Four modes. One of them is real.
Autopilot is how much of your stream you hand over. Today you can hand over nothing — every prediction on PLLAY has been run by a person. Here is the ladder we are building, and exactly where we are on it.
How much you hand over
Manual
You decide everything.
You write the question, you open it, and you say what happened. Nothing moves without you. This is how every prediction on PLLAY has ever been run.
Copilot
It drafts. You approve every time.
The system watches for a moment worth asking about and drafts the question. Nothing reaches your audience until you press go, and nothing settles until you confirm it.
Not yet — The permission layer this needs is built and switched off. What is missing is the part that proposes — no agent has ever drafted a question on PLLAY.
Guarded Autopilot
It acts inside limits you wrote. It stops at the edge.
You write the limits in plain language — which games, how often, how large. Inside them it runs without you; at the edge of them it stops and asks. Money and settlement stay behind the same rules either way.
Not yet — The rules engine that would enforce your limits exists in the database and has never evaluated a live action.
Full Autopilot
It runs the whole loop while you play.
The end state: your stream generates its own events, resolves them against evidence, and pays out, with you reviewing afterwards rather than during.
Not yet — Not built, and not currently being built. We will not describe this as available until a creator has actually run a stream on it.
What a model is not allowed to decide
Everything in the left column is enforced by the database, not by a prompt. It holds whatever a model outputs, and whatever a bug in an agent tries to do. That is what makes the ladder above safe to climb: a wrong model output is a bad suggestion, never a bad payment.
Decided by rule
Who is allowed to do what
Row-level security, per row, per user
When entries close
Server clock, not the viewer's device
Who is eligible to enter
Policy checks that run before the entry is written
Every movement of credit
Double-entry ledger; balances are derived, never edited
Whether a result may settle
A state machine that rejects illegal transitions
Compliance and payout gates
Deny by default, in the database
Decided by a model
Noticing a moment worth asking about
Detection
Drafting the question and the choices
Generation
Suggesting when to run one
Recommendation
Reading the scoreboard after a match
Computer vision, on the games it has been calibrated for
Every one of these is a suggestion to a person or a proposal to a rule. None of them moves credit on its own.
What is actually running
The part of this that works is the reading. On 14 games it has been calibrated for, the resolver watches the end of a match and either reads the result or declines to. Across 38 recorded trials it settled 31 and declined 7, with 0 wrong settlements. Those trials are recorded video, not live broadcast — it has never been pointed at a live stream.
It reads scoreboards. It does not watch a person and judge what they did. When it cannot read a result cleanly it hands the decision to you, and that path is not a fallback we hope to remove — it is the reason the rest is safe to run.
Credits, not dollars — PLLAY is pre-launch and cash rails are off.
See every trial in the recordWhat each label means
These badges appear across the site. They are the same six everywhere, and they mean this:
- Live
- Working today for any qualified user, self-serve.
- Beta
- Working, with real users, behind a flag or a waitlist.
- Pilot
- Running with named partners only, not generally available.
- Coming Next
- Actively being built. Not usable yet.
- Simulated Demo
- A demonstration of the concept. Not running on real data.
- Enterprise Roadmap
- Planned for enterprise engagements. Not built.
The mode that exists is the one you can start with
Run your first prediction the way every prediction on PLLAY has been run — by deciding it yourself. You will be the reason the next mode up is worth building.