# Rerun verification receipt

Independent read-only agent: `experiment_inventory` (same model family, separate
agent execution). It performed arithmetic and configuration checks; it did not
modify code, tune candidates or make the main thread's selection decisions.

## Archived pool

Passed: 10,010 original normalized configurations, unique indices 0–10,009,
current source/data/model hashes and recomputed evaluation keys. Recomputed every
promotion verdict and validated all 28 assessed sensitivity reports. Exactly one
qualifier: index 4335, original hybrid9090, identical model hash and every parameter
to revision2 observation. Fresh incumbent/qualifier evaluations reproduce checked
metrics and complete sensitivity reports. Statuses: 14 pass, 14 fail, 9,982 core
ineligible and unassessed. Historical eligibility is not forward evidence.

## Overlays and saved heads

Passed: independently reconstructed 160 periods across 16 saved traces without
using the shared accounting helper. P&L, drawdown, calendar Calmar, entry/exit
counts and WFO eligibility agree, maximum error 1.78e-15. Original score-structure,
10% entry-loss and saved head policies reproduce scores/actions within 1e-10.
FROZEN incumbent params and MODEL_HASH match; baseline/null traces are identical.
No qualifying replacement. Unassessed overlay rows are not sensitivity passes;
WFO sentinel six means insufficient coverage, not six observed losing folds.
Saved slow heads retain unpurged-label training limitations.

## Main-thread checks

State: all 360 return/DD/Calmar/trade-count rows match canonical accounting;
legacy mean-position exposure convention is disclosed, not used for selection.
Learned exit: complete 315-row grid, 175 strictly purged cutoff records, 15,634
saved predictions reconstructed from serialized coefficients/trees and candidate
state. ATR: reused independent action/ATR/fee/event auditor with current output,
repository and protected-file path bindings; all 40 metrics, 48 pairs, 96,570 bars
and 1,023 event/horizon records pass. Individual machine-readable receipts are
in each study directory.

No new production code was implemented for this research batch. Prior-turn
parity-tool/test edits and unrelated user work remain outside this batch's edits.
The batch used current functions via private evaluator instances for older studies
whose CLI summaries are obsolete. Full traces and model/settings artifacts allow
reconstruction; no promotion method was called to write a winner.