Every Fusion run ends the same way: the blinds come off. While judges score, their packets carry no engine, provider, or model names; the only identity map sits with the run lead, inside the parent run manifest. When the run closes, an outcome survey unblinds every role, and those attributed results accumulate into the win ledger: a strength record kept per engine, per provider, and per model.
That two-step order is the design. Blind verdicts are fair; attributed outcomes are useful. The ledger exists because we insist on both, in that sequence, and it feeds a concrete decision we make on every future run: which agents to seat in which lanes for which kinds of work.
Blind while the run is live
During a run, a judge sees an identity-free packet. No engine names, no provider names, no model names, nothing that hints at who wrote the diff being scored. We keep it that way because blind packets keep judges honest: a verdict about code should be a verdict about the code. The one exception is the run lead, who holds the identity map in the parent run manifest so the run can be reassembled later. Nobody else can look it up, and nobody needs to.
The survey that lifts the blinds
After the run closes, an outcome survey walks through every role and attaches names. The lead, each candidate, each judge, the integrator, QA, ship, closeout: all of them get unblinded. A verdict that read "this candidate fails this criterion" during the run becomes a fact about a specific model from a specific provider on a specific engine. The judging is already finished, so the attribution cannot bend it.
What the ledger accumulates
Each unblinded run adds to the record at three levels:
| Level | Question it answers over time |
|---|---|
| Engine | Which agent runtimes hold up on our work |
| Provider | Which vendor's stack delivers |
| Model | Which specific model wins which kind of task |
We call this the win ledger. It is not a leaderboard for its own sake; it is roster intelligence. When we configure the next run, the ledger tells us which agents have earned a lane for this kind of work and which have not. And because attribution comes from the manifest rather than from hunches, we never guess which model wrote the code; a ledger built on guesses would be worse than no ledger at all.
Why the order cannot flip
The sequence is the point, and each half protects the other.
Attribute during judging and fairness dies. Judges start scoring reputations instead of diffs, and everything downstream of the verdict, including the ledger itself, inherits the bias. Stay blind forever and learning dies. Verdicts evaporate at closeout, and the next roster gets chosen on brand impressions instead of recorded outcomes.
Where this goes
As the ledger deepens, roster seating stops being a matter of taste. We want the question "who should compete on this slice" to have an evidence-backed answer that cites past runs the way a receipt cites criteria. The blinds exist so that when they come off, what gets written down is worth keeping.