File-ops hard gate
12/12vs stock 9/12measuredQwen3-4B ReST Fused
On the frozen paired run, ReST improved file-ops depth from 9/12 to 12/12, removed eight unexpected side effects, and completed that slice 2.42x faster wall-clock. The same candidate fell from 30/45 to 25/45 on frontier-qualified breadth. The honest win is routed specialization, not a general-planner claim.
Headline Numbers
Evidence class applies to each value: measured = retained evaluation or runtime result; historical = recorded result not freshly reproduced; derived = calculated or decided from records; observed = current artifact state; not-measured = an explicit evidence gap.
Qualified breadth
25/45vs stock 30/45measuredDepth wall speedup
2.42x360.50s → 148.91sderivedSafety events removed
1011 total → 1 across both gatesmeasuredCompetitive Context
| System | Metric | Score | Size / Class | Comparable? | Readout |
|---|---|---|---|---|---|
| posttrainllm Qwen3-4B ReST | file-ops depth | 100% | 4B, 8.06GB stored | Direct | Fresh run: 12/12, zero unexpected side effects, and 13.40 tok/s. |
| Stock Qwen3-4B | same file-ops depth fixture | 75% | 4B | Direct | Fresh run: 9/12, eight unexpected side effects, and 12.10 tok/s. |
| ReST breadth result | frontier-qualified breadth | 55.6% | 4B | Direct | Below stock's 66.7%; this is why the frozen general-successor decision is reject. |
Direct rows share this artifact's eval setup. Directional rows are useful market context but should not be read as leaderboard claims.
Fresh requalification
| Gate | Stock | ReST candidate | Readout |
|---|---|---|---|
| File-ops depth | 9/12 | 12/12 | +25 points |
| Qualified breadth | 30/45 | 25/45 | -11.1 points |
| Depth wall time | 360.50s | 148.91s | 2.42x faster |
| Depth peak RSS | 6.77GB | 8.13GB | +1.36GB |
Release Blockers
General breadth regressed
The fresh candidate lost five of 45 frontier-qualified breadth cases versus stock.
Unblock: Do not unblock this candidate as a general model; route it only to the file-ops envelope.
Not a Pace planner
Pace uses a different intent envelope and its own six-dimension ship gate.
Unblock: Re-distill on Pace's action surface and clear the Pace gate before runtime wiring.
Fine-Tune Report Card
The portable before/after proof for this artifact: baseline, candidate, deltas, regressions, slices, performance, eval validity, and the ship/retry/reject decision — with every value labeled measured, derived, historical, skipped, or not recorded. Compiled offline from the recorded factory evidence; no model was run to produce it.