Baseline
Arithmetic mean of six fixed replay-window losses evaluated by immutable ZERO.3.
EXP-002 · completed 2026-07-19
Budget the trajectory.
The cumulative guard held its boundary, then closed the learning path: 66 commits, eight consecutive rollbacks, no public checkpoint.
No-go
The frozen question
Can a cumulative direct replay guard keep public replay regression at or below 2% while preserving enough faculty learning to pass every quantity gate?
EXP-001's local guard accepted 200 of 200 candidates and missed 2.685% cumulative forgetting. EXP-002 changes the commit authority, not the task.
Next result
Scaling extended the path, but did not reopen it.
EXP-003 prospectively retried each frozen AdamW proposal at eight deterministic learning-rate scales. Five additional updates committed under the same 1.5% authority, then eight attempts exhausted every scale before a public checkpoint.
Commit authority
Six slices. One immutable baseline. Every attempt.
Candidate
The same six windows evaluated after AdamW proposes weights and moments, before commit.
Decision
Rollback above 1.5% cumulative increase or on any non-finite composite.
Forecasts frozen before the run
| Forecaster | Go / no-go / failure |
|---|---|
| Mechanistic guard model | 28% / 70% / 2% |
| Q2.3 trace model | 14% / 83% / 3% |
The predicted binding risk materialized. In the scientific retry, both forecasts settled to no-go: Brier 0.1688 for the mechanistic model and 0.0494 for the Q2.3 trace model. The preserved pre-training failure settled separately as execution failure: 1.5288 and 1.6494.
Result
The boundary held. Progress stopped.
74 attempts
66 candidates committed; attempts 67–74 exceeded 1.5% and rolled back.
1.5463% max
The largest accepted increase was 1.4253%; the largest rejected candidate was 1.5463%.
0 public reads
No public checkpoint, promotion evaluation, or replication seed was opened.
Still sealed
Promotion data, seeds 1 and 3, ZERO.4 promotion, and the Solomon bridge.
Stop rules
Eight consecutive rejects, two public replay violations, Pareto staleness, or attempt exhaustion.
Evidence
Every decision logs its composite, six source values, guard rule, and rollback digest.
Evidence
Settled and independently checkable.
The live ledger passes all admission gates and verifies 16 objects across 17 hash-linked events. An initial sandbox compile failure is preserved separately; it occurred before any optimizer attempt. The unchanged package then completed outside the sandbox and settled no-go.