Continual retests
Hot-swap which retests count toward the fold, how deep the lane reaches down the ranking and where it draws the bottom edge, completed-wave aggregation, spare-capacity retests, and how the lane yields to an open benchmark rollout — no platform redeploy. The fold rule changes what validators weight; strict is the audited rollback path. Every change is an append-only revision that stores the whole policy.
Current scoring policy
Controls how deep the retest lane reaches, when completed waves affect rankings, whether evidence ties pool their occupied emission slots, and whether spare validator capacity may advance the next bounded wave.
- Revision
- 0
- Fleet gate
- Not ready
- Aggregate fold
- Inactive
- Tie + ceiling payout
- Disabled
- Retest lane
- Running
- Retest cohort
- Top 5→ 0 admittedof 0 ranked
- Propagation
- Within 5s
How tied emission slots pay
Pooling keeps the same total miner weight and the same emission-set membership. Exact effective-score ties pool immediately; a non-exact tie needs aligned shared-seed evidence and must remain inside the paired uncertainty band. Missing evidence cannot widen a payout group.
Which retests count toward a seed
This changes what validators weight. The continual mean behind official_composite is taken over the seeds this rule admits, so widening it widens the estimator, re-orders the tail, and moves emission shares. Strict is the historical fold and the rollback path — one audited revision, no redeploy.
How deep to retest
5 is the emission set — the agents that earn. Going deeper rescores ranked challengers on the same wave seeds, so one arrives at the top 5 with confirmation depth already banked instead of starting a fresh sweep. Emissions, the weight fold, and wave completion stay on the top 5 at every setting; the extra members ride each seed only once the emission set is claimed.
Only 0 ranked agents exist on the active benchmark right now, so the cohort is all of them until the field grows.
Where to draw the bottom edge
A rank cutoff cannot express a tie: rank n is retested and rank n + 1 is not, even when the two hold the same composite and only first-seen separates them. Tie tolerant keeps the same cutoff and then admits anyone below it whose composite is within the band of it, so the band narrows on its own as retests accumulate.
During a benchmark rollout
Retests rescore the active version. While a rollout collects, they compete for the same validator slots the incoming cohort needs to reach quorum.