The daily cycle #

The engine has two kinds of clocks. Market-sensitive jobs use America/New_York, so they stay tied to the opening bell when daylight saving changes. Nightly and weekend drivers use UTC, so their New York wall-clock time moves by one hour between Eastern Daylight Time (EDT) and Eastern Standard Time (EST).

The table below shows New York time. A slash means “during EDT / during EST.”

“Tuesday–Saturday” in the overnight rows is the UTC calendar. In New York, those runs belong to the preceding Monday–Friday market session.

Before and during the session #

The pre-open check exists because an overnight decision can be stale by the opening bell. It sees a bounded update and answers only keep or cancel. A failed or unavailable check cannot manufacture a new entry.

The event cycle begins after the market has opened. Its unusual minute pattern is exact: 9:35, 9:50, then every 15 minutes on :05, :20, :35, and :50 until 3:50. It reads facts whose availability time is no later than the decision cutoff. These event decisions are shadow evidence; they do not trade intraday.

The hourly and four-hour observers ask related questions at fixed New York times. Missed windows are not replayed. Replaying them later would let the observer see facts that were unavailable at the original time. Each observer holds one database connection for the complete run, including its retained quote and cross-check writes, and waits no more than 60 seconds for that connection.

After the close #

The main nightly driver is deterministic. It refreshes the universe, collects prices, chooses the latest session with broad enough coverage, screens candidates, reconciles actions, advances the paper league, updates frozen forward records, and queues supporting work. Each stage records its start, finish, duration, and exit status both in the normal log and in logs/stage-timings.jsonl; tools/stage_timings.py summarizes recent runs.

After the league and sync, the independent price verifier materializes its database slice and then uses four network workers behind one global request pacer. Its network phase runs alongside the farm drain. Earnings collection is bounded to unknown, near-term, or seven-day-stale names on an ordinary night, with a full eligible-universe pass each Monday. Repeated intraday bars with unchanged payloads reuse their first fact revision, and new facts are written as batches.

The AI work comes later. First the engine retains inputs and source receipts. Then the daily opportunity agent runs. Half an hour later, candidate-wide scoring runs with matched rule scores and refreshes the paper-book and evaluation views. Scoring re-renders the league before it validates evidence. Evidence failure exits 75 after that render, and the service does not restart on that status. Keeping these jobs separate makes their information cutoffs and failure states visible.

The historical research archive also runs in bounded slices. On weekdays it has five fixed-UTC slots; on weekends it has six four-hourly slots. A caught-up TradingView symbol waits for two to four completed sessions and requests them together at the unchanged request rate; long backfills keep their existing chunk bound. Because this archive is retrieval-time research data, it never overwrites operational prices or prices a simulator fill.

Read-only snapshots follow producer cadence rather than every timer indiscriminately. Event runs publish at most once per 120 minutes, pre-open publishes none, and nightly, scoring, and TradingView archive runs remain unthrottled. When no write-ahead log exists, the snapshot path durably copies the raw database and verifies it; a WAL uses DuckDB's consistent copy path.

Weekends #

The deterministic weekend jobs also use UTC:

Every scheduled driver has an advisory lock and a completion grace period. A status page can say running, interrupted, stale-running, or overdue, but it does not silently restart a job. Failures are investigated on demand by an agent. There is no push alerting.