Architecture

Trajectory Memory Ledger separates durable ingestion from research and training. 1. Read `events-YYYY-MM-DD.jsonl` 2. Skip envelopes already covered by the date-scoped cursor 3. Group gateway envelopes by `flow_id` 4. Build one trajectory card per completed flow 5. Normalize the card to schema v2 6. Score it with the six-signal reward model 7. Append to the JSONL ledger under an exclusive file lock 8. Write cursor and Prometheus metrics atomically Daily event files may restart `seq` at 1. A global monotonic cursor would silently drop events after a date rollover. - batch extraction - SFT export - train/validation splits - reward ablations - selection experiments - model training and evaluation - paper metrics This split keeps the operational collector reliable while leaving research tooling flexible.