my-NexusTradingForexBot/docs/TASK-09-70D-CANDIDATE-VALIDATION-FINAL.md

TASK-09 — 70D CANDIDATE VALIDATION / SHADOW / GOVERNANCE READINESS — FINAL REPORT

Agent: Hermes-70DValidation (AGENT-09) · TASK-09-70D-CANDIDATE-VALIDATION · 2026-08-19 Role: 70D Validation, Candidate Lifecycle & Governance Readiness Engineer This task performs NO promotion. Promotion remains human-authorized only (INV-015).

1. GIT STATE (verified, not assumed)

Item	Value
Branch	main
Local HEAD	e7586f9ea4c86b8438696b2869ed0b36b20599c6
origin/main	e7586f9ea4c86b8438696b2869ed0b36b20599c6
ls-remote refs/heads/main	e7586f9ea4c86b8438696b2869ed0b36b20599c6
ahead / behind	0 / 0 (in sync at verification time)
Working tree	DIRTY — 112 entries (parallel swarm WIP: model_lifecycle AI-Hub diagnostics, liquidity_runtime, schema_v2, tests, scratch probes). NOT touched by this task.

Parallel commits observed during this task (chronological): 3f3f3d9 (Forensic-70D-UI STEP-01/02), 987c550 (Forensic-70D-UI STEP-07 debug console), e7586f9 (AGENT-05 BUG-106 fix). The TASK-13 report's 67b77e5 and TASK-08's 11e3402 are both OLD heads — superseded.

2. BUG-106 — FIRST TECHNICAL GATE

• Status: FIXED (committed e7586f9 by AGENT-05, verified independently by TASK-09).
• Root cause (verified in code at HEAD): compute_70d_frame passed all_bars[:i+1] (entire history) to compute_liquidity_features per row → O(n²). The live governor caps at 4000 bars; the dataset builder saw unbounded history (also broke train==live parity).
• Fix at HEAD: LIQUIDITY_HISTORY_LIMIT = 4000 bounded causal tail (all_bars[max(0, i+1-4000):i+1]) in BOTH compute_liquidity_frame and compute_70d_frame. Matches the live engine's _completed_bars 4000-bar trim (train==live parity restored).
• Independent benchmark (TASK-09, this report — scratch tool outside repo):

idx	unbounded	bounded (4000)	ratio
4000	0.976s	1.031s	0.9x
6000	2.420s	1.224s	2.0x
10000	6.626s	1.348s	4.9x
15000	14.898s	1.561s	9.5x

Asymptotics: unbounded ~quadratic, bounded ~linear after the 4000-bar crossover. Semantics preserved: same causal window, same engine, same features/clipping/normalization.

• Authoritative benchmark artifacts: docs/BUG-106-PERFORMANCE-FIX.md + scratch benches (committed e7586f9).

3. PARITY (dataset == replay == inference == runtime)

• GREEN. Full TASK-03 parity suites run by TASK-09 at HEAD: test_70d_contract_parity_task3.py, test_70d_dataset_parity_task3.py, test_70d_parity_task3.py, test_70d_replay_parity_task3.py, test_70d_perf_task3.py, test_70d_runtime_hook_task3.py → 68 passed, 3 skipped (skips = explicit perf probes, by design).
• artifacts/validation/70d_liquidity_parity.json (existing, authoritative): real-broker probe on data/raw/XAUUSD_M5.parquet, 1000 bars, 25 timestamps checked, 0 mismatches, exact=True; 6-scenario full-matrix; schema hash 235b8fcc… tolerance 1e-12 (deep-history rounding-only, documented).
• Golden fixtures: tests/golden/70d_liquidity_parity/parity_golden.json.

4. 70D CONTRACT VERIFICATION (from executable source, not docs)

• schema_id: scalp_v3 (canonical 70D), registered in features/schema.py (Base 0..49 | News 50..59 | Liquidity 60..69).
• feature_schema_hash: 235b8fccc96b7e0e (stamped at build, verified by parity).
• dimension 70, feature count 70, ordering Base→News→Liquidity, names canonical (bsl_distance_atr … post_sweep_displacement at 60..69).
• algorithm version: liquidity_engine:70d-v1.0.0.

5. DATASET (real)

• dataset_id: ds_d3f35b12d63148da · hash aad73c8fde905223e1405a79ba9eb592397e93da4d09fc4e1e4e314f016465b0
• source: data/raw/XAUUSD_M5.parquet (real broker, last 1200 bars)
• symbol XAUUSD · M5 · 1146 rows (train 802 / val 171 / test 173)
• schema scalp_v3 · feature_schema_hash 235b8fcc… · dimension 70
• label_schema triple_barrier_3class_v1 · purge 3 / embargo 3 · split hash 6acd7b73…
• DATA QUALITY (from 70d_liquidity_parity.json real_dataset block): all_finite True, all_in_range True, duplicate_timestamps 0, schema_hash_ok True.
• Manifest temporal_range shows 1970-01-01 — a KNOWN writer formatting artifact (naive-datetime serialization); real probe timestamps verified exact (0 mismatches). Flagged, not silently repaired.

6. A/B/C EXPERIMENT

• Status: NOT_RUN — no candidate model trained. The TASK-04 fair-benchmark protocol and TEST-70D-MODEL-01..25 (incl. fairness gates) exist and pass/skip truthfully, but the A/B/C run itself requires a real 70D candidate training cycle which remains unexecuted in the swarm. No protocol modification occurred.
• TEST-70D-MODEL-01..08 fairness gates verified present (same dataset/labels/split/purge).

7-9. WALK-FORWARD / OOS / ROBUSTNESS / CALIBRATION

• All NOT_RUN — no candidate. Infrastructure verified: WalkForwardTrainer (artifact_save_path default = wf_candidate, BUG-104 guard), verify_70d_artifact, governance verify_candidate 14 gates. No random temporal splits (purge/embargo enforced by dataset factory). No OOS observed → no OOS invalidation risk.

10. CANDIDATE

• NO_CANDIDATE. Registry (experience_model_registry) holds ONLY the 50D Champion (3 rows, all primary_scalp_scalp_v1_50d, CHAMPION). No 70D candidate artifact, manifest, or lifecycle entry exists anywhere. Candidate path separation: candidate artifacts land under artifacts/model_generation/models/ (wf_candidate etc.); Champion lives at artifacts/models/scalp/XAUUSD/v1.0.0/ — physically isolated, verified.

11. CHAMPION (BUG-104 incident state — unchanged by TASK-09)

• Path: artifacts/models/scalp/XAUUSD/v1.0.0/model.pt
• SHA256 (independently recomputed this task): 9105cef7d93e23b8a7d529ef797efba9283f36bc8190d2729909c1fc95634d2f
• model.meta.json: RESTORED_CANDIDATE (bench_a_v1-derived 50D scalp_v1) — the frozen original f0f70efb… is UNRECOVERABLE (documented docs/CHAMPION_ARTIFACT_INCIDENT_20260819.md).
• Status: internally consistent (registry row, meta, and file hash agree on 9105cef7); the file's current hash MATCHES the recorded restore hash — TASK-09 did NOT alter it. Operator decision still pending per INV-015 (restore external / retrain / accept).
• Champion write-protection: VERIFIED — TEST-70D-MODEL-32 (default save path != live path) and TEST-70D-MODEL-31 (champion write protection) pass; WalkForwardTrainer default is artifacts/model_generation/models/wf_candidate/model.pt.

12. SHADOW

• Infrastructure: shadow70 package present (runtime/load-validator/health/drift/store/worker), import-verified; tests: test_shadow70_safety + test_shadow70_runtime → 45 passed (after a transient parallel-edit flake; rerun clean).
• Runtime state: NO_VALIDATED_CANDIDATE (truthful — no 70D candidate exists to attach).
• Evidence: shadow70_observations = 2 rows, both smoke_24 fixtures (not real candidate evidence). Shadow sample floor: NOT SATISFIED (no candidate → no observations).
• Safety: shadow package imports NO adapter/order_manager/risk engine (INV-018); test_shadow70_safety suite green. No broker execution path exists in shadow code.
• Virtual performance: none computed (no candidate).

13. GOVERNANCE

• verify_candidate: 14 gates (artifact_exists → liquidity_contract_valid), read-only.
• promotion_preview / rollback_preview / emergency controls: present, import-verified.
• model_governance_state: 0 rows. model_promotion_audit: 0 rows. No promotion executed.
• Required gates for THIS candidate: SKIP/INSUFFICIENT on training/walk-forward/OOS/ robustness/calibration/shadow — under the contract (PASS=PASS, FAIL=FAIL, SKIP=NOT PASS, INSUFFICIENT=NOT PASS) this means NOT_ELIGIBLE for promotion, by design.

14. REGRESSIONS

• BUG-108 (AUDIT-0007 release_metadata fresh-DB): test_database_migrations_phase18 → 38 passed (fresh DB reaches current version 7 incl. AUDIT-0007). Repair effective.
• Governance suite: test_model_governance_phase16 → 59 passed.

15. GOVERNANCE STATUS MATRIX (TASK-09 verdict)

Gate	Status	Explanation
Technical artifact	PASS	BUG-106 fixed + benchmarked (9.5x); parity green
Dataset	PASS	ds_d3f35b real, finite, in-range, no dup timestamps
Training	NOT_RUN	no 70D candidate trained (NOT PASS)
Walk-forward	NOT_RUN	no candidate (NOT PASS)
OOS	NOT_RUN	no candidate (NOT PASS)
Robustness	NOT_RUN	no candidate (NOT PASS)
Calibration	NOT_RUN	no candidate (NOT PASS)
Shadow	NO_VALIDATED_CANDIDATE	2 smoke rows only; floor not met
Drift	NO_VALIDATED_CANDIDATE	no candidate to monitor
News contract	PASS	news_10d_from_context + neutral-zero when disabled; parity-verified
Liquidity contract	PASS	liquidity_engine 70d-v1.0.0; parity exact on real data
Champion protection	PASS	BUG-104 regression green; paths isolated
Migration (BUG-108)	PASS	38 passed incl. fresh-DB AUDIT-0007

16. PROMOTION VERDICT

NO_CANDIDATE  (no 70D candidate artifact exists)

For the governance readiness of a candidate that does not exist:

INSUFFICIENT_EVIDENCE  (nothing trained/validated/shadowed)

NEVER PROMOTED — and nothing CAN be promoted until a real candidate carries evidence through the verify gates. Promotion remains human-authorized (INV-015).

17. REQUIRED ARTIFACTS

• artifacts/validation/70d_governance_readiness.json (this task, committed)
• artifacts/validation/70d_candidate_manifest.json → N/A (no candidate — documented in readiness JSON; not fabricating a manifest)
• Existing authoritative artifacts reused (no duplicates created): artifacts/validation/70d_liquidity_parity.json, docs/BUG-106-PERFORMANCE-FIX.md, docs/CHAMPION_ARTIFACT_INCIDENT_20260819.md, docs/70D_LIQUIDITY_PARITY_REPORT.md, docs/MODEL_BENCHMARK_70D_LIQUIDITY.md.

18. QUALITY GATES

• Full gate NOT run (parallel WIP in tree; gate failures would be indistinguishable from swarm churn). Scoped verification run instead (all GREEN): parity 68/3, shadow70 45, governance 59, migrations 38, BUG-104 guard 1. The project gate will be exercised on the swarm's consolidated tree by its owners.
• Transient flake observed and reported honestly: shadow70 suite 4-fail on first run (parallel file edit), clean 45-pass on rerun — NOT hidden, NOT a real regression.

19. COMMIT LOG (TASK-09)

See git log: AGENT-09 commits carrying this report + readiness JSON + handoff (docs only — no production code touched; swarm WIP preserved).