SniperGold_ML/ml/p3/p3_s222_discrepancy_impact

P3-S22.2 — DISCREPANCY IMPACT ASSESSMENT & REPAIR DECISION (research-only)

Session    : P3-S22.2 — DISCREPANCY IMPACT ASSESSMENT & REPAIR DECISION
Status     : COMPLETE (research-only; NO repair, NO production modification)
Policy     : docs/CODE_VERIFICATION_AND_SILENT_BUG_POLICY_v1.md
Disposable : YES — verification/audit artifacts only; nothing here is part of
             the production MQL5, FEATURE_CONTRACT, labels, or ML pipeline.
Owner      : chiki2bum2 (authoritative repo)
Scope      : ADJ-1 (F2 mit_state label) and ADJ-2 (M30 index-pair vs UTC-clock)
             impact on F2 semantics, F3 Candidate Setup, F4 gating, Candidate
             Setup population, P3-S16 labels, P3-S18/P3-S20 ML, production.

This namespace holds the independent oracles, deterministic replay, tests, machine-readable evidence, and reproducibility manifest for the P3-S22.2 impact assessment of the two discrepancies discovered in P3-S22.1.

---

What was assessed

Discrepancy	Observed	Potential effect under test
ADJ-1	production FVG/OB zone state never assigns mit_state=PARTIALLY_FILLED(1); partial fills stay UNMITIGATED(0)+partial_filled	does the representation change F2 semantics, F3 creation, features/labels?
ADJ-2	research M30 gate uses index-pair resampling of M15, not UTC-clock M30 buckets	does the M30 construction difference propagate into F4/F3/Candidate Setups/ML?

Scripts (run in order)

Script	Part	Role
s222_partA_adj1.py	A	independent zone-contract oracle + 14 cases + 6 mutations; ADJ-1 impact
s222_partB_adj2.py	B	independent UTC-clock M30 oracle + gap taxonomy + 5 mutations; ADJ-2 bar/gate measurements
s222_partC_replay.py	C	deterministic counterfactual replay (C-A: mit encoding; C-B: UTC M30 gate) on the full scope
s222_partC_spotcheck.py	C	mechanism spot-check of the C-B population difference
s222_partD_ml.py	D	ML-impact determination (frozen P3-S18/S20 rows unchanged; population sensitivity quantified)
s222_partE_decision.py	E	machine-readable repair decision matrix
s222_run_repro.py	F	reproducibility runner (2 runs, byte-identical except generated_utc)

Outputs

output/s222_partA_adj1.json           ADJ-1 oracle/cases/mutations
output/s222_partB_adj2.json           ADJ-2 UTC-clock oracle measurements
output/s222_partC_replay.json         counterfactual replay summary
output/s222_partC_rows.csv            row-level population diff (909 rows)
output/s222_partC_spotcheck.json      mechanism spot-check
output/s222_partD_ml.json             ML impact determination
output/s222_partE_decision.json       repair decision matrix
output/s222_reproducibility.json      reproducibility manifest
output/s222_oracle_utc_m30_cache.npz  deterministic cache of the UTC M30 oracle

Headline evidence

ADJ-1 (Part A + C-A)

• 14/14 cases: production (mit=0+flag) and contract (mit=1) are consumer-ALIGNED and F3-IDENTICAL; 8/14 exact-label departures (the ADJ-1 label itself).
• Mutations 6/6 detected where contractually expected.
• Part C counterfactual (production mit encoding on the full scope): IDENTICAL — 686/686 chain identities, per-bar outcome signature equal, 0 feature-row changes.
• Classification: A — REPRESENTATIONAL ONLY / NO DOWNSTREAM IMPACT.

ADJ-2 (Part B + C-B)

• UTC-clock M30 oracle: 100,764 M30 bars vs 98,516 index-pair; full-feed gate diffs 38,399 bars (19.49%); scope diffs 37,651 (19.31%); diff attribution: intraday-break/weekend/holiday gaps and continuous stretches after a phase shift. Mutations 5/5 detected.
• Part C counterfactual (UTC m30 gate into the SAME frozen F3 chain): DIFFERS — 686 committed in-scope vs 707 UTC-clock in-scope; 202 committed chain identities NOT reproduced, 223 different identities appear; per-bar outcome signature differs; shared rows keep m30_gate feature unchanged (structural identity h4==m30==direction on surviving rows).
• Spot-check: shared (484) setups have 0% creation-bar gate differences; baseline-only (202) have 62%; UTC-only (223) formed under differing gates.
• ML rows (frozen P3-S18/P3-S20): NOT modified. The historical population is M30-construction-dependent: 202/686 committed in-scope row-identities are not reproduced under runtime-consistent UTC-clock M30, and 223 different ones appear.
• Classification: C — HISTORICAL BEHAVIOR IMPACT CONFIRMED (Candidate Setup creation/population level under runtime-consistent semantics).

Decision (support only; NOT implemented)

ADJ-1 : ACCEPT WITH CONTRACT CLARIFICATION (option 2) — document that the two
        encodings are functionally equivalent; no code repair justified.
ADJ-2 : REPAIR REQUIRED — SEPARATE AUTHORIZATION (option 3) — rebuild the
        RESEARCH M30 gate construction on the UTC grid and re-derive Candidate
        Setups / P3-S18 / P3-S20 as a separate authorized phase. NO production
        MQL5 repair is needed (runtime reads broker M30, which is UTC-grid).

Rules

• NO production MQL5 modification; NO FEATURE_CONTRACT/label/TP-SL/horizon change; NO retraining; NO calibration; NO nonlinear model; NO external data; NO Tickstory/Dukascopy; NO deployment; NO trading; NO optimization.
• All mutation/alternate implementations live in disposable copies only.
• Historical P3-S16..P3-S22 artifacts are NOT modified (additive docs only, per the new handover).
• Reproducibility: every artifact is deterministic; the runner executed each part twice and found byte-identical substantive outputs (only generated_utc differs).

End of P3-S22.2 namespace README.

Provenance

Docs:

• docs/P3_S22_2_DISCREPANCY_IMPACT_ASSESSMENT.md
• docs/P3_S22_2_REPAIR_DECISION.md
• docs/SESSION_HANDOVER_2026-08-25_P3_S22_2_DISCREPANCY_IMPACT_ASSESSMENT.md

Governing policy: docs/CODE_VERIFICATION_AND_SILENT_BUG_POLICY_v1.md.

Inputs: frozen committed caches/feeds (M15 npz, gate-series cache, P3-S18 rows), frozen F3 oracle/engines (canonical_oracle / F3SetupEngine), frozen agent rules (spec_tests_engine2_gates.py). No network, no terminal execution.