High Signal Podcasts (on-record)

Source-backed index of public statements by notable people. High Signal sub-product, own layout, live at podcasts.highsignal.app. The unit is the claim, not the episode.

Every published claim has a verbatim transcript excerpt, explicit attribution status, date, and source link. Identified speakers are linked to people; unresolved speakers are labeled and excluded from people counts. Timed YouTube links are limited to claims extracted from that video's captions. Missing evidence is preferred over a confident guess.

Stack

• API: Hono on Cloudflare Workers (workers/api)
• DB: Cloudflare D1 + Drizzle (packages/db)
• Raw storage: R2 bucket on-record-raw
• Ingest: Python / uv (python/ingest), GitHub Actions cron
• Web: Astro SSR at podcasts.highsignal.app

Quickstart

pnpm install
uv sync --project python/ingest --dev
pnpm db:migrate:local
pnpm --filter @on-record/api dev   # wrangler on :8787
pnpm dev:web                       # Astro SSR on :4321
pnpm ingest -- --stage extract --focus recs
pnpm ingest -- --stage recover-speakers --dry-run
pnpm ingest -- --stage extract # batches 20; no per-episode claim ceiling
pnpm quality

Copy .env.example to .dev.vars (worker) and .env (ingest). Do not commit secrets.

Production D1 on-record-db and R2 on-record-raw already exist. Deploys are dispatch-only (.github/workflows/deploy.yml). New remote migrations still need operator approval.

Trusted production corpus

Production contains 25 shows and 10,325 episode rows. The public trusted 23-show subset exposes 8,414 catalog episodes, 1,209 transcript episodes, 11,624 published claims from 935 people across 1,190 source episodes, and 294 quote-safe named-reference evidences in 281 canonical groups. D1 retains 300 reference rows before the public quote-safety pass. Of the transcript episodes, 1,092 (90.3%) have at least 10 claims and 19 have none. TBPN and Odd Lots remain stored but are withheld from all public routes because their diarized speaker labels are not safe to map to people yet.

Build the narrow reviewed v9 bundle with pnpm release:build-reviewed, or the full trusted corpus bundle with pnpm release:build-trusted. The trusted bundle is an incremental upsert with no data-deletion statements and overlays the v9 manual review decisions before adding the broader exact-evidence corpus. Production deployment remains dispatch-only.

Qualified release candidate

The 2026-08-29 v10 snapshot is the qualified production release source. It contains 30,561 public claims across 1,208 transcript episodes and 956 verified people. Coverage averages 25.3 claims per episode; 1,177 episodes have at least 10 claims, 753 have at least 20, and 309 have at least 30.

The named-item index exposes 1,144 evidence rows in 963 groups, including 539 book evidences grouped into 416 titles, 199 app evidences in 149 groups, 51 tool evidences in 50 groups, and 194 other evidences in 191 groups. Book counts are distinct verified people, while 15 book evidences from unverified speakers are labeled separately. Both focused book queues rerun with zero remaining extraction candidates, and integrity checks are clean for primary evidence, quote duplication, attribution, and public cross-kind item duplication.

Qualification does not imply deployment. Source, CI, migration, import, deployment, and public-verification receipts are tracked separately in HANDOFF.md.

The broader expansion uses the separate v10 working snapshot. Extraction has no default per-episode ceiling: strong complete claims can use deterministic exact-excerpt rules, while named recommendations and ambiguous candidates keep the model gate. Claims whose transcript does not establish the speaker carry speaker_unverified; they remain searchable but never count as a person. The responsive research UI and API trust boundary are qualified and live. The previous release used a time-travel backup, uploaded all 1,208 reviewed R2 objects, applied an incremental no-delete D1 bundle, and deployed API and web through the dispatch-only workflow.