From raw inputs to reviewed daily digest.
The MVP pipeline is deliberately post-market. It optimizes for correctness and reviewability, not speed.
Daily flow
Ingest market data. Normalize daily OHLCV and delivery fields. Reject invalid OHLC, bad dates, and duplicate listing-date rows.
Ingest filings. Classify event category, calculate materiality, extract key values, and list missing values such as margin or execution timeline.
Compute moves. Calculate return, market-relative return, sector-relative return, z-scores, volume anomaly, and interestingness score.
Build evidence items. Convert market and event facts into claim-ready units with source names, timestamps, confidence, and evidence IDs.
Build evidence packet. Attach asset metadata, move diagnostics, events, peer/sector context, hypotheses, allowed claims, blocked claims, and confidence recommendation.
Generate draft card. Produce structured sections only from the packet. Generated cards default to needs_review.
Verify. Check banned language, causality overstatement, unsupported numbers, evidence references, uncertainty, and sources.
Human review. Reviewer approves, rejects, or edits. Edited approvals rerun the verifier.
Publish digest. Only published cards appear in public card lists and daily digest output.
Batch endpoint
POST /internal/pipeline/post-market
{
"date": "2026-04-25",
"market_return_pct": 0.8,
"sector_returns": { "aviation": 1.4 },
"minimum_interestingness_score": 60,
"max_cards": 20,
"candidates": []
} Output
- Top unusual moves for the day.
- Evidence packets for generated candidates.
- Draft cards that remain blocked from public views until reviewed.
- Verifier reports and audit artifacts.
- Daily digest grouped by card type.