Where does the data come from?

Three sub-pipes inside the TradingView app's app/market_data/ module:

• OHLCV bars via yfinance — daily refresh for every watchlist ticker (240+ bars). Lands in ohlcv_bars (composite PK on symbol+interval+ts).
• Macro series via yfinance + FRED API — daily 01:00 UTC, gated by MACRO_ENABLED. Lands in macro_series.
• Earnings calendar — rolling 150-row universe (NASDAQ + Street Tier-1+2 + 8-K Item 2.02 confirm), 90-day TTL. Lands in earnings_calendar_rows.
• IV percentile + earnings dates per ticker upserted into ticker_market_data.

Door consumer: Lakshmi only.

The legacy unversioned POST /webhook endpoint receives Pine Script alerts as text. These are notification-semantic — no dedup, fast lane.

• Lands in alerts table
• Fans out to tv_context_items with kind=webhook (7-day TTL, deduped by SHA256 within a rolling window).

Door consumer: Lakshmi only.

This is not the webhook pipe. When the operator pastes a chart screenshot into the app's TV Context route (POST /v1/tv-context/screenshot), a vision pipeline kicks in:

1. Tesseract OCR extracts tickers
2. Claude vision summarises the chart
3. Qwen2-VL extracts structured chart references
4. Sidecar markdown written to ~/Documents/knowledge-vault/Sources/.../<ticker>_<HMS>_<id>.md
5. Unknown tickers flow into ticker_review_queue

Door consumer: Lakshmi only. (Per CLAUDE.md demo-discipline: "webhooks and screenshots are two pipes — document them as such.")

The vault-indexer's youtube_channel.py ingester polls every channel listed under Videos/<author>/_channel.yaml. Cadence: hourly, gated by VIDEO_INGEST_ENABLED.

For each new video, a 3-stage pipeline runs:

1. Frame detection — sample candidate keyframes
2. Tesseract OCR — pull text overlays
3. Qwen2-VL (MLX on Apple Silicon) — caption frames + extract structured chart YAML

The output is a markdown at ~/Documents/knowledge-vault/Videos/<author>/<week>-<slug>.md. Unknown tickers seen in vision frames go into ticker_review_queue.

Door consumer: any door whose vault scope includes Videos/<subdomain>.

The vault-indexer's ingest_edgar.py polls Edgar daily for watchlist tickers, gated by EDGAR_INGEST_ENABLED. Writes per-filing markdown to ~/Documents/knowledge-vault/Filings/<ticker>/<accession>.md. Idempotent on accession number.

Door consumer: Lakshmi only (finance vault scope).

An aggregator pipeline that imports tier-1/tier-2 conviction lists, politician buys (House STOCK Act, Senate EFD), insider buys (Form-4), and options flow.

Each snapshot is dated: The Street/snapshots/YYYY-MM-DD/. Files: tier-1-conviction.md, tier-2-conviction.md, politicians-buys.md, insiders-top.md, options-bullish.md, plus _index.md and methodology.md.

Chunking unit: per-ticker H2 section (~600 tokens by design).

Status: all 6 data sources currently enabled: false in _source.yaml; manual snapshots exist (last: 2026-05-08).

Every workout the operator does in the Lovable verocity app writes one row to Supabase workout_logs. Fields include planned vs actual sets, RPE, total_seconds, avg_hr_bpm.

Derived signals are computed by a SQL view v_drift_signals with seven components (pace, block_lag, gap_breach, load_drop, rpe_drift, recovery, conditioning).

Door consumer: Zeus.

Append-only file at Zeus/03_execution/measurements.yaml. Operator writes bodyweight + waist circumference weekly.

A stdlib helper body_recomp_snapshot.py reads this and emits a body_recomp_sub JSON signal that becomes the 10% second-layer weight on the fitness drift composite (D-022).

Door consumer: Zeus only.

Ganesh has no DB. Signals come from heterogeneous markdown via a single helper:

• Ganesh/02_library/active_sprints.md — current slot deadlines
• Ganesh/_inbox/ — unprocessed drops (7-day rotation)
• Ganesh/03_execution/learning_log.md — append-only session log
• Ganesh/02_library/notes/ — processed notes with applied_in: graduation tracking

The helper ganesh_state_snapshot.py reads the markdown and emits a 5-component drift signal JSON.

Door consumer: Ganesh only.