Integrity

Settlement you can defend.
Numbers you can check.

A payout is only as good as the result behind it. So every settled result carries its evidence, its score and the threshold it cleared, and we publish how the system performs, holds and voids included.

Cross-checkedevery result, two signalsHeldanything uncertain is held for reviewNeverauto-pays a conflicted read

Anything uncertain is held for review and checked before it settles, so an unclear result is never settled on a guess. VALORANT: human confirmation required while automatic settlement is under review. Fortnite, League of Legends, Dota 2, PUBG: BATTLEGROUNDS and Rocket League are available and live today. “Available and live today” means the title is offered now for recorded-video settlement, not live-stream settlement. See /coverage#live. Game names are trademarks of their owners, and a listing is not a partnership, sponsorship or endorsement. See /coverage#titles. A clear read on recorded video settles automatically for Fortnite, League of Legends, Dota 2, PUBG: BATTLEGROUNDS and Rocket League. Other titles are in Beta.

Methodology and records →

Why we publish

A record is worth more
than a claim.

Partners

Settlement you can defend.

When a user or a brand questions a result, you answer with the receipt, the evidence and the threshold it was held against, not with a judgment call.

Investors

An auditable quality record.

The method and its limits are published. You can check them instead of taking our word for it.

Brands

Pay out against a verified record.

The partner releases a pool only once its result is settled. A clearance fee, where one is agreed, is charged only on pools that settle; held and voided pools pay nothing.

What we publish

Six measures, in the open.
Holds and voids included.

Auto-settle rate

Results the model settled at or above the threshold, with no person.

Production
Publishing from first production run

Wrong auto-settles

Automatic settles whose outcome did not match the human-checked one.

Production
Publishing from first production run

Hold rate and void rate

Results held below the threshold or sent for confirmation, and results voided with no usable outcome.

Production
Publishing from first production run

Reversal count

Settled results overturned in a dispute and recorded as a new entry.

Production
Publishing from first production run

Calibration

How far the model’s stated confidence sits from how often it is right on human-labelled results.

Production
Publishing from first production run

Certification status

Where each title stands against the graduation criteria, per title.

Current
VALORANT: human confirmation required while automatic settlement is under review. Fortnite, League of Legends, Dota 2, PUBG: BATTLEGROUNDS and Rocket League are available and live today. “Available and live today” means the title is offered now for recorded-video settlement, not live-stream settlement. See /coverage#live. Game names are trademarks of their owners, and a listing is not a partnership, sponsorship or endorsement. See /coverage#titles.

Updated as production results come in.

Certification

These titles
settle automatically.

A title is called certified only when a certification run is linked. None are linked on this page. A title graduates only when its production record clears every criterion below.

Fortnite

Available · automatic on recorded videoWhat Live means

Valorant

Human confirmation required

Dota 2

Available · automatic on recorded videoWhat Live means

League of Legends

Available · automatic on recorded videoWhat Live means

PUBG

Available · automatic on recorded videoWhat Live means

Rocket League

Available · automatic on recorded videoWhat Live means
Graduation criteria, per title
CriterionBar
VolumeEnough production results on the title, within a recent window, to measure it.
AccuracyAgreement with human labels at or above PLLAY’s bar, measured only once enough labelled results exist.
CalibrationStated confidence stays close to how often the title is right.
ReversalsNo reversals in the window.

These criteria cover recorded video for the titles named above. Thresholds and progress counts are not published. No accuracy figure is published here.

Ground truth

Checked against a person,
not against the model.

Blind labelling

Accuracy and calibration are measured against outcomes labelled by a named PLLAY reviewer. Only a person can record a label; a name that looks like an agent or a bot is refused.

The review page opens in blind mode: Vision’s result and the labelling agent’s suggestion stay hidden until the reviewer picks the outcome they see in the clip. A label that differs from the suggestion needs a written note.

Blind mode is on by default and a reviewer can switch it off. The benchmark’s human checks (as of 2026-08-10) were made before this review page existed. Production labels will appear here as they are exported.

The rule

If evidence is thin,
we hold.

Rule

If evidence is insufficient, the event is held for review or voided.

Recorded video settles on a clear read. PLLAY settles from recorded match video.
PathWho settlesBelow thresholdReviewerHuman-override %
VODModel, automated at or above thresholdheldNot requiredPublished on the tape
LivePLLAY settles from recorded match video.heldPLLAY settles from recorded match video.Not offered

The server sets the threshold. Create does not accept it. Held is a waiting state, not an error.

Receipts and record

Every settled result
gets a receipt.

Receipt
settlement_idUnique id for the outcome. Kept after a reversal.
outcomeThe result that was settled.
confidence · thresholdThe model score and the threshold it was held against. A settle below the threshold is refused.
evidence_uriThe evidence window: clip, frames and signals the call was made on.
model_versionThe exact model build that produced the call.

Each result is recorded as an entry. A correction is a new entry, not an edit: a reversal gets its own id, and the original settlement_id is kept. Held and voided results carry a reason instead of a settlement_id.

Public hash anchor: Off. The code exists, no chain is configured and nothing is anchored. Funds do not settle on-chain. A receipt records a decision. Oracle returns a result; it does not move money.

Evidence bundle

Each read links its evidence.
Check it yourself.

Attached to each read. A hold with no read has none.
framesThe clip window the call was made on
signalsDetector outputs and rule inputs that fired
confidenceScore and the threshold it was held against
model_versionExact model build that produced the call
reviewer_idPresent when a human confirmed or overrode
sourcesPlanned: optional corroborating inputs (secondary clip, official feed, sensor). When enabled, agreement and conflicts are recorded on the receipt.
settlement_idUnique id for the outcome; the receipt
Disputes

Contest a call.
An overturn writes a new record.

Within 72h

Dispute

The partner files through the API (POST /v1/events/{id}/dispute) or emails support@pllay.io, within 72 hours of the result. One dispute per result.

Target 4h

Acknowledge

Acknowledged. A PLLAY reviewer who did not make the original call takes it and checks the evidence record.

Target 48h

Decide

Decided with a written reason. Upheld: no change. Overturned: PLLAY overturns the result and issues a correction.

On overturn

Record

A new record is written; the original is never edited. Settlement API: event.reversed (reversal_settlement_id). Partner API: resolution.corrected (correction_id).

These times are the target process. They are not a contractual guarantee.

States
openlockedheldsettledreversedvoided
Held: the read is not clear, or the evidence is insufficient. A waiting state, not an error: wait for event.settled or event.voided.
Next step

Settle your first 500 results free.

Put PLLAY’s settlement layer under your own arcade, contests or predictions. 90-day launch season, then $1 per settled result. 100 settled results are free every month, and there is no monthly spend cap. You keep users, funds and KYC.

Request a keySee pricing

Ready to see how results are verified for your audience?