my-NexusTradingForexBot/docs/forensics/MODEL_READINESS_REPORT.md

MODEL READINESS REPORT — 2026-09-04T15:40+03:30

Canonical contract: ScalpNet v3 · (B,32,70) · 70D · scalp_v3 · 3 classes NO_TRADE/BUY/SELL · ds_70d_clean_m1_20260904 · CLEAN_HISTORICAL · hash 3ae68…/235b8fccc96b7e0e Candidate under test: artifacts/models/scalp/XAUUSD/70d_liquidity/model.pt (live champion path) Tooling: .venv/Scripts/python.exe Python 3.11.16 · torch 2.13.0+cpu · torch.load(..., weights_only=True) · polars

---

1. Executive Verdict

Axis	Grade	One-liner
Dataset	PASS	ds_70d_clean_m1_20260904 — 99,946 rows × 101 cols — sha256 3ae687eaaa1f32a64c6d8acc1ab92d4ab9bceb0949d11cfe9e83ea852e3260fe matches task — feature_schema_id scalp_v3 / label_schema triple_barrier_3class_v1 / class_count 3 / seq_len 32 / production_eligible true / all_gates_pass true
Contract (code)	PASS	temporal_contract.FEATURE_DIM=70 + schema_contract.DIMENSION=70 + CANONICAL_SEQ_LEN=32 + LABEL_SCHEMA_3CLASS_V1 class_count=3 — single source per axis — see CONTRACT_AUDIT_REPORT.md; ACTIVE_SCHEMA_ID=scalp_v1/50D is intentional documented lag
Artifact coherence	P0 FAIL — ARTIFACT CONTRACT INCOHERENCE	Metadata claims 3, tensor is 4 — classifier.weight [4,32] vs model.meta.json model_head_classes 3 — bundle is incoherent — BLOCKED FROM LIVE, BLOCKED FROM PROMOTION
Training integrity	INCONCLUSIVE (blocked by P0)	Fresh-init comparison not runnable on incoherent tensor shape; parameter_movement_frac 0.774 quoted in stale provenance is not evidence for this bundle
Behavioral health	BLOCKED — do not run	Probes on a 4-logit tensor through a 3-class decoder are meaningless (false NO_TRADE collapse or masked logit 3) — must be run only on a fresh coherent 3-class bundle emitted by the live 34×10 retrain
OOS / walk-forward / calibration	BLOCKED	t70d_seq_v1/v2 TCN 70D 3-class has OOS, but live ScalpNet v3 70D 3-class has not emitted a coherent artifact to measure
Offline↔live equivalence	BLOCKED	(B,32,70) contract is sound; no coherent scalp_v3 3-class ScalpNet to replay through
Governance	FAIL	model.meta.json dataset_id null — no provenance link to ds_70d_clean_m1…; feature_schema_hash binding exists but head mismatch voids it

Overall: P0 ARTIFACT CONTRACT BLOCK — behavioral/OOS/calibration/equivalence probes are WITHHELD until the live retrain emits a coherent bundle. Do not promote, do not trade live.

---

2. Artifact Forensics (machine-verifiable)

Full 33-artifact dump: artifacts/forensics/model_artifact_forensics_20260904.md + .json (this report summarizes).

Artifact	Size	SHA256	Params	input_projection	Head	Meta claim	Verdict
70d_liquidity/model.pt (live)	1,334,268	c8c0b5b06d4c094dc04c9e8ff45cbfffc6f3fb396d42e3df46449068b1dbfd2b	331,492	[128,70]	classifier [4,32]	70 / 3 / model_head_classes 3 / scalp_v3 / seq 32	P0 — claims 3, tensor 4
70d_liquidity/model.pt.bak_20260904	1,335,531	763a25f61fe6b7d35da79fc3d2432b5fce59fdd2fc9237e816de20ec79e88d98	331,492	[128,70]	[4,32]	—	P0 (same lineage, different weights)
70d_liquidity/model.pt.bak2_1788498799	1,334,268	c8c0b5b0… (identical to live)	331,492	[128,70]	[4,32]	2,051 B legacy meta → rewritten 4,293 B model_head_classes:3 without retraining	P0 (hand-edited meta)
70d_news/model.pt	1,335,531	2b98f333…	331,492	[128,70]	[4,32]	model_head_classes 4	Self-consistent 4 — P0 if used as 70D 3-class
t70d_full_retrain/model.pt	1,334,268	c8c0b5b0… (≡ live)	331,492	[128,70]	[4,32]	scalp_v3 70×3	P0 — identical tensor to live
wf_candidate/model.pt	1,335,403	ec84ed21…	331,459	[128,70]	classifier [3,32]	scalp_v4 70×3 / production_eligible false / label_origin UNKNOWN	Geometry PASS / governance FAIL
t70d_seq_v1/v2/model.pt	~954,896	—	236,803	[128,70] TCN	head.3 [3,64]	TCN_ATTENTION_V1 / scalp_v3 70×3 / seq 32	PASS (TCN family, not ScalpNet v3)
50d_main/model.pt	1,325,291	342901…	328,932	[128,50]	[4,32]	model_head_classes 4 / scalp_v1	Legacy 50D — out of scope

Scaler: model.scaler.npz → mean [70] / std [70] — dimension correct — not sufficient vs head mismatch.

Provenance signal: retrain_70d_liquidity_provenance.json (elapsed_sec 62.33 / trainable_rows 26947 / gate EVIDENCE_WRITTEN) claims BEHAVIORAL_HEALTH_PASS (n_cls 3 / max_prob 0.39 / logit_std 0.18 / wait_mass 0.0) — cannot be evidence for this 4-class tensor — provenance itself flags meta_has_dataset_id false / meta_dim_70 false. Stale metrics, wrong bundle.

2.1 Why P0 cannot be repaired by renaming

• Hand-editing model.meta.json (model_head_classes 4→3) is already proven to have happened (2,051 B → 4,293 B without weight change) — re-editing does not fix the tensor.
• wf_candidate [3,32] is scalp_v4 / UNKNOWN lineage, production_eligible false, missing feature_schema_hash — promoting it would be a governance bypass.
• Only a fresh 34×10 walk-forward emission (purged 15 + embargo 15, compute_70d_frame_fast, per-fold scaler fit on TRAIN only, CANONICAL_CLASS_COUNT 3) produces a promotable bundle.

---

3. Dataset — PASS (verified, not trusted)

Field	Value	Verified how
dataset_id	ds_70d_clean_m1_20260904	dataset_manifest.json
rows	99,946 (total 99946 / train 69962 / val 14991 / test 14993)	polars read_parquet + manifest.row_counts
sha256	3ae687eaaa1f32a64c6d8acc1ab92d4ab9bceb0949d11cfe9e83ea852e3260fe	sha256sum — matches task context
feature_schema_id	scalp_v3	manifest.feature_schema_id + verification.json
feature_schema_hash	235b8fccc96b7e0e	manifest.feature_schema_hash — matches model.meta.json
label_schema	triple_barrier_3class_v1 — NO_TRADE 0 / BUY 1 / SELL 2 — WAIT is policy-only	manifest.label_schema_id / contract.class_count 3
class_count	3	same
label_distribution	0:14898 / 1:6261 / 2:5788 (~55% NO_TRADE, imbalanced — expected for purged barrier)	manifest.label_distribution
lineage	CLEAN_HISTORICAL / label_origin CLEAN_HISTORICAL / governance_override_required false / production_eligible true	manifest.label_origin / contract.production_eligible
sequence_windows	seq_len 32 / max_gap_us 900000000 (15 min in manifest vs 10 min canonical — see note) / windows_total 26916 / windows_valid 22436 / tensor [26916,32,70]	manifest.sequence_windows
purge/embargo	purge 15 / embargo 15 / labeler embargo 3 / no_trade_stride 2	manifest.purge_parameters
eval_rows	26,947 walk-forward trainable	same as retrain_70d_liquidity_provenance.walk_forward.trainable_rows
builder	compute_70d_frame_fast (BUG-106, byte-identical to canonical)	manifest.contract.builder
all_gates_pass	true (verify_70d ok / gap_safe ok / label_integrity ok / schema_hash ok)	verification.json
git_commit	76254b9e	manifest.git_commit

Note — gap mismatch: manifest.sequence_windows.max_gap_us=900000000 (15 min) vs temporal_contract.CANONICAL_MAX_GAP_US=600000000 (10 min) — manifest window uses the dataset-generation gap; live/sequence builder canonical is 10 min. Windows valid at 15 min will be re-gated at 10 min by live — this narrows valid windows (acceptable narrowing, not widening). Documented, not a blocker.

---

4. Contract — PASS (code), intentional lag documented

Summary from CONTRACT_AUDIT_REPORT.md (285 lines):

Axis	SSoT	Live ACTIVE	Verdict
70D	temporal_contract.FEATURE_DIM=70 + schema_contract.DIMENSION=70 / SCHEMA_ID scalp_v3	features/schema.py:95 ACTIVE_SCHEMA_ID=scalp_v1 / 50D	PASS — intentional lag until governance promotion
32	temporal_contract.CANONICAL_SEQ_LEN=32 aliased by sequence.py	live_engine.py:468 → CANONICAL_SEQ_LEN / artifact seq_len	PASS — one literal, aliases only
3 classes	LABEL_SCHEMA_3CLASS_V1 class_count 3 + architectures.CANONICAL_CLASS_COUNT=3	4 tolerated via LEGACY_HEAD_CLASSES=4 + model_class_contract._mask_wait_logit	PASS with fenced legacy — training is 3
Ordering	schema_contract.canonical_feature_names() (50+10+10)	same	PASS — no duplicate tuple
Label map	LABEL_SCHEMA_3CLASS_V1.numeric_mapping {0,1,2}	WAIT=3 is policy-derived, never a training label	PASS
Scaler	npz(mean,std) + feature_schema_hash bound everywhere	live_engine.ScalerBundle	PASS — pre-train-pure, per-fold fit

No competing 50/70/32/4 literal drives training geometry outside the SSoT.

---

5. Training Integrity — INCONCLUSIVE (blocked by P0)

A 70D 3-class ScalpNet cannot be proven trained until a coherent 70D 3-class tensor exists. The live tensor is 4-class, so initial vs final weight delta / loss curves / fold metrics measured on 3-class assumptions are inapplicable.

What is proven about the pipeline (not this bundle):

• compute_70d_frame_fast is byte-identical to compute_70d_frame (200 synthetic M1 bars: max |feat_0..4 diff| 0.0, cols 80/80, rows 146=146) — forensic/recover-three-model-fast already integrated — 70D variants now use the fast path.
• t70d_seq_v1/v2 TCN 70D 3-class artifacts prove the 70D 32/70 OOS gate runs end-to-end.
• wf_candidate [3,32] proves the LEGACY ScalpNet 70D 3-class head compiles (331,459 params = 331,492−33 for 3*32+3 vs 4*32+4).

Required once the live 34×10 emits a coherent [3,32] bundle:

• initial vs final weight stats (prove not fresh-init), weight delta (max_tensor_diff style), loss curves per fold, fold-level metrics, class distribution / prediction distribution / logit stats, artifact sha256 / git_commit / seed / scaler hash per fold, training command provenance.

---

6. Behavioral Health — WITHHELD (do not run on incoherent bundle)

Required probes per §9 — to be run only on the fresh coherent candidate:

1. zero input 2) all +3 3) all -3 4) single-feature 5) feature-group 6) random valid 7) repeated identical 8) small perturbation sensitivity 9) class-logit distribution 10) probability distribution 11) batch vs single 12) determinism

Gates to fail on: near-uniform (mean_max_prob ≈ 0.28 / low logit_std), class collapse (all argmax one class), constant logits, insensitivity to meaningful features, exploding logits, NaN/Inf, wrong shape/mapping, scaler/order flipped.

Reference degeneracy: mean_max_prob ≈0.28 / very low logit_std / little response — from stale probes run on an incoherent/untrained bundle — must not be copied as pass criteria for the new bundle; use documented tolerance only.

Provenance stale metrics (not evidence): max_prob_mean 0.3909 / logit_std_mean 0.1837 / wait_mass 0.0 / margin_sensitivity 0.0618 / parameter_movement_frac 0.774 — quoted but rejected as evidence because the measured tensor cannot emit 3 logits.

---

7. OOS / Walk-Forward / Calibration — WITHHELD

Temporal split: purge 15 + embargo 15 (label horizon 15) — walk-forward is 34 folds × 10 epochs × batch 256 (--seed 42) over trainable_rows 26947 (purged, not 99,946 raw) — no train/test leakage when executed via WalkForwardTrainer.

To report once coherent: accuracy / balanced_accuracy / macro F1 / per-class precision+recall / confusion matrix / OOS metrics / fold dispersion / calibration error (ECE) / reliability diagram / entropy / class distribution / rejection-or-NO_TRADE behavior / confidence distribution.

Current OOS that exists: t70d_seq_v1/v2 TCN — not comparable to ScalpNet v3 (different arch), so not acting as gate.

---

8. Offline↔Live Equivalence — WITHHELD

Chain to prove (both paths): real market data → canonical 70D builder → sequence builder (32,70) → scaler (70) → model (3) → logits (3) → probs → decision.

To capture: one or more exact (B,32,70) inputs and replay identical tensor through offline three_model.train_variant forward path and live live_engine._infer_probabilities / models/scalp_net.ScalpNet — assert same feature ordering / same scaler / same dtype / same orientation / same artifact sha / same logits±ε / same probs±ε / same decision / no hidden live transform.

Contract is sound — temporal_contract proves TRAIN INPUT (B,L,70) == LIVE INPUT (1,L,70) — but no coherent 70D 3-class ScalpNet to exercise it.

---

9. Governance — BLOCKED

Valid lifecycle is candidate → challenger → champion.

Gate	Status	Evidence
No manual rename / direct champion write	PASS (currently)	Incoherence was a hand-edited meta, not a champion overwrite — still prohibited
Candidate registration with provenance	FAIL	model.meta.json dataset_id null / lineage null / git_commit null — no link to ds_70d_clean_m1_20260904
Challenger→champion promotion	PASS-weak	model_governance_routes.py:1245 execute_promotion requires actor+model_id+approval_token + freeze lock + hash re-verify — but token is presence-only, not HMAC-bound (see §11)
Champion overwrite protection	PASS	live_engine.hot_swap_model validates+bakes before swap under _bundle_lock — correct, minus path/auth
Path allow-list	FAIL	arbitrary model_artifact_path accepted (see security audit)
Deserialization	FAIL	weights_only=False (see §11)

---

10. Active Retrain — DO NOT DISTURB

Signal	Value
Command	scripts/dev/train_70d_liquidity_production.py --dataset-id ds_70d_clean_m1_20260904 --folds 34 --epochs 10 --batch 256 --seed 42
PIDs (15:40 check)	26528 (parent 18832) + 25032 (child of 26528) — both alive via Get-CimInstance Win32_Process
Provenance	elapsed_sec 62.33 earlier checkpoint — gate EVIDENCE_WRITTEN / variant 70d_liquidity / trainable_rows 26947
Action	Monitor only — no kill/restart/duplicate — when it completes, verify model.pt head is [3,32] + model.meta.json dataset_id=ds_70d_clean_m1_20260904 / feature_schema_hash=235b8fccc96b7e0e / seq_len 32 / class_count 3 / git_commit set

---

11. Security & Capital Safety — BLOCKED (2× P2)

Full file: SECURITY_AUDIT_SEC_CAPITAL_DATA_BROKER.md. Summary:

Area	Grade	Detail
torch.load / weights_only	FAIL	9 production loads: live_engine.py:2684,2727,2755 + governance/load_gate.py:66 + runtime.py:110 + integrity.py:269,314,426 + checks_features.py:421 + shadow/challenger.py:127 + model_governance_routes.py:582 + streaming_replay.py:135 — no weights_only=True — RCE before validation
Path traversal	FAIL	live_engine.hot_swap_model:1505 + diagnostics_state_routes.model_hot_swap:1636 accept arbitrary string → Path.exists() → torch.load — no resolve().is_relative_to(ARTIFACTS_ROOT)
Auth on model_hot_swap	FAIL	diagnostics_state_routes.py:1626 — zero Depends/JWT/Bearer — server.py:425 CORS ["*"]
Promotion token	FAIL-weak	execute_promotion token is presence-only, not HMAC(actor,model_id,expiry) + single-use
Silent exception swallowing	PASS	audit_repository.py DDL except: pass documented — broker layers degrade to UNAVAILABLE, not silent trade hides — minor: live_engine._expected_num_features_for_artifact:2688 silently falls back to FEATURE_DIM
Degraded-data trading	PASS	live_freshness_gate() + DEGRADED→BLOCKED→NO_TRADE — DEGRADED never trades

---

12. Remaining Blockers (ranked — do not downgrade)

Pri	ID	Status	One-liner
P0	MDL-INCOHERENCE	BLOCKED	70d_liquidity metadata 3 vs tensor 4 — must be re-emitted by live 34×10 retrain — no manual rename
P1	SEC-CAPITAL-DESER	BLOCKED	torch.load without weights_only=True — 9 call sites
P1	SEC-CAPITAL-PATH+AUTH	BLOCKED	arbitrary model_artifact_path + unauthenticated model_hot_swap + CORS *
P2	DATA-GOV-PROVENANCE	FAIL	model.meta.json dataset_id/lineage/git_commit null
P2	REL-BUG-160+BUG-239	DEFERRED	release.yml pre-stage + ISCC flag — separate PR stack nse_qa_head_wt
P3	BUILD-SITE-FLAG	DEFERRED	scripts/docs/build_site.py 600-line FLAG_BUILD_INDEX dead tail + docs-enhance wiring — P3 commit 979624fc pending publish
P3	WORKTREE-PRUNE	LOW	scratch/rungate/* 3 prunable wts — git worktree prune only after review

---

13. Safe Level

SAFE TO CONTINUE VALIDATION (off-live)
SAFE FOR PAPER ONLY  (no live trading on any 70D ScalpNet artifact)
BLOCKED FROM LIVE
BLOCKED FROM MERGE (Batch B/C — model swap / lifecycle / risk)
BLOCKED FROM PROMOTION (any candidate while P0 holds)

No architecture change proposed — wrong head is not a capacity problem.

---

14. Next Verifiable Step (after retrain emits coherent bundle)

1. sha256sum model.pt  → new sha != c8c0b5b0
2. torch.load(weights_only=True)  → classifier.weight shape [3,32]
3. model.meta.json  → dataset_id ds_70d_clean_m1_20260904 + hash 235b8fccc96b7e0e + seq_len 32 + class_count 3 + git_commit != null + seed 42 + num_folds 34
4. behavioral 12-probe suite (machine + human JSON)
5. OOS walk-forward 34-fold metrics + calibration (ECE) + confusion matrix
6. offline↔live (B,32,70) replay parity  (same scaler, same logits±1e-5)
7. candidate→challenger→champion promotion proof (hash-gated, HMAC token after fix)

All until then: monitor PID 26528/25032, preserve 6 stashes, do not promote.

Generated from direct tensor reads, manifest reads, and process probes — not from agent summaries.