Aliwkarimi98/SniperGold_ML
Watch
1
0
Fork
You've already forked SniperGold_ML
 1
forked from chiki2bum2/SniperGold_ML
Code Pull requests Activity
SniperGold_ML/docs/P3_S22_2_DISCREPANCY_IMPACT_ASSESSMENT.md

11 KiB
Raw Permalink Blame History

P3-S22.2 — DISCREPANCY IMPACT ASSESSMENT

Date       : 2026-08-25
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
Repair     : NONE implemented in this phase (repair decision matrix only)
Handover   : docs/SESSION_HANDOVER_2026-08-25_P3_S22_1_VERIFICATION_DEBT_CLOSURE.md
             (newest authoritative handover by Git ancestry; read completely
              FIRST; predecessor chain traced through P3-S22 -> S21.R -> ...)
Starting SHA : 0f1421b0d0bd6e9c5c858ee9bc1fb02a18f3599f  (local==origin/main)

Scope

Assess, with independent machine-readable evidence, whether the two discrepancies found in P3-S22.1 materially affect:

  1. F2 zone state semantics,
  2. F3 Candidate Setup creation,
  3. F4 MTF gating,
  4. Candidate Setup population,
  5. P3-S16 v1 labels,
  6. P3-S18/P3-S20 baseline ML results,
  7. production-facing behavior.

The two discrepancies:

  • ADJ-1 — production AF_FVGZoneState/AF_OBZoneState never assign mit_state=PARTIALLY_FILLED(1); partial fills remain UNMITIGATED(0) with the auxiliary partial_filled flag.
  • ADJ-2 — the committed research M30 gate is built by index-pair resampling of the M15 feed (m30_from_m15), which drifts from the runtime UTC-clock M30 at session/data gaps.

Method

Each discrepancy was assessed independently with:

  • an independent oracle (fresh implementation in this phase, not the audited code under test),
  • controlled synthetic fixtures (human-auditable),
  • full-scope deterministic measurement on the frozen research feed,
  • counterfactual replay of the FROZEN F3 chain (canonical_oracle / F3SetupEngine, P3-S17R2 FULL PARITY) with only the disputed input changed,
  • mutation tests on disposable copies,
  • two-run byte reproducibility (runner).

Evidence: ml/p3/p3_s222_discrepancy_impact/output/*.json/csv + ml/p3/p3_s222_discrepancy_impact/README.md.

PART A — ADJ-1 IMPACT ASSESSMENT

Independent oracle and cases

A fresh canonical zone-contract evaluator was written in this namespace (partial fill = mit=1, ACTIVE; full fill = terminal; FVG wick full-fill; OB close-through strict). 14 cases cover the 12 mandated scenarios (untouched, fully mitigated, partial, partial-then-further, partial-then-invalidated, pre-entry partial, post-entry partial, narrow/wide zone, boundary touch, bull/bear symmetry, multiple consecutive partial fills, OB partial and OB close-exactly-on-boundary).

For each case the following were compared:

  • production state representation ({mit, partial, invalidated}),
  • oracle semantic state (contract {mit=1 for partial}),
  • consumer availability (IsActive / invalidation),
  • F3 Candidate Setup creation with BOTH encodings.

Results (s222_partA_adj1.json)

cases_total          : 14
consumer_aligned     : 14/14
f3_creation_identical: 14/14
exact_label_equal    : 6/14
adj1_label_departures: 8/14      (the ADJ-1 label itself)
mutations            : 6/6 detected (M_A1..M_A6)

The 8 label departures are exactly the partial-fill cases: production encodes mit=0 + partial_filled=true, the contract oracle encodes mit=1. In every one of those cases the consumer availability and the F3 creation result are identical.

Downstream determination

  • F2 consumers (AF_FindFVG/AF_FindOrderBlock) and the F3 layer use IsActive()/zone_consumable, which treat UNMITIGATED(0) and PARTIALLY_FILLED(1) identically.
  • The F3 zone_mit_at_creation field is informational; no feature/label uses mit_state (FEATURE_COLS is geometry-based: width/offset/dist/ages/gates).
  • Mutation suite confirms the oracle DETECTS label-level mutations (force/suppress PARTIALLY_FILLED, alter flag) and semantic mutations (mitigation-bar shift ±1, boundary condition), proving the verification has power: the ADJ-1 label difference is real but behaviorally inert.

Classification

ADJ-1 : A — REPRESENTATIONAL ONLY / NO DOWNSTREAM IMPACT

PART B — ADJ-2 IMPACT ASSESSMENT

Independent UTC-clock M30 oracle

utc_clock_m30 (s222_common.py) aggregates the M15 feed into fixed UTC 30-minute buckets by wall-clock open time (open = floor(t/1800)*1800, close = open+1800). The frozen Context-agent rule is then evaluated per UTC M30 bucket and mapped to a per-M15-bar gate by the frozen as-of rule (newest CLOSED M30 bar with close_time <= t+900).

Oracle self-test: bucket opens are multiples of 1800; 8 consecutive M15 bars form exactly 4 UTC buckets. Agent pipeline validation: feeding the oracle the committed index-pair M30 OHLC reproduces the committed per-bucket dirs with 0 mismatches (98,516/98,516) — the oracle machinery is byte-faithful and the only difference is the M30 bar construction.

Measurements (s222_partB_adj2.json)

feed bars                : 197,032     (scope 194,985)
M30 bars index-pair      : 98,516
M30 bars UTC-clock       : 100,764     (diff +2,248)
full-feed gate diffs     : 38,399 bars (19.49%)
scope gate diffs         : 37,651 bars (19.31%)
diff attribution (full):
  continuous (after a phase shift) : 36,320
  session_gap                      :  2,079
mutations               : 5/5 detected (M_B1..M_B5)

Gap taxonomy of the feed: intraday breaks (3,600/4,500/7,200 s), weekend gaps (~48-51 h), holiday/DST (~24-25 h) — the index-pair construction keeps its pairing phase fixed in data-index space, so after any missing-bar/odd-gap the synthesized M30 bucket boundaries no longer align to wall-clock 30-min marks, and the phase persists across continuous stretches (which is why the majority of gate diffs sit inside "continuous" segments).

Propagation question

Does the M30 construction difference propagate into F4 gate decisions or Candidate Setup creation? Answered by Part C replay: the F4/F3 layer reads the m30 gate as a semantic input, and the gate values differ on 19.5% of bars; the counterfactual (below) shows this DOES change which Candidate Setups are created. H4 and M15 gates are not affected.

Classification

ADJ-2 : C — HISTORICAL BEHAVIOR IMPACT CONFIRMED
        (Candidate Setup creation / population level under
         runtime-consistent UTC-clock M30 semantics)

PART C — HISTORICAL IMPACT REPLAY (deterministic counterfactual)

The frozen chain builder VPR.chain_parity (same kernels, same F3 engines) reproduces the committed population exactly when fed the committed gates (694 created / 686 in-scope, per-bar mismatch 0), confirming the replay baseline. Two independent counterfactuals were then applied to the SAME feed:

C-A — ADJ-1: production mit encoding

Every partially-filled active zone re-encoded to production form (mit=0 + partial flag), everything else identical.

entity comparison : 686/686 common, 0 only-base, 0 only-adj1
outcome signature : identical
feature rows      : 0 changed
verdict           : IDENTICAL

C-B — ADJ-2: UTC-clock M30 gate

The committed index-pair m30 gate replaced by the independent UTC-clock gate; all other inputs (h4, m15, sweep, choch, zone) byte-identical.

entity comparison (chain identity): n_a(committed)=686  n_b(utc)=707
  common chain identity : 484
  committed only        : 202      (would NOT be created under UTC M30)
  utc only              : 223      (would be created instead)
outcome signature      : DIFFERS
shared m30_gate feature changed rows : 0
verdict                : DIFFERS

Spot-check (s222_partC_spotcheck.json): for the 484 shared setups the m30 gate at the creation bar is IDENTICAL (0% diff) and only ~10% of window bars differ; for the 202 committed-only setups 62% have a different creation-bar gate and ~42% of window bars differ; the 223 UTC-only setups formed under similarly differing gates. This confirms the population change is driven by the m30 gate VALUE differences, not an artifact.

Interpretation. The research chain is the historical source of the 686-row P3-S18/P3-S20 population. Under runtime-consistent (UTC-clock) M30 semantics the frozen F3 rules create a different population (707 in-scope identities; 202 committed identities not reproduced; 223 new). Therefore ADJ-2 is not representational: it changes Candidate Setup creation on the validated historical scope.

PART D — ML IMPACT

  • No retraining, no calibration, no artifact rewrite (frozen P3-S16/P3-S18/P3-S20 results byte-identical; git status confirms only new research artifacts).
  • ADJ-1: 0 feature rows, 0 label rows affected (Part C 686/686 identical; mit_state not a feature/label).
  • ADJ-2: the FROZEN P3-S18/P3-S20 rows are unchanged. However, the historical Candidate Setup population they derive from is M30-construction-dependent:
    • 202 of 686 committed in-scope chain identities (29.4%) are NOT reproduced under the independent UTC-clock oracle;
    • 223 different identities appear;
    • the m30_gate FEATURE of surviving (shared) setups is unchanged (structural identity h4==m30==direction holds), so the row-LEVEL change is a population-membership change, not a feature-value change on survivors.

If the research M30 construction were ever repaired (UTC-grid rebuild), the Candidate Setup population and therefore P3-S18/P3-S20 input rows would change; any model result re-derivation is a separate authorized experiment and was NOT performed here.

PART F — SILENT-BUG PROTECTION (summary)

Dimension Applied
Independent oracle yes (fresh contract oracle A; UTC-clock M30 oracle B; agent-pipeline 0-mismatch validation)
Boundary conditions yes (14 synthetic cases; bucket fixtures; boundary touch)
Temporal isolation yes (as-of rule; closed-bar; no future bar)
Row alignment yes (chain-identity key; CSV row diff; spot-check)
Metamorphic yes (bull/bear symmetry; pair-phase shift; continuous-feed identity index-pair==UTC)
Mutation 6/6 (Part A) + 5/5 (Part B)
Reproducibility 2 runs byte-identical except generated_utc (runner)

Evidence grades: independent verification (oracles, mutations, replay) and validated evidence (agent pipeline reproduction 0-mismatch + Part C spot-check mechanism) are recorded in the JSON outputs.

External-data / production / deployment status

Production MQL5 : NONE changed (verified by git).
FEATURE_CONTRACT / labels / TP-SL-horizon : NONE changed.
ML : NONE (no retraining / calibration / nonlinear / deployment).
External data : NONE (no Tickstory/Dukascopy / substitution).
Terminal/tester : NONE executed in this phase (all evidence is Python on
                  committed frozen research artifacts).
Deployment/trading : NONE.

Final classification

ADJ-1 : A — REPRESENTATIONAL ONLY / NO DOWNSTREAM IMPACT
ADJ-2 : C — HISTORICAL BEHAVIOR IMPACT CONFIRMED
        (Candidate Setup creation/population under runtime-consistent
         UTC-clock M30 semantics; frozen P3-S18/P3-S20 rows unchanged)
Overall decision on repair: see docs/P3_S22_2_REPAIR_DECISION.md

End of P3-S22.2 impact assessment.