# Prioritized experiment reruns — September 24, 2026

Execution: research/knowledge work. This is the durable queue and decision log.
Use the corrected evaluator and the latest preserved TradingView exports. Do not
promote, alter winners/Pine, relax gates, or treat inspected history as forward data.
Preserve both observation freezes. Existing unrelated working-tree edits stay intact.

## Priority and stopping rules (declared before this batch)

1. **Archived grouped/random-sparse/hybrid comparison:** replay the exact 10,010
   configurations already tested, including the paired incumbent. The new execution
   sensitivity guard directly changes the old rejection mechanism. No extra random
   configurations, seeds or threshold search. Report eligibility by model and all
   failures; compare against the newly replayed incumbent only.
2. **Score-structure and entry-loss overlays:** replay the original confirmation-off,
   higher-low entry, lower-high exit and combined width 2/3/4 policies plus the fixed
   10% entry-loss experiment. The older score-structure report predates corrected
   measurement. Use current accounting rather than its obsolete summary function.
3. **Slow-regime residual heads:** replay saved seeds 8080/8181/8282 and the null
   control. Do not retrain or tune. Determine whether accounting/data changes rescue
   the existing models; retain the limitation that replay does not repair training.
4. **State/timer and learned-exit policies:** refresh their exact predeclared sets on
   all five BTC timeframes. These already used corrected accounting; the reason for
   this one repeat is refreshed data and longer historical coverage. No new policy
   variants or hyperparameter search. Learned heads use the existing purged protocol.
5. **BTC 4H volatility stops:** repeat the two predeclared ATR variants and matched
   controls once. These already failed corrected validation; retain the original
   validation rule and do not rescue them using the later period.
6. **Synthesis:** record keep/shelve/defer for every family, verify protected files
   and input/source hashes, preserve evidence, and specify the next decision.

Rank by historical eligibility and risk-adjusted evidence, not largest raw return.
Every family gets one bounded pass; failures do not authorize expanding the search.
No live change without separate matching trade-list parity and prospective evidence.
The September 25 observation boundary remains separate from this retrospective work.

## Execution record

| Priority | Status | Result / decision |
|---|---|---|
| 1 | Complete | 10,010/10,010 replayed; exactly one qualifying configuration: original hybrid9090, archive index 4335, confirmation −730. Already frozen for observation. |
| 2 | Complete | All 11 alternatives fail eligibility. Width-3 entry gate remains a coverage-limited candidate; confirmation-off, exit/combined overlays and 10% entry-loss do not rescue performance. |
| 3 | Complete | Seeds 8080/8181/8282: DD 84.875%, Calmar 0.171–0.174, insufficient WFO. Shelve saved heads; no retraining. Original unpurged labels remain a limitation. |
| 4 | Complete | State: 40/40 IS Calmar declines again. Learned exit: no consistent historical improvement; linear/tree early management worsen 21/20 of 30 pairs. Later gains remain sparse/inspected. Shelve tested sets. |
| 5 | Complete | Neither ATR variant nominated. Original validation deltas reproduce: median P&L −2.48/−4.69 pp vs baseline; drawdown deterioration breaches limits. Shelve. |
| 6 | Complete | Synthesis and independent audits complete; all 941 protected hashes and both freezes unchanged. Verified archive: results/research_checkpoint_2026_09_24/experiment_reruns.tar.xz. |

Evidence root: `results/mlp_experiments/reruns_20260924/`. Each completed stage
will retain its own outputs. If interrupted, inspect this queue and completed
stage artifacts before resuming; never overwrite earlier results or restart a
completed stage merely because a session ended.

An independent read-only agent inventories old studies and identifies stale
measurement paths; the main thread executes the ordered studies and owns decisions.