it } from 'vitest'
 import {
  buildVodUrl,
  clockToSeconds,
+  legislatureForDate,
  matchExcerpt,
  normalizeSpeakerName,
  normalizeTranscription,
+  parseDurationToSeconds,
  parseEventExcerpts,
+  parseOfficialKeywords,
+  parsePresidingOfficerTransitions,
  parseVodStatus,
+  resolvePresidingOfficer,
  selectSpeechEvents,
  speechDate,
+  speechSourceKey,
  speechTimeOfDaySeconds,
 } from '../../scripts/lib/camaraSpeeches.mjs'

@ -251,6 +257,146 @@ describe('matchExcerpt', () => {
  })
 })

+describe('speechSourceKey', () => {
+  it('joins datetime, type and phase into the stable identity', () => {
+    expect(
+      speechSourceKey({
+        dataHoraInicio: '2023-02-07T17:28',
+        tipoDiscurso: 'BREVES COMUNICAÇÕES',
+        faseEvento: { titulo: 'Breves Comunicações' },
+      }),
+    ).toBe('2023-02-07T17:28|BREVES COMUNICAÇÕES|Breves Comunicações')
+  })
+
+  it('degrades missing parts to empty strings', () => {
+    expect(speechSourceKey({})).toBe('||')
+    expect(speechSourceKey(null)).toBe('||')
+  })
+})
+
+describe('parseOfficialKeywords', () => {
+  it('splits the raw newline string and keeps each keyword', () => {
+    expect(parseOfficialKeywords('Saúde, SUS\r\nEducação\r\n\r\nEscola')).toEqual([
+      'Saúde, SUS',
+      'Educação',
+      'Escola',
+    ])
+    expect(parseOfficialKeywords(null)).toEqual([])
+  })
+})
+
+describe('legislatureForDate', () => {
+  it('maps dates to the legislature ranges at the boundaries', () => {
+    expect(legislatureForDate('2023-02-01')).toBe('57')
+    expect(legislatureForDate('2023-01-31')).toBe('56')
+    expect(legislatureForDate('2015-02-01')).toBe('55')
+    expect(legislatureForDate('2015-01-31')).toBe('54')
+    expect(legislatureForDate('2011-01-01')).toBe('54')
+  })
+
+  it('returns null outside the supported range or for invalid input', () => {
+    expect(legislatureForDate('2010-12-31')).toBeNull()
+    expect(legislatureForDate('')).toBeNull()
+    expect(legislatureForDate('2023-2-1')).toBeNull()
+  })
+})
+
+describe('parseDurationToSeconds', () => {
+  it('parses the excerpt card and VOD spellings', () => {
+    expect(parseDurationToSeconds('0h04\'03"')).toBe(243)
+    expect(parseDurationToSeconds('1h00\'00"')).toBe(3600)
+    expect(parseDurationToSeconds('0:04:07')).toBe(247)
+    expect(parseDurationToSeconds('4:07')).toBe(247)
+  })
+
+  it('rejects empty, malformed and out-of-range values', () => {
+    expect(parseDurationToSeconds('')).toBeNull()
+    expect(parseDurationToSeconds(null)).toBeNull()
+    expect(parseDurationToSeconds('abc')).toBeNull()
+    expect(parseDurationToSeconds('1::30')).toBeNull()
+    expect(parseDurationToSeconds('1:99:99')).toBeNull()
+    expect(parseDurationToSeconds('0h99\'00"')).toBeNull()
+    expect(parseDurationToSeconds('1:99')).toBeNull()
+  })
+})
+
+describe('parsePresidingOfficerTransitions', () => {
+  const transitionHtml = (heading: string, times: number[]) => `
+<h4 class="g-l-assista__categoria-outros-videos">${heading}</h4>
+<ul class="g-l-assista__lista-outros-videos">
+${times
+  .map(
+    (tMs) =>
+      `<li><a id="link-trecho-video" href="https://www.camara.leg.br/evento-legislativo/67091?a&#x3D;558641&amp;t&#x3D;${tMs}&trechosOrador=&crawl=no" class="chamada__link-trecho" titulo="X">y</a></li>`,
+  )
+  .join('\n')}
+</ul>`
+
+  it('reads the transition groups, using the newest clip of each as the change moment', () => {
+    const html = [
+      transitionHtml(
+        'Troca da mesa Presidente Bohn Gass por Participante Pompeo de Mattos',
+        [1675815220717, 1675814951403],
+      ),
+      transitionHtml(
+        'Troca da mesa Presidente Pompeo de Mattos por Participante Bohn Gass',
+        [1675814710153],
+      ),
+    ].join('\n')
+
+    expect(parsePresidingOfficerTransitions(html)).toEqual([
+      { from: 'Pompeo de Mattos', to: 'Bohn Gass', tMs: 1675814710153 },
+      { from: 'Bohn Gass', to: 'Pompeo de Mattos', tMs: 1675815220717 },
+    ])
+  })
+
+  it('returns [] when the page has no transitions or the markup changes', () => {
+    const html = excerptHtml(67091, {
+      audioId: 5,
+      tMs: 1000,
+      speaker: 'JORGE SOLLA',
+      party: 'DEPUTADO (PT-BA)',
+      startTime: '17:30',
+      duration: '0h01\'00"',
+    })
+    expect(parsePresidingOfficerTransitions(html)).toEqual([])
+    expect(parsePresidingOfficerTransitions('')).toEqual([])
+  })
+})
+
+describe('resolvePresidingOfficer', () => {
+  // Real transition groups of event 67091 (2023-02-07), newest-clip times.
+  const transitions = [
+    { from: 'Bohn Gass', to: 'Pompeo de Mattos', tMs: 1675815220717 },
+    { from: 'Gilberto Nascimento', to: 'Pompeo de Mattos', tMs: 1675800828283 },
+    { from: 'Pompeo de Mattos', to: 'Adriana Ventura', tMs: 1675804114707 },
+    { from: 'Adriana Ventura', to: 'Pompeo de Mattos', tMs: 1675804616513 },
+    { from: 'Pompeo de Mattos', to: 'Bohn Gass', tMs: 1675814710153 },
+  ]
+
+  it('returns the incoming officer of the latest transition before the moment', () => {
+    // Solla spoke at 17:28 BRT (20:28 UTC) — Gilberto→Pompeo at 17:13 BRT.
+    expect(resolvePresidingOfficer(transitions, Date.parse('2023-02-07T20:28:00Z'))).toBe(
+      'Pompeo de Mattos',
+    )
+    // After Bohn→Pompeo (21:13 BRT), Pompeo presided again.
+    expect(resolvePresidingOfficer(transitions, Date.parse('2023-02-08T00:16:00Z'))).toBe(
+      'Pompeo de Mattos',
+    )
+  })
+
+  it('returns the outgoing officer before the first transition', () => {
+    expect(resolvePresidingOfficer(transitions, Date.parse('2023-02-07T19:00:00Z'))).toBe(
+      'Gilberto Nascimento',
+    )
+  })
+
+  it('returns null without transitions or a usable moment', () => {
+    expect(resolvePresidingOfficer([], 1675801680000)).toBeNull()
+    expect(resolvePresidingOfficer(transitions, Number.NaN)).toBeNull()
+  })
+})
+
 describe('normalizeTranscription', () => {
  it('accepts the OpenAI-compat segments shape', () => {
    expect(
--- a/tests/unit/speechFacets.unit.spec.ts
+++ b/tests/unit/speechFacets.unit.spec.ts
@ -0,0 +1,158 @@
+// @vitest-environment node
+
+import { describe, expect, it } from 'vitest'
+
+import { mergeFacetClassification, parseLlmFacetResponse } from '@/lib/speechFacets'
+import { classifySpeechByGazetteer, matchMunicipalityMentions } from '@/lib/speechGazetteer'
+
+describe('classifySpeechByGazetteer', () => {
+  it('detects topics, explicit scopes and municipality mentions', () => {
+    const result = classifySpeechByGazetteer({
+      summary: 'Investimentos em saúde e educação na Bahia.',
+      keywords: ['SUS', 'Escola'],
+      transcript: 'Defendemos o SUS, a escola pública e a universidade em Feira de Santana.',
+    })
+
+    expect(result.topics).toContain('saude')
+    expect(result.topics).toContain('educacao')
+    expect(result.scopes).toContain('bahia')
+    expect(result.municipalities.map((entry) => entry.slug)).toContain('feira-de-santana')
+    expect(result.people).toEqual([])
+    expect(result.programs).toEqual([])
+    expect(result.projects).toEqual([])
+  })
+
+  it('matches whole words only (no "sus" inside "suspeito")', () => {
+    const result = classifySpeechByGazetteer({
+      transcript: 'O suspeito foi ouvido e liberado.',
+    })
+
+    expect(result.topics).not.toContain('saude')
+  })
+
+  it('infers the Bahia scope from a municipality mention alone', () => {
+    const result = classifySpeechByGazetteer({
+      transcript: 'A cidade de Vitória da Conquista recebeu o investimento.',
+    })
+
+    expect(result.scopes).toEqual(['bahia'])
+  })
+})
+
+describe('matchMunicipalityMentions', () => {
+  it('consumes the longest name (São Félix do Coribe never also yields Coribe)', () => {
+    const slugs = matchMunicipalityMentions('Ele passou por São Félix do Coribe.').map(
+      (entry) => entry.slug,
+    )
+
+    expect(slugs).toEqual(['sao-felix-do-coribe'])
+  })
+
+  it('does not split a longer name into a shorter one (Condeúba is not Conde)', () => {
+    const slugs = matchMunicipalityMentions('A comitiva esteve em Condeúba.').map(
+      (entry) => entry.slug,
+    )
+
+    expect(slugs).toEqual(['condeuba'])
+  })
+
+  it('requires a capitalized proper noun (prose "saúde" is not the municipality Saúde)', () => {
+    expect(matchMunicipalityMentions('A saúde pública é prioridade.')).toEqual([])
+    expect(
+      matchMunicipalityMentions('A agenda passou por Saúde, no interior da Bahia.').map(
+        (entry) => entry.slug,
+      ),
+    ).toEqual(['saude'])
+  })
+
+  it('ignores institutional phrases that contain a municipality name', () => {
+    expect(matchMunicipalityMentions('O Sistema Único de Saúde atende a população.')).toEqual([])
+    expect(matchMunicipalityMentions('O Banco Central elevou os juros.')).toEqual([])
+    expect(matchMunicipalityMentions('O Palácio do Planalto confirmou a agenda.')).toEqual([])
+    expect(
+      matchMunicipalityMentions('Ele esteve em Central, no interior da Bahia.').map(
+        (entry) => entry.slug,
+      ),
+    ).toEqual(['central'])
+  })
+
+  it('maps a Salvador mention to its 19 zone entries', () => {
+    const slugs = matchMunicipalityMentions('O comício foi em Salvador.').map((entry) => entry.slug)
+
+    expect(slugs).toHaveLength(19)
+    expect(slugs.every((slug) => slug.startsWith('salvador-ze-'))).toBe(true)
+  })
+})
+
+describe('parseLlmFacetResponse', () => {
+  it('maps labels and values to the taxonomy, dropping unknown values', () => {
+    expect(
+      parseLlmFacetResponse({
+        topics: ['Saúde', 'seguranca-publica', 'inexistente'],
+        scopes: ['Bahia', 'Mundo'],
+        people: ['Dilma Rousseff', 'dilma rousseff', ''],
+        programs: ['Minha Casa, Minha Vida'],
+        projects: ['PEC 45'],
+      }),
+    ).toEqual({
+      topics: ['saude', 'seguranca-publica'],
+      scopes: ['bahia'],
+      people: ['Dilma Rousseff'],
+      programs: ['Minha Casa, Minha Vida'],
+      projects: ['PEC 45'],
+    })
+  })
+
+  it('maps the long labels too', () => {
+    const parsed = parseLlmFacetResponse({
+      topics: ['Direitos Humanos e Assistência Social', 'Infraestrutura e Transporte'],
+    })
+
+    expect(parsed?.topics).toEqual(['direitos-humanos', 'infraestrutura'])
+  })
+
+  it('returns null for a response without any known key', () => {
+    expect(parseLlmFacetResponse(null)).toBeNull()
+    expect(parseLlmFacetResponse('texto')).toBeNull()
+    expect(parseLlmFacetResponse({ foo: 1 })).toBeNull()
+  })
+
+  it('caps and truncates the free-text mentions', () => {
+    const parsed = parseLlmFacetResponse({
+      people: Array.from({ length: 20 }, (_, index) => `Pessoa ${index}`),
+      programs: ['x'.repeat(200)],
+    })
+
+    expect(parsed?.people).toHaveLength(12)
+    expect(parsed?.programs[0]).toHaveLength(120)
+  })
+})
+
+describe('mergeFacetClassification', () => {
+  const gazetteer = classifySpeechByGazetteer({
+    transcript: 'O SUS e a escola pública em Feira de Santana.',
+  })
+
+  it('unions topics and prefers the validated LLM scopes', () => {
+    const merged = mergeFacetClassification(gazetteer, {
+      topics: ['cultura'],
+      scopes: ['brasil'],
+      people: ['Lula'],
+      programs: [],
+      projects: [],
+    })
+
+    expect(merged.topics).toEqual([...gazetteer.topics, 'cultura'])
+    expect(merged.scopes).toEqual(['brasil'])
+    expect(merged.people).toEqual(['Lula'])
+    expect(merged.classifiedBy).toBe('llm')
+  })
+
+  it('keeps the gazetteer provenance without an LLM response', () => {
+    const merged = mergeFacetClassification(gazetteer, null)
+
+    expect(merged.classifiedBy).toBe('gazetteer')
+    expect(merged.scopes).toEqual(gazetteer.scopes)
+    expect(merged.municipalities).toEqual(gazetteer.municipalities)
+  })
+})
--- a/tests/unit/speechSearch.unit.spec.ts
+++ b/tests/unit/speechSearch.unit.spec.ts
@ -0,0 +1,23 @@
+// @vitest-environment node
+
+import { describe, expect, it } from 'vitest'
+
+import { normalizeForSearch, uniqueByNormalizedForm } from '@/lib/speechSearch'
+
+describe('normalizeForSearch', () => {
+  it('strips accents, lowercases and collapses whitespace', () => {
+    expect(normalizeForSearch('  Saúde   Pública ')).toBe('saude publica')
+    expect(normalizeForSearch('São Félix do Coribe')).toBe('sao felix do coribe')
+    expect(normalizeForSearch('')).toBe('')
+  })
+})
+
+describe('uniqueByNormalizedForm', () => {
+  it('dedupes by normalized form keeping the first spelling', () => {
+    expect(uniqueByNormalizedForm(['Saúde', 'saude', 'SUS', ' sus '])).toEqual(['Saúde', 'SUS'])
+  })
+
+  it('drops empty values', () => {
+    expect(uniqueByNormalizedForm(['', '  ', 'Bahia'])).toEqual(['Bahia'])
+  })
+})