Rumor-driven actions and checkpoints
Qualifying rumor signals rerun the normal lineup or waiver rules. Use this page to see why a signal produced an action, no move or a deferred retry.
Which signals can act?
| Rule | Behavior |
|---|---|
| Read order | Increasing signal ID after the team's durable watermark |
| Confidence | At least RUMOR_ACTION_CONFIDENCE_THRESHOLD; default 0.75 |
| Duplicate player/action | Latest qualifying signal in the batch wins |
| Low-confidence/non-action rows | Still account for them so they cannot stall the cursor |
| Suggested action | Evaluation | Published subset |
|---|---|---|
boost_waiver_priority | Normal live waiver pools, >3 margin and IR gates | Moves adding a signaled player |
bench_watch | Shared pregame lineup evaluation | Swaps removing a signaled current starter |
Rumors add no score bonus and force no move. A qualifying signal with no normal-rule improvement becomes no_beneficial_move; unrelated moves wait for scheduled evaluation.
Decision evidence adds signal ID, sentiment and confidence to VORP/source evidence. Missing inputs and identities are separate outcomes from no improvement.
What starts a run?
| Trigger | When | No-benefit signal |
|---|---|---|
| Schedule | Daily 21:45, or the tuesday-weekly chain's last step | Settled |
rumor.signal_created event | After the 11:00 digest, once the catalog subscribes the job | Held for a later run |
| Manual or HTTP | On request | Settled |
The digest emits the event because it writes rumor_signals, the table this job reads. sync_rumors only writes raw_signals, so it emits nothing. The schedule stays as the fallback, because an event is lost if the broker doesn't take it.
An event run reads VORP from the previous evening's refresh, which may not reflect the news behind a signal yet. It still publishes any move that clears the normal rules. A signal with no move stays behind the watermark instead of settling, so the 21:45 run evaluates it on fresh VORP. Signals settled in the meantime aren't evaluated or published again. Once a held signal is 24 hours old, the next run settles it whatever its trigger, so a paused or missed 21:45 run can't stall the cursor. signals_held in the outcome details counts the signals that run held, so a second event run on the same day re-evaluates and counts the same signals again.
What survives a restart?
| Checkpoint field | Purpose |
|---|---|
| Team and version | PostgreSQL ownership; optimistic version check rejects overlapping updates |
| Watermark | Last contiguous fully accounted signal ID |
| Signal progress | Preparation, action links and recorded reasons |
| Prepared actions | Immutable payload, stable gateway key, attempt times and outcomes |
| Failed-job markers | Retain job failure state for seven days |
Preparation commits before HTTP publication. A crash after delivery replays the same key; database transactions never span gateway calls.
The legacy Redis rumor_action:last_processed_id initializes the first team checkpoint, then becomes a best-effort mirror. Pending work stops the watermark; accepted and already-queued outcomes both complete an action.
When does retry stop?
| Work | Limit |
|---|---|
| Publication | Three failures |
| Signal evaluation | Three failures and 24 hours. A held signal has no time limit here, because a hold is not a failure |
| Waiver action | 24-hour retry window |
| Lineup action | 20-minute retry window; gateway key lasts 30 minutes |
| Different season/week | Drop prepared action rather than replay |
| Saved IR + add | Recheck AUTO_PUBLISH_IR_MOVES at delivery |
Terminal failures retain reasons in execution evidence. Completed checkpoint entries can be pruned below the watermark.
How do you inspect the outcome?
| Result | Meaning |
|---|---|
| Returned claims/swaps | Newly published actions, plus watermark |
success | All work accounted for, including no-op or duplicate outcomes |
partial_failure / failure | Unaccounted work, missing identity or publication failure |
| Already queued | Gateway recognized an earlier publication; not proof of ESPN execution |
Inspect signal/action counters and drop reasons in Activity before attempting recovery.
Related
- Lineup rules
- Waiver and IR gates
- Implementation: rumor_action_service.py, rumor_action_checkpoint.py