Skip to content
Validation

What the evidence actually shows.

Vision AI analyzes broadcast evidence and recommends or verifies outcomes when confidence is sufficient. It has been validated on recorded video. Automated live-stream settlement is currently in private validation. When an outcome cannot be confirmed, PLLAY holds or refunds rather than guessing.

The bench record

Every trial, including the ones it declined.

38
Recorded broadcasts
31
Settled automatically
7
Declined to settle
0
Wrong settlements

Across 14 game titles. These are recorded broadcasts run through the full pipeline — not live traffic, and not a sample of production activity. No pool has ever been settled by this system on a live stream.

What it did on each trial

Four outcomes. Only one moves money.

The interesting number is not how often it settled. It is what happened the rest of the time — because a system that always produces an answer is a system that guesses.

AUTO

31

Confidence cleared the bar. The outcome is settled.

Moves money

CONFIRM

1

A candidate outcome exists but the signals disagreed. A person confirms it.

Does not move money

VOID

2

The evidence establishes that no valid outcome occurred. Stakes are refunded.

Does not move money

HOLD

4

No candidate outcome could be read at all. Nothing settles and nothing is guessed.

Does not move money

Method

How a reading becomes a decision.

Optical character recognition on the broadcast itself

Every trial re-runs the full pipeline on the video — frame extraction, OCR, then reconciliation of the readings against each other. There is no fallback path that guesses when OCR fails.

One adapter per game, not one model for everything

An adapter describes where a specific title puts its score, its clock and its end-of-match banner. 22 adapters exist on disk; 15 have bench rows behind them. A title with no adapter is not supported — it is not "lower accuracy", it simply does not run.

The tier is the output, not the outcome

The pipeline emits a tier and a reason. It does not call a settlement function. What acts on that tier is a separate gate, and on live streams today that gate is a person.

A decline is a success

Declining to read an unclear broadcast is the designed behaviour, not a miss. That is why the counts below are reported per tier rather than as a single accuracy percentage — a percentage would score a correct refusal as a failure.

Not validated

What this evidence does not cover.

Stated as plainly as the record above, and for the same reason. A validation page that only lists successes is marketing wearing a lab coat.

Live streams

Private Pilot

Every trial is recorded video. Automated settlement on a live broadcast is in private validation and is not switched on.

Anything that is not a game broadcast

In Development

The adapters read game scoreboards and end-of-match banners. Reading a person on camera — a rep, a shot, a physical outcome — is not built and has no bench rows.

Titles without an adapter

In Development

7 adapters exist on disk without bench rows behind them, and any title with no adapter at all does not run. Coverage is per-title, not general.

When an outcome cannot be established, PLLAY holds the pool or refunds every stake. A bad read is designed to cost PLLAY a settlement rather than cost a player their entry.

Related: Platform, Trust and Vision AI.

Partner With PLLAY