# Experiment reruns after measurement fixes — September 24, 2026

**The fixes rescue one old candidate: the original BTC 6H hybrid9090.** Replaying
the complete archived pool finds no additional qualifying replacement. It is
already one of the three frozen observation policies. No live model changed.

The [priority queue](mlp_experiment_queue_2026_09_24.md) was recorded before this
batch. All five experimental stages have now run on the latest preserved chart
exports with the original candidate sets. No thresholds, model choices or
acceptance gates were expanded in response to results. The analysis remains
retrospective; September 25 prospective observation is a separate experiment.

## 1. Archived model/settings pool — keep original hybrid9090

Replayed exactly **10,010** configurations: 1,001 per model, consisting of the
incumbent and grouped, random-sparse and hybrid models with seeds 9090/9191/9292.
These are the same configurations tested previously, not a new random search.
Every row uses the current data/accounting contract and a freshly replayed incumbent.

Of 28 candidates passing the initial trade/risk/subperiod/WFO hurdles, 14 pass
the execution-sensitivity guard and 14 fail it. The other 9,982 are explicitly
unassessed for sensitivity because they already fail a core hurdle. After all
remaining comparison requirements, **one row qualifies**: original hybrid9090,
archive index 4335. Its confirmation threshold remains −730; every policy
parameter matches the hybrid in the observation bundle.

| Policy | IS return | Drawdown magnitude | Calendar Calmar | Historical eligibility |
|---|---:|---:|---:|---|
| Incumbent control | 2846.94% | 50.50% | 1.00449 | Reference |
| Original hybrid9090 | 1037.79% | 31.87% | 1.07604 | Pass |

The hybrid trades lower total return for lower drawdown and higher Calmar. Its
raw proximity diagnostic is now 3, yet the actual perturbation test passes. This
reinforces why raw proximity alone was the wrong veto. No other grouped,
random-sparse, hybrid or retuned-incumbent configuration passes all requirements.

The hybrid's inspected March–September period returns 25.70% with just three
entries; that is descriptive evidence, not sufficient prospective validation.
The repeated historical searches and inherited training/selection exposure remain.

Evidence: `archived/configurations.json`, `rows.jsonl`, `incumbent.json`,
`comparison.json`, `summary.json`. A sequential start was checkpointed after 513
rows and only remaining rows resumed through six subprocess workers. The exact
archive index coverage is checked; no row is dropped or counted twice.

## 2. Score structure and entry loss — no new qualifying overlay

Replayed baseline, confirmation-off, entry-only/exit-only/combined pivots at widths
2/3/4, and the fixed 10% entry-loss policy. The old score-structure summary still
contains obsolete date handling; this batch uses the current evaluator and saved
full traces instead of that summary. Policies themselves are unchanged.

| Candidate | IS return | Drawdown magnitude | Calmar | Decision |
|---|---:|---:|---:|---|
| Exit confirmation off | 1057.13% | 57.06% | 0.60583 | Shelve |
| Higher-low entry width 2 | 1913.13% | 43.54% | 1.00857 | Fails risk/coverage |
| Higher-low entry width 3 | 1596.66% | 37.21% | 1.10101 | Keep existing observation; insufficient WFO |
| Higher-low entry width 4 | 2189.59% | 45.12% | 1.02358 | Fails risk/coverage |
| Lower-high exit widths 2/3/4 | 2.64–80.52% | 41.18–46.03% | 0.00692–0.16131 | Shelve |
| Combined widths 2/3/4 | −6.56–35.93% | 28.74–45.02% | −0.02854–0.08423 | Shelve |
| Fixed 10% entry-loss | 1522.29% | 55.92% | 0.71883 | Shelve |

Entry width 3 still has only one qualifying WFO fold. The `WFO_Neg_Folds=6`
sentinel for insufficient coverage does not mean six observed losing folds.
These alternatives fail core gates before a new sensitivity assessment; their
`unassessed_ineligible` states are not sensitivity passes. The already frozen
width-3 policy's separate sensitivity evidence remains in its observation bundle.

Evidence: `overlays/results.json`, `comparison.json`, twelve compressed full traces.

## 3. Archived slow-regime heads — shelve

Replayed saved seeds 8080/8181/8282 without fitting, plus the exact null-head
control. Candidate IS returns are 206.19%, 211.12% and 208.81%; all have 84.875%
drawdown and Calmar 0.17124–0.17385 versus the control's 1.00449. None has enough
qualifying WFO folds. Their 31.79% inspected later return has only one entry.

The old residual-head trainer did not purge ten-bar label endpoints at training
and validation boundaries. Replaying saved heads does not repair that defect.
This is a negative screen of those artifacts, not clean model validation. A
different future training study must correct the split before fitting; there is
no evidence here to justify retraining these heads now.

Evidence: `slow_heads/results.json`, `comparison.json`, four full traces. Null
scores match the canonical network within 1e−10 and null/control traces agree.

## 4. State policies and learned exits — negative results persist

State policies: all eight alternatives reduce IS Calmar on all five BTC
timeframes again (**40/40**). Macro memory, cooldown and hold timers do not earn
a live trial. The complete 360-row result grid remains available.

Learned exits: refitted only the original ridge/tree heads using their existing
strictly purged, per-window protocol; no hyperparameter search or new features.
The six historical windows across five timeframes give 30 paired comparisons:

| Variant | Return better / worse / tied | Calmar improved |
|---|---|---:|
| Constant delay 12 | 5 / 25 / 0 | 6 |
| Linear market veto | 2 / 8 / 20 | 2 |
| Linear state veto | 2 / 21 / 7 | 3 |
| Linear combined veto | 2 / 13 / 15 | 2 |
| Linear early management | 6 / 21 / 3 | 4 |
| Tree market veto | 3 / 15 / 12 | 5 |
| Tree combined veto | 5 / 13 / 12 | 5 |
| Tree early management | 7 / 20 / 3 | 3 |

Linear/tree early management median historical return changes are −26.01/−14.08
percentage points. Some later-period gains remain interesting but cannot rescue
failed historical evidence. For example, 8H delay returns 30.71% versus 2.37%
baseline with only three entries; 1D delay returns 32.79% versus 1.61% with five
candidate entries. These histories have been inspected and are not a holdout.

Evidence: `state/` and `learned_exit/` with metrics, summaries and audits. All
360 state return/DD/Calmar/trade metrics agree with canonical accounting. Legacy
exposure columns use mean contemporaneous position rather than canonical lagged,
time-weighted exposure; maximum state difference is 0.6623 percentage points and
exposure is not used in these selection judgments. Learned results retain the
same reporting convention. No exposure equivalence is claimed.

## 5. BTC 4H ATR stops — neither passes

Repeated the original ATR 5/7.5 designs and matched fixed controls, including
training-only calibration and validation-before-later nomination. Neither was
nominated. Median validation return changes versus baseline remain −2.48/−4.69
percentage points and median Calmar changes −0.27/−0.46. Worst drawdown
deteriorations are 2.04/5.90 points, breaching the fixed limit. No multiplier or
ATR period was adjusted. Shelve these designs.

Evidence: `volatility/`, including all 40 metrics, 48 paired rows, 96,570 bar
records, 1,023 event/horizon records and the independently reconstructed audit.

## Verification and next decision

The independent read-only agent reconstructed all 160 measurement windows across
the sixteen overlay/head traces, including fees, initial equity, actions and WFO
coverage. Maximum metric discrepancy was 1.78e−15. The learned-exit audit verified
175 strict label cutoffs and reconstructed 15,634 saved ridge/tree predictions.
The prior ATR auditor was reused with explicit current output/root/protected-file
paths; its strategy and arithmetic checks passed. Preflights passed; all current
exports have a missing final MVRV value and their existing loaders excluded that
unconfirmed last row. No future-data performance was manufactured.

The agent separately verified exact coverage of all 10,010 archived settings,
recomputed every eligibility verdict, and freshly reproduced the incumbent and
sole qualifier including full sensitivity reports. All 941 protected file hashes
and both observation bundles remain unchanged. The checkpoint preserves fourteen
model artifacts, all five chart exports, winner snapshots and 28 source files.

**Next priority: continue the existing frozen comparison of incumbent, original
hybrid9090 and higher-low width 3.** No new candidate needs adding. Obtain matching
current TradingView trade-list exports when available to finish execution parity;
the five winning-model score checks already passed. Keep observed new data separate
from retrospective reruns and retain the minimum observation coverage rules.

Do not repeat these same historical searches again without a specific new defect
or new hypothesis. This batch did not revisit unrelated legacy strategies, widen
architectures, rerun broad upstream training or introduce new assets. Those need
their own bounded design rather than inheriting a claim that this batch validated
them. The known slow-head label issue is recorded, not silently declared fixed.

Evidence root: `results/mlp_experiments/reruns_20260924/`. A verified checkpoint
archive `results/research_checkpoint_2026_09_24/experiment_reruns.tar.xz` and
`experiment_reruns_manifest.json` preserve this batch alongside earlier evidence.