Vision reads the outcome.
Thin evidence holds.
Vision reads a recorded clip. At or above the threshold it resolves on its own. Below it, the result is held for a reviewer. Vision proposes the outcome from live video and a reviewer confirms it. Every live result carries a reviewer_id, and one without it is refused.
The recorded-video benchmark
| Title | Clips | Matched expected | Auto-settled | Wrong auto-settles |
|---|---|---|---|---|
| Madden NFL 26 | 3 | 3 | 1 | 0 |
| NBA 2K26 | 2 | 2 | 2 | 0 |
| Counter-Strike 2 | 3 | 3 | 1 | 0 |
| Dota 2 | 2 | 2 | 1 | 0 |
| Fortnite | 3 | 3 | 3 | 0 |
| Call of Duty: Warzone | 2 | 2 | 2 | 0 |
| Valorant | 2 | 2 | 2 | 0 |
| Rocket League | 2 | 2 | 2 | 0 |
| Apex Legends | 2 | 2 | 2 | 0 |
| League of Legends | 4 | 4 | 4 | 0 |
| Rainbow Six Siege | 3 | 3 | 2 | 0 |
| PUBG | 2 | 2 | 2 | 0 |
| Overwatch 2 | 5 | 5 | 4 | 0 |
| Rocket League Sideswipe | 3 | 3 | 3 | 0 |
- Each clip was run through the full Vision pipeline from scratch, with no fallback.
- Each clip has a ground-truth outcome checked by a person. A hold counts as correct only where a hold was the expected result.
- The headline check is a wrong auto-settle: an automatic settle with an incorrect outcome.
- Recorded video only. There is no live-video result in this set.
- Public clips chosen by the team, 2 to 5 per title. Not a random sample and not production traffic.
- It shows the pipeline on these clips. It is not an accuracy rate for any title and not a production test.
You set the bar per event.
Set on create
The partner sets threshold (0 to 1) on each event when it is created. 0.9 is the value in our examples, not a default.
At or above
Recorded video settles on the model.
Below
The event holds. A settle below the threshold is refused (422 below_threshold).
Vision reads the recorded clip frame by frame. It settles only when the readings agree. If they disagree it holds for confirmation, and if the clip is unreadable it voids. It never guesses and never settles silently. The settlement contract records the decision as a confidence against a threshold.
No guess, no payout.
held
Nobody is paid until it is resolved. A reviewer is next.
voided
The clip is unreadable or there is no winner on frame. The pool refunds.
If evidence is insufficient, the event is held or voided. We do not guess.
Recorded video first. Live with a reviewer.
| VOD | Live | |
|---|---|---|
| Who settles | Model, autonomous at or above threshold | Model proposes; reviewer confirms |
| Below threshold | held | held |
| Reviewer | Not required | reviewer_id required on every settle |
The live path always requires a reviewer_id: settling a live event without one returns 422 live_unreviewed. Certified automated live is not available today (Phase 2).
Phase 2 is planned and not available today: live verification on the remaining titles, and certified automated live with no reviewer. Automated live is gated on authenticated production tests on record. Phase 2 has no date.
Contest a call on the record.
Contest
Partner flags the settlement_id in the Console or via legal@pllay.io within 72h of event.settled.
Acknowledge
Integrity reviewer assigned. Evidence bundle re-run against the recorded model_version.
Decide
Upheld: no change. Overturned: POST /v1/events/:id/reverse.
Ledger
A compensating credit is written as a new record referencing the original settlement_id. The original is never edited. event.reversed fires.
These are target times, not a contractual guarantee.
Live reviewer staffing and response times are scoped per deployment.
Recorded-video intake is built and gated to enabled partners with an active plan. It is not launched for partners yet.