99.8% limiting public rate
Operation, exact request, committed answer, and exact artifact passed at 99.8%; the other four quantity rates were 100%.
EXP-004 · settled 2026-07-19
Direction changed.
Gates held.
The global replay-tangent projection carried seed 2 through 700 full-scale commits, then the selected public checkpoint and sealed promotion split both passed.
Prospective go
Committed700 / 700
SelectedUpdate 500
Promotion splitPassed once
Decision
The intervention reopened the constrained learning path.
All 700 optimizer attempts committed at full scale. None backtracked or rejected. The tangent projection changed 423 candidate directions; 277 candidates already had a non-increasing tangent and stayed unchanged.
The largest directly measured committed replay increase was 0.050892%, far below the unchanged 1.5% authority. Seven public checkpoints were evaluated and six were jointly feasible. Update 500 entered the frozen frontier and remained selected when training stopped after 200 further commits without a Pareto improvement.
Selected evidence
Public pass, then one sealed test.
1.183323% replay
The selected checkpoint stayed below the frozen 2% public replay ceiling with a 1-point minimum faculty margin.
99.6% limiting promotion rate
The disjoint promotion split opened exactly once and passed every quantity gate with zero rejected-state mutations.
Projection and trajectory
| Metric | Value |
|---|---|
| Attempts / commits / trial evaluations | 700 / 700 / 700 |
| Full-scale / backtracked / rejected | 700 / 0 / 0 |
| Projected / unprojected trials | 423 / 277 |
| Maximum / mean displacement removed | 14.337324% / 3.673116% |
| Maximum absolute post-projection tangent dot | 0.000000621835 |
| Public / jointly feasible checkpoints | 7 / 6 |
Forecasts settled to go
| Forecaster | Frozen go / no-go / failure | Brier |
|---|---|---|
| Tangent-mechanism model | 32% / 66% / 2% | 0.8984 |
| Q2.3–Q2.5 empirical trace model | 16% / 81% / 3% | 1.3626 |
Both independent forecasts put their largest probability on no-go before execution. The prospective go is therefore a material update, not a retrospective fit.
16 objects
All content-addressed ledger objects verified.
17 events
The contribution-to-settlement chain verified.
8 source hashes
Contract, result, logs, model, checker, and driver are pinned.
Evidence
Exact commits, exact artifacts, one live run.
Execution used frozen Zero commit 412ab70a. Observed artifacts and the selected quantized model are published at result commit 94c7e1d0. ilXyr prospectively froze the experiment at aac5197e and resolved run 09400765… as go. This contract did not require an external OSF preregistration.
Read the machine-readable evidence →
Inspect the upstream result →
Follow-through
The family replication passed.
The seed-2 go earned EXP-005. Seeds 1 and 3 later executed independently and both resolved go. Q2.6 cleared its frozen multi-seed family rule, making the seed-2 artifact promotion-eligible as ZERO.4.