# SNIPERGOLD_ML SESSION HANDOVER

## Session Metadata

```text
Date:              2026-08-21
Timezone:          UTC+7 (MetaTrader server offset +7, local Asia UTC+7)
Session purpose:   P1 correctness repair (SB-01, SB-02, SB-06) + formal session close & handover
Starting Forge HEAD: bb5eec05efc4b8493154e25c02b6f019ec50365f
Ending Forge HEAD:  (recorded in section 11 after push)
Repository:         https://forge.mql5.io/chiki2bum2/SniperGold_ML.git
Local workspace:    D:\TradingTerminal\HFM Metatrader 5\MQL5
Publish mirror:     D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\AlgoForge\publish\AlgoForge
```

Repository layout note (identified, NOT merged):

```text
authoritative local source : D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\SniperGold_ML  (train_model.py, repaired\*)
research artifacts         : D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\AlgoForge\ml\parity\* ; MQL5\Files\* ; MQL5\Profiles\Tester\*
legacy project             : D:\TradingTerminal\HFM Metatrader 5\MQL5 (git root HEAD 609bcfd, README only; audit-only, never published)
publish mirror             : D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\AlgoForge\publish\AlgoForge (authoritative Forge source)
```

---

## 1. Session Objective

This session performed the **P1 correctness repair** on SniperGold_ML:

```text
SB-01  validation AUC bug in train_model.py            -> FIXED + TESTED
SB-02  temporal label overlap (purge gap = H_LABEL)    -> FIXED + VERIFIED
SB-06  runtime-vs-training feature parity harness      -> MEASURED + VERIFIED
```

All three are complete, each with a mandatory regression test and full provenance
(git SHA / dataset hash / config hash / seed / Python version recorded in the evidence
logs and P1_REPAIR_REPORT.md). The session was **stopped before P2** by explicit
instruction. No LSTM v2 / Informer / MS-GARCH / meta-label gate / threshold or
hyperparameter optimization was started. The freeze model was not modified.

---

## 2. Verified Repository State

```text
Forge HEAD (before close): bb5eec05efc4b8493154e25c02b6f019ec50365f (branch main)
Forge HEAD (after close):  see section 11 (verified via git ls-remote after push)
Local state:               D:\TradingTerminal\HFM Metatrader 5\MQL5
                           git root main @ 609bcfd077d13a82260367f35613d9ef42c69b29
                           ('Initial commit', 2026-01-16); only README.md tracked;
                           entire project untracked by design (audit-only repo).
Publish mirror state:      D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\AlgoForge\publish\AlgoForge
                           branch main @ bb5eec05efc4b8493154e25c02b6f019ec50365f
                           working tree CLEAN before close (verified: git status)
Sync status:               publish mirror HEAD == Forge HEAD (verified via git ls-remote origin HEAD)
Uncommitted state:         publish mirror: clean before close commit.
                           Local root: untracked research artifacts (expected; never published).
```

Verification commands executed during close:

```text
git rev-parse --abbrev-ref HEAD        -> main
git rev-parse HEAD                     -> bb5eec05efc4b8493154e25c02b6f019ec50365f
git status --porcelain                 -> (empty)
git ls-remote origin HEAD              -> bb5eec05efc4b8493154e25c02b6f019ec50365f
```

---

## 3. P1 Results

### SB-01 — Validation AUC bug

Old (bug):

```text
auc(yv.ravel(), Pv[:,0])
validation ≈ 0.5005
best_epoch = 0
```

The old metric flattened the two-head labels and compared them against only one
head's predictions -> validation AUC ~ 0.5005 (≈ random). Early stopping therefore
selected epoch 0; the freeze model (TEST 0.6270/0.6207) is a ~1-epoch model — an
artifact of the broken metric, not of the methodology.

New (fixed):

```text
per-head validation AUC  (val_auc_2head: AUC(y_long,P_long) and AUC(y_short,P_short), mean for early stopping)
VAL LONG  = 0.6796
VAL SHORT = 0.6796
TEST LONG = 0.6582
TEST SHORT = 0.6582
best_epoch = 55
```

Regression Test A: **PASS** (synthetic 2-head labels: per-head AUC = 1.0; legacy
`ravel` metric = 0.6250 != 1.0 -> proves two-head labels ≠ flattened labels against
one head).

Baseline lama `0.627/0.6207` is now classified **SUPERSEDED**.

### SB-02 — Temporal label overlap

```text
OLD:
train = 35652
test  = 11885
overlap exists: max_train_outcome_end = 44995 >= min_test_outcome_start = 44973

NEW:
train = 35652
purge = 13     (gap = H_LABEL = 24 bars, purge bars removed)
test  = 11872
split_bar = 44972
```

Inequality verification (explicit check, enforced by `purge_verify` which raises
RuntimeError on overlap):

```text
44995 < 44997   -> OK, no train/test outcome overlap
```

Regression Test B: **PASS** (synthetic split fails on the OLD scheme, passes on the
NEW purged scheme).

### SB-06 — Runtime-vs-training feature parity

```text
timestamp join   = 14850 / 14850
feed identical   = close max|d| = 0.0000
```

Per-feature groups (mismatch rate on joined samples):

```text
Formula-clean   (rate = 0.0000):  f4_internal  f5_bias  f7_sweep  f12_dsign  f16_mom20
Edge-small      (rate < 0.004):   f13_dmag 0.0007 | f3_swing 0.0017 | f8_choch 0.0035 | f9_chochok 0.0025
Window-sensitive(A-window: EA 600-bar runtime window vs training full history):
                                  f6_eqpos 0.5492 | f14_dhigh 0.3104 | f15_dlow 0.2974 | f17_range 0.5492
Formula-mismatch(B-formula):      f10_eqh 0.9846  | f11_eql 1.0000
Unresolved HTF  (B-formula):      f0_htf1 0.3609  | f1_htf2 0.4516  | f2_htf3 0.5130
Derived         (follows above):  f18_conf 0.5189
```

Key findings:

```text
1. A-window mismatch is MATERIAL, not cosmetic:
     range_atr : EA 56.99  vs py_full 10.44  (2026-06-16 13:30)
     eqpos     : EA -21.14 vs py_full -0.86
   The README claim that the window difference "practically does not change results"
   is DISPROVED by the parity evidence. This is the plausible explanation for the
   runtime cross-feed AUC drop 0.627 -> 0.5305.
2. "mom fix" = INVALID. An attempted EA edit (close[tc-21] -> close[tc-20]) was
   reverted: the parity harness proved the ORIGINAL formula was correct (the tc
   parameter receives tcv = total, so close[tc-21] = 20 intervals). After revert,
   f16_mom20 mismatch rate returned to 0.0000. Lesson: the parity harness caught
   my own wrong "fix".
3. Top-40 mismatch classification: 32 x A-window, 8 x B-formula.
```

---

## 4. Current Evidence Classification

```text
0.627/0.621 freeze                 = SUPERSEDED  (broken SB-01 metric + SB-02 overlap)
0.6582/0.6582 post-purge           = TRUSTED as corrected offline estimate
runtime cross-feed 0.5305/0.5487   = INCONCLUSIVE (contaminated by now-measured parity gap)
PF 1.32 (308 trades)               = INCONCLUSIVE
parity report                      = TRUSTED
historical pre-purge evidence      = PRESERVED (pre_purge_run.log + renamed INVALID artifact)
```

Explicit:

```text
0.6582/0.6582 is NOT a deployment claim.
Runtime/training parity is not yet fully resolved.
```

---

## 5. Current Architecture State

```text
Engine 1          = MTF data/cache            (implemented, verified)
Engine 2          = N/C/E/P SMC + fuzzy       (4 independent signal agents + weighted aggregator, implemented)
Engine 3          = renderer                  (display only, reads Engine 1-2 output, implemented)
MLP               = static, SEPARATE from Engine 2 (19->12->2, per-head Platt calibration,
                    exported .mqh; NOT an Engine 2 component)
LSTM              = No-Go experiments (Phase 4 sub-session 3; gate per protocol G1-G7)
Regime-aware ML   = not implemented
Meta-label gate   = not implemented
```

The target architecture is NOT described as implemented anywhere in this document.

---

## 6. Open Blockers for P2

### P2-A — HTF Bias

```text
f0-f2 mismatch = 36-51%
root cause     = UNRESOLVED
```

A faithful Python replica of the EA BTTFBias logic on the same data does NOT
reproduce the EA output (probe_htf.py / probe_ea_bttf.py / probe_htf2.py /
probe_htf3.py; probe_htf4.py hit a degenerate case with no M15 bars at the gold
daily break). The EA tester's D1 cache content differs from the training feed.
BTTFBias input/cache state must be traced.

Status: **BLOCKING — SOURCE OF TRUTH = UNRESOLVED**

### P2-B — EQH/EQL

```text
f10_eqh mismatch = 98.5%
f11_eql mismatch = 100%
```

Python training uses same-type pivot pairs + ATR at the pivot bar; EA/v4.4 uses
consecutive-in-list pivot pairs + current ATR. Current policy: legacy runtime/v4.4
semantics = current production reference. But semantics are NOT to be considered
theoretically correct yet.

Status: **BLOCKING DESIGN DECISION** (do not silently align)

### P2-C — Runtime Context Window

```text
Affected: f6_eqpos, f14_dhigh, f15_dlow, f17_range
Finding:  Python/full history vs EA/600-bar runtime window
```

Material magnitude proven (see SB-06); this is the largest economic impact item.

Status: **BLOCKING** — the production runtime window must become an explicit contract.

---

## 7. Source-of-Truth Policy for Next Session

```text
Production runtime semantics are the default source of truth
for features actually consumed by the deployed MQL5 runtime.

Exception:
If runtime semantics are proven to be a semantic bug,
do not silently copy them into training.

Instead:
LEGACY-BUG
-> define corrected semantics
-> update runtime
-> update training
-> establish parity
```

Per blocker:

```text
HTF bias       : UNRESOLVED
EQH/EQL        : legacy runtime/v4.4 semantics = provisional production source
Context window : runtime deployment window = provisional production contract
```

---

## 8. Experimental Freeze

Until P2 is complete:

```text
NO LSTM v2
NO INFORMER
NO MS-GARCH
NO REGIME GATE
NO META-LABEL GATE
NO HYPERPARAMETER OPTIMIZATION
NO THRESHOLD OPTIMIZATION
NO PF OPTIMIZATION
NO NEW TRADING CLAIM
```

The MLP architecture (19->12->2) and seed 42 must be preserved for baseline comparison.

---

## 9. Next Session Entry Criteria

The next session starts with **P2 SOURCE-OF-TRUTH ADJUDICATION**, in order:

```text
P2.1  HTF bias forensic trace        (debug dump EA BTTFBias input/cache; resolve f0-f2)
P2.2  EQH/EQL semantic adjudication  (decide pivot-pair + ATR convention; f10/f11)
P2.3  runtime context window         (decide window contract; f6/f14/f15/f17)
P2.4  FEATURE_CONTRACT               (write the explicit feature contract document)
P2.5  full 19-feature parity         (re-run harness until all features clean by contract)
P2.6  corrected baseline             (retrain/evaluate on parity-corrected feed)
P2.7  runtime probability parity     (CSV mode-0 AUC parity runtime vs training)
```

P2.1–P2.5 must NOT be skipped to jump directly to a new model.

---

## 10. Required Artifacts from This Session

Committed to this repository (repo-relative paths, SHA256 verified at close):

```text
docs/SESSION_HANDOVER_2026-08-21_P1.md
repaired/P1_REPAIR_REPORT.md                                  85FCE2D34BD6BDD7A5B1A67A3708A48E0438BE7430D536A9F2E6086BD86E191A
repaired/pre_purge/pre_purge_run.log
repaired/pre_purge/SniperGold_ML_prepurge_INVALID_overwritten_by_postedit_run.mqh
repaired/post_purge/post_purge_run.log
repaired/post_purge/SniperGold_ML_postpurge.mqh               9D8906A89DB6580A01A1F1747E6E24D4C9D3F849910AEDCE2AFE873B6F1FE44E
ml/parity/parity_harness.py                                   5C04ADD5CA1AB1DA306BF25BA7DBC4968D447F6DAD8730DDF8ABA75A1C90037F
ml/parity/RUNTIME_TRAINING_PARITY_REPORT.md                   300E5DCFF205BC35FFA369C4D888D0A9C935233117EA839EF13FEBEA1AC2382D
ml/parity/AlgoForge_bt_features_XAUUSD_M15.csv                DCA3AC50D7B6465365D6EBE6DF43F96CAC91809153630C1A88B188480C387D44
ml/train_model.py                                             1B967C764638A7B1FED66D3B7C80A88E2FF0AE8311E5C1D7F406870CBC8B02E5
MQL5/Experts/AlgoForge_Backtest_Baseline.mq5                  A19891E40A5EA0D400F54E54ED42BDCD1D517C0EE1FE1FA9F121795B7B3E4FBA
```

Freeze model (NOT modified during P1, NOT committed, remains in local workspace):

```text
D:\TradingTerminal\HFM Metatrader 5\MQL5\Include\SniperGold_ML.mqh
SHA256 C0D5CAE6663F457D2C87750BC8690B47EF4557D0426BFB4183280DC1D1A2F1E5
header 'AUC validasi : 0.5005 | AUC long: 0.6270 | AUC short: 0.6207' = SB-01 bug evidence
```

Additional local research artifacts (NOT committed, remain in local workspace):

```text
D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\SniperGold_ML\repaired\diag_epoch0.py
D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\SniperGold_ML\repaired\probe_npz.py
D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\SniperGold_ML\repaired\probe_feed.py
D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\SniperGold_ML\repaired\probe_mom_eq.py
D:\TradingTerminal\HFM Metatrader 5\MQL5\Shared Projects\SniperGold_ML\repaired\probe_htf.py / probe_htf2.py / probe_htf3.py / probe_htf4.py / probe_ea_bttf.py
D:\TradingTerminal\HFM Metatrader 5\MQL5\Profiles\Tester\AlgoForge_Parity.XAUUSD.M15.20260101_20260820.000.ini   (F17A1B00...)
D:\TradingTerminal\HFM Metatrader 5\MQL5\Files\SniperGold_ML\features_XAUUSDc.npz                              (E60285F5... training cache, immutable)
D:\TradingTerminal\HFM Metatrader 5\MQL5\Files\AlgoForge\Data\features_XAUUSD.npz                              (547072AB... new cache, NOT used in P1)
```

---

## 11. Commit & Push

```text
Branch:          main (publish mirror)
Commit message:  chore: close 2026-08-21 P1 repair session with handover
Reviewed diff:   only closure/evidence artifacts listed in section 10 (plus updated EA)
Final Forge HEAD: 3e38d0e0f45109be0184bf1ed0120b7d8a532558 (verified via git ls-remote origin HEAD; canonical close commit = d55a481, follow-up doc commit = 3e38d0e)
```

---

## 12. Final Checkpoint

```text
Final Forge HEAD:   3e38d0e0f45109be0184bf1ed0120b7d8a532558 (canonical close commit d55a481; follow-up doc commit 3e38d0e)
Commit:             d55a481 (close) + 3e38d0e (checkpoint finalization)
Author:             chiki2bum2
Timestamp:          2026-08-21T19:33:40+07:00
Repository clean:   YES (git status: nothing to commit, working tree clean)
Push verified:      YES (git ls-remote origin HEAD == 3e38d0e == local HEAD at verification)
```