# Miroscope — BUILD PROGRESS Live build log. Small saved steps. If a session compacts, resume from here. Product: MT5 Expert Advisor that trades ONLY the 1-2-3 reversal, manages trades (not just alerts), multi-timeframe bias (D1->H1, H4->M15), scale-in entries, partial + runner (15-off/5-running = 75/25), pre-entry invalidation that cancels a setup. SAFETY: live trading is OFF by default (`InpEnableLiveTrading=false`). No broker connect, no account creds, no real orders until Michael explicitly authorizes. Not compiled/tested yet — MQL5 compiler is not on this box. Every claim stays honest: written != compiled != backtested != live. Main file: source/Miroscope.mq5 ## Checklist - [x] Preserve the original submitted scanner (source/Tommy-123_Pattern_Scanner_V8.0_Full.mq5, untouched) - [x] Phase 0: code review, all findings proven (PHASE0-REVIEW.md) - [x] Phase 1: cleaned scanner P1, defects fixed (PHASE1-NOTES.md, tests/TEST-RESULTS.md) - [x] Phase 2: executing EA Miroscope_EA_P2.mq5 (detector + bias engine + setup objects + on-chart display) - [x] Phase 2: trade execution HARD-GATED OFF (InpLiveTrading=false AND InpAllowOrders=false, all 12 order sites audited) - [x] Phase 2: EA logic ported to Python and executed (ea/TEST-RESULTS-P2.md) - [x] Phase 3: risk engine (risk/risk_engine.py) + seeded simulation (risk/RISK-RESULTS.md) - [x] Phase 4: backtest on REAL historical data (100 trades, honest loss — see report) - [ ] Phase 4: paper trading on live market data — BLOCKED (no broker, no MetaTrader, no demo account) - [ ] Phase 4b: strategy-tuning loop (backtest says current spec loses; trading decisions for Michael) - [ ] Phase 5: live deployment — NOT AUTHORIZED, never without Michael's explicit go - [x] Phase 6: runbook + handoff docs (docs/RUNBOOK.md) - [ ] Compile in MetaEditor — NOT DONE (no MQL5 compiler on this box; needs Michael) ## Log - 2026-09-14 21:11 EDT — Starting real build from the 15 requirements + the submitted scanner. ChatGPT draft not needed as a base; building fresh from spec. - 2026-09-17 08:09 EDT — Client-facing build plan written (Miroscope-Build-Plan.md + PDF): 7 milestone-gated phases, Phase 0 done through Phase 6 handoff, honest week estimates, no live trading until paper trading passes. Real phases, no padding. - 2026-09-22 01:30 EDT — Project closed per Michael (deal off, no payment, no contract). Final archival package: full May-September correspondence reviewed, sanitized archival report added (docs/Miroscope-Archival-Report-2026-09-22.md + PDF), obsolete "(TD)" naming fixed, both document PDFs regenerated from current markdown and visually inspected page by page. ZIP rebuilt and integrity-checked. - 2026-09-22 02:05 EDT — Michael ordered the build restarted: phases 0 and 1, internal engineering only, live trading stays OFF, no broker, no real money. Work lives in the Miroscope side chat. Phase 0 re-verified against the actual source (was previously recorded done): exactly 2 duplicate function definitions confirmed, both hard MQL5 compile errors, both resolved per the author's own "SINGLE DEFINITION" comments. Findings: docs/PHASE0-REVIEW.md. Phase 1 delivered: source/Miroscope_Scanner_P1.mq5 (1,703 lines, 36 functions, zero duplicates, braces/parens balanced, scanner only, zero trade-execution calls, alert logic intact). Static audit only: NO MQL5 compiler on this box, "compiles clean" NOT claimed, MetaEditor check still required. Detection core ported to Python (tests/test_123_detection.py) and EXECUTED against synthetic OHLC: 16/16 pattern cases pass (5 bullish BUY, 5 bearish SELL, 3 no-break near-miss silent, 3 no-rejection near-miss silent), entry/SL/TP math verified at realized RR 3.0000. Raw core on 80 chop bars: 8 pattern-shaped hits with all gates forced open; the real scan layers trend/zone/RSI/volume/ML gates in front. Full results: tests/test-results-2026-09-22.txt. - 2026-09-22 02:10 EDT — CORRECTION, own the miss. The 02:05 entry above was my first pass in the side chat and it was shallow: it caught only the duplicate functions and wrongly said "everything else is structurally sound." A deeper parallel pass (from the main chat, artifacts at the build root) found four more real defects, and I verified each one against the source myself before accepting it: (F2) SortPairsByPriority reorders g_pairs[] while indicator handles stay keyed to the old order, so ATR/RSI/MA get attributed to the wrong symbols from tick 2 on; (F3) with ML off, mlConfidence=0.5 < 0.55 gate, so STRONG/FULL statuses are unreachable; (F4) CheckRSICondition is a tautology, always true for any RSI 0-100 (the working gate is the inner one in DetectPatternForPair); (F5) GetPipSize 4-digit branch returns point*100, 100x off; (F6) six vestigial inputs/globals declared but never read. Canonical review: PHASE0-REVIEW.md (build root). P1 was rebuilt from the original by build_p1.py with fixes for F1, F2, F3, F5 + Miroscope rebrand: source/Miroscope_Scanner_P1.mq5 (57,281 bytes, 1 def per function, g_displayOrder sort fix, ML gate fix, pip fix, zero trade calls — all re-verified by grep on disk). Canonical notes: PHASE1-NOTES.md. Canonical simulation: tests/TEST-RESULTS.md + tests/test_scanner_logic.py (24/24 textbook detections at the exact bar, 0 wrong-direction, 0 false fires on 4 malformed controls, all four defects reproduced numerically and fixes verified). My shallow docs/PHASE0-REVIEW.md was deleted (it was wrong); docs/CHECKSUMS-P1.txt regenerated against the real artifacts. Standing honest boundary unchanged: no MQL5 compiler here, MetaEditor compile check still required before this touches a live terminal. - 2026-09-22 02:35 EDT — Michael: keep building through the phases until a genuine wall. Phase 2 delivered: ea/Miroscope_EA_P2.mq5 (649 lines, 24,262 bytes) — 1-2-3 detector on trigger TF, bias engine (lane 0: D1 bias/H1 trigger, lane 1: H4 bias/M15 trigger), setup objects (entry/SL/TP1 1.5R/TP2 3R/invalidation/scale-in), scale-in entries (60/40 risk split), partial-plus-runner (75% at TP1, SL to BE+0.1R, runner to TP2), pre-entry invalidation kill, 12-bar expiry, 21:00 session cutoff + flatten, -3% daily loss halt, on-chart display. Trade execution HARD-GATED OFF: InpLiveTrading AND InpAllowOrders must both be true (both default false); all 12 order call sites audited behind the gate. Manual audit: braces 62/62, parens 360/360, 25 functions, zero duplicates. NOT compiled (no MQL5 compiler here). EA decision logic ported to ea/test_ea_logic.py and EXECUTED: T1 12/12 detected with setup values exact to 1e-9; T2 0 against-bias setups; T3 invalidation kill at 0.0R; T4 expiry at 0.0R; T5 full lifecycle (fill, scale-in, TP1 partial, runner TP2) at +3.025R; T6 scale-in then SL at -1.0R combined (correct risk allocation); T7 lane engine 46+84 setups, 100% direction/bias agreement. Full results: ea/TEST-RESULTS-P2.md. Notes: PHASE2-NOTES.md. - 2026-09-22 02:40 EDT — Phase 3 delivered by build worker: risk/risk_engine.py (stdlib only, import-safe) — position sizing from risk %, 60/40 scale-in plan, 75/25 partial-plus-runner, BE+0.1R, 12-bar expiry, invalidation kill, 21:00 session cutoff, -3% daily loss halt, 5 setups/day cap. Seeded simulation (seed 20260922, $10k, 1% risk, 40 synthetic setups): lot samples 0.160-0.355 (16.9-37.6 pip stops); 11 killed pre-entry, 8 expired, 21 filled, 10 scale-ins; exits: 8 full SL, 6 TP1+TP2, 4 TP1+BE, 3 TP1-then-BE; total -$532.00 on synthetic random resolutions (says nothing about edge — real verdict is Phase 4). Docs: risk/RISK-ENGINE.md, risk/RISK-RESULTS.md. All functions spot-verified against spec after delivery. - 2026-09-22 03:05 EDT — Phase 4 data hunt (build worker): real data obtained. backtest/data/eurusd-h1.csv — 10,605 H1 rows, 2025-01-01 to 2026-09-22 UTC, Yahoo Finance EURUSD=X (volume unusable, FX spot). Plus 1 validated day of Dukascopy M1 (cross-checked vs Yahoo 1m, ~3 pip mean diff). histdata.com POST hangs, stooq unreachable, Dukascopy throttles after ~15 reqs (full 20-month M1 pull infeasible). See backtest/data/DATA-SOURCE.md. - 2026-09-22 03:10 EDT — Phase 4 BACKTEST ON REAL DATA. backtest/run_backtest.py runs the full ported stack (P1 detector + P2 setup/lifecycle + P3 risk) over the 20-month EURUSD H1 set, lane 0 (D1 bias -> H1), $10k, 1%/setup, 21:00 UTC cutoff, 5/day cap, -3% daily halt. No spread/commission/slippage modeled. RESULT: 100 trades, 31 wins, win rate 31.0%, profit factor 0.78, expectancy -$13.82/trade (-0.139R), max drawdown 27.59%, return -13.82% (final $8,618.17). Exit diagnostic: 69 full SL (-1R), 16 TP1+TP2 (up to +3.02R), 15 TP1 then runner stopped at BE (~+1.4R); median trade -1.0R; max 2 concurrent. HONEST READ: as specified, the strategy loses on this dataset — 31% reach TP1 but only 16% ever see TP2 while 69% take the full stop; the 1.5R/3R structure needs a win profile the pattern is not producing here. Per the build plan this means fixing the logic (target/stop/gate tuning, re-run on same data), which is a trading-decision loop for Michael, not started here. Paper trading BLOCKED: no broker, no MetaTrader, no demo account in this environment. Full report: backtest/BACKTEST-REPORT.md. Phase 6 runbook: docs/RUNBOOK.md. # 2026-09-25 06:45 EDT — VARIANT TESTING LOOP AUTHORIZED (Michael) Michael pasted 11 ChatGPT strategy variants (1-2-3 pattern family). Directive: port each TESTABLE variant into a Python harness, run on the SAME 20-month EURUSD H1 dataset as the Phase 4 baseline, keep what survives, cut the rest. Research proposes, the backtest disposes. Coordinator dispatched; one 15-min watchdog (miroscope-variants-watchdog) standing by. No live trading, no broker, no real money anywhere in this loop. - 2026-09-25 ~11:00 UTC — Variant backtest loop COMPLETE (coordinator subagent). Tested 8 H1 ChatGPT strategy variants (V1a/V1b/V2/V3/V4/V5/V7/V9/V11) on the same 10,605-bar EURUSD H1 dataset as Phase 4, $10k / 1% risk, no spread (comparable). Engine: variants/run_variants.py (stdlib only, deterministic; indicators and partial-fill accounting unit-verified; no lookahead: fractals confirmed 2 bars late). Ranked by expectancy: V3 +0.635R (PF 5.13, n=13), V7 +0.183R (n=3), V11 +0.068R (PF 1.12, n=23), V4 -0.029R, V5 -0.068R, V2 -0.088R, baseline -0.139R, V9 -0.396R; V1a/V1b 1 trade each (untestable: RSI-at-P1 filter passes ~1-15x in 20 months). VERDICT: no variant beats the baseline convincingly. V3 is interesting but n=13 (winners cluster in the 2025 EURUSD rally; needs out-of-sample data). V11 marginally positive at n=23. V4 breakeven; V2/V5/V9 lose; V1/V7 untestable as specified. M15 variants V6/V8/V10 NOT tested (no M15 data; not faked). Nothing here justifies paper trading. Full report: variants/VARIANTS-REPORT.md. Plan: variants/PLAN.md. Progress log: variants/PROGRESS.md. STOPPED per brief: no paper trading, EA untouched. # 2026-09-25 07:00 EDT — ROUND 2 AUTHORIZED (Michael) Round-2 ChatGPT workup received (10 experiments, pre-registered ladders, overfitting discipline section). Directive: port and backtest all round-2 runs on the same dataset/accounting as round 1. Spec: variants/ROUND2-SPEC.md. ~20 runs: V3 family x4, V1 ladder x2, V11 slope ladder x3, V4 3x3 timing/depth matrix x9, V2-EMA20, V9 control. Report must show expectancy CURVES across ladders (stable plateau beats isolated peak). Coordinator dispatched; one 15-min watchdog (miroscope-variants-r2-watchdog). No live trading, no broker, no real money. # 2026-09-25 ~07:05 EDT — ROUND 2 COMPLETE All 20 round-2 runs backtested on the same EURUSD H1 data/accounting as round 1. Engine extended (not rebuilt): variants/run_variants.py gained round-2 classes; round-1 classes untouched and re-run to exact reproduction. Mechanization per variants/PLAN-R2.md (round-2 "1 tick" = 0.01*ATR14 per spec; no lookahead). Unit-checked one run per family before batching; R math spot-verified by hand. Ranked by expectancy: V3-L5 +0.502R (PF 2.76, n=120), V1-RSI45 +0.490R (n=30), V3-L3 +0.465R (n=142), V3-NX +0.442R (n=13, no-expiry control), V1-RSI40 +0.394R (n=12), V2-EMA20 +0.378R (PF 2.24, n=107), V9c +0.172R (n=52), V4 matrix +0.031..+0.119 (D0.25 best in all 3 rows, weak ridge), V11-A/B/C identical +0.111R (n=30; slope knob non-binding). V3-NX shows the 18-bar expiry was worth ~+0.19R of round-1 V3's +0.635R; entry still positive without it. Sensitivity curves: V1 monotonic (2->12->30 trades, expectancy holds/rises), V3 threshold 5vs3 stable plateau, V11 flat (dead knob), V4 gentle D0.25 ridge. NO isolated single-rung peak anywhere: overfit discipline held, curves reported. VERDICT: first triple-digit-sample positives of the program (V3-L5, V3-L3, V2-EMA20). V1-RSI45 and V9c alive but thin (n=30/52). All in-sample; nothing justifies paper trading. Next honest step (Michael's call, NOT started): out-of-sample (more pairs/years) on V3-L5/L3 and V2-EMA20. Full report: variants/VARIANTS-REPORT-R2.md. Plan: variants/PLAN-R2.md. Progress: variants/PROGRESS-R2.md. STOPPED per brief: no paper trading, EA untouched. # 2026-09-25 07:20 EDT — ROUND 3 AUTHORIZED (Michael) Round-3 ChatGPT validation protocol received: pre-registered SURVIVED/INCONCLUSIVE/DEAD thresholds (+0.10R, PF 1.30, >=20 trades every pair, >=6/8 positive pairs, worst pair > -0.25R), fixed 8-pair test order, trap protocols, 3 minimal robustness variants, decision tree for round 4. Directive: freeze the protocol and run it. Spec: variants/ROUND3-SPEC.md (locked, no param changes). Downloaded 8 OOS H1 pair files from Yahoo (~10.5k bars each, full 07-17 UTC coverage). 64 runs: 5 variants x 8 pairs + 3 robustness x 8 pairs. Coordinator dispatched; one 15-min watchdog (miroscope-variants-r3-watchdog). No live trading, no broker, no real money. # 2026-09-25 11:23 UTC — ROUND 3 OOS VALIDATION COMPLETE (coordinator) 64 runs (5 variants + 3 robustness x 8 unseen pairs). Pre-registered verdicts: V3-L5 SURVIVED (+0.389R, PF 2.15, 964 trades, 8/8 pairs positive), V3-L3 SURVIVED (+0.375R, PF 2.08, 1,158 trades, 8/8 positive), V2-EMA20 SURVIVED (+0.285R, PF 1.79, 915 trades, 8/8 positive), V1-RSI45 INCONCLUSIVE (USDCAD n=19, one trade short of the 20 minimum), V9c INCONCLUSIVE (agg PF 1.21 below the 1.30 bar). All robustness perturbations held; no worry-condition triggered. Report: variants/VARIANTS-REPORT-R3.md. Decision tree points to round 4 (temporal validation of V3-L5 + V2-EMA20) — NOT started; Michael's call. No live trading, no broker, no real money. # 2026-09-25 07:33 EDT — ROUND 4 AUTHORIZED (Michael) Round-4 ChatGPT temporal-validation protocol received: 3 variants (V3-L5, V2-EMA20, V3-L3), V1-RSI45 benched (no post-selection relaxation), EURUSD excluded from verdicts (reference only), 4 chronological blocks with entry-block attribution, gold second-asset-class test for all three, tightened TEMPORALLY ROBUST/INCONCLUSIVE/FRAGILE verdicts, 10 round-5 branches. Locked in variants/ROUND4-SPEC.md BEFORE any run. Downloaded gold-h1.csv (GC=F, 9,837 bars). ~96 run-units: 3 variants x 7 FX pairs x 4 blocks + EURUSD reference + 3 x gold x 4 blocks. Coordinator dispatched; one 15-min watchdog (miroscope-variants-r4-watchdog). No live trading, no broker, no real money. # 2026-09-25 07:41 EDT — ROUND 4 TEMPORAL VALIDATION COMPLETE (coordinator) 27 runs (3 variants x (7 FX pairs + EURUSD reference + gold), full-series runs with entry-block attribution, frozen definitions, EURUSD hard-excluded from verdicts). Pre-registered FX verdicts: V3-L5 TEMPORALLY ROBUST (4/4 blocks pass, agg +0.402R, PF 2.21, 7/7 pairs positive, 830 trades); V3-L3 TEMPORALLY ROBUST (28/28 pair-block cells positive, agg +0.376R, PF 2.09, 1,000 trades); V2-EMA20 TEMPORALLY ROBUST (4/4 blocks pass, agg +0.285R, PF 1.79, 805 trades; weakest cells in short block D, none breaching the -0.50 guard). Gold: V3-L5 GOLD-REJECTED (1/4 blocks meet N>=30; expectancy positive in all 4 blocks), V3-L3 GOLD-INCONCLUSIVE (3/4 blocks pass), V2-EMA20 GOLD-REJECTED (2/4 blocks meet N>=30; expectancy positive in all 4). Gold rejections are sample-size, not sign; not used to rescue or condemn FX. Report: variants/VARIANTS-REPORT-R4.md; raw trades: variants/results-r4.jsonl; verdicts: variants/verdicts-r4.json. Round-5 branches firing: 1 (forward validation, frozen), 4 (threshold plateau on new data), 8 (V2 stays FX-specific). NOT started; Michael's call. No live trading, no broker, no real money. # 2026-09-25 07:44 EDT — ROUND 5 LOCKED (Michael approved) Round-5 ChatGPT forward-validation protocol received and locked in variants/ROUND5-SPEC.md BEFORE any forward bar. Archive extended to full round-5 freeze: backtest/data/archive-2026-09-25/ now holds raw data + frozen specs (R2-R5) + code + reports + results, 24 files SHA-256 hashed in MANIFEST-ROUND5.txt, sealed read-only. CORRECTION recorded: round 3 ran 8 non-EURUSD pairs; round 4 ran 7 (EURJPY dropped in transcription, verdicts stand on pairs tested); round 5 forward universe = all 8 non-EURUSD pairs. Forward window: 2026-09-21 → 2027-03-31 fixed. Formal verdict needs 6 complete months + N>=150/variant + 15 trades on 7/8 pairs. CONFIRMED: E>=+0.10R, PF>=1.30, 6/8 positive pairs, worst pair >-0.25R. Plateau test: both L3/L5 confirmed + |ΔE|<=0.10R. Rejection = permanent retirement, one audit diagnostic only, no new variants. Interim looks monthly (first 2026-10-31), descriptive only. Gold excluded from round 5. No forward runs possible yet — forward data accumulates from 2026-09-21 onward. Paper-trading gate design specified in workup; remains a human decision, not started. Video request still pending (~1 hour mark). # 2026-09-30 14:42 EDT — VARIANCE LAB LIVE (Michael directive: never lose progress) - P2B was the original attempt. It had an error code and was retired. P2C is the sole current base and control. Do NOT attribute old demo trades to P2B. - P2D built 2026-09-30: exact P2C with tradedBias = -BiasDirection() before DetectPattern(). Magic 20260930. Both gates false. Source: ea/Miroscope_EA_P2D.mq5 (1,017 lines). Page: /variance/p2d/. Compile pending, backtest pending, verdict pending. - P2E built 2026-09-30: exact P2C with raw bias untouched, baseSig inverted to sig = -baseSig after DetectPattern(). Magic 20260931. Both gates false. Source: ea/Miroscope_EA_P2E.mq5 (1,019 lines). Page: /variance/p2e/. Compile pending, backtest pending, verdict pending. - P2D vs P2E are genuinely different: P2D changes which signal qualifies (bias flip before detection). P2E keeps the qualifying signal and reverses only the resulting action. - Variance Lab: /variance/ lists control P2C (10,605 H1 bars, 100 trades, 31% win, PF 0.78, expectancy -0.139R, max DD 27.59%, net -13.82%) and both variants. - Vault: /vault/ password-protected (splash) holds every Miroscope document. Cloudflare is the vault, local disk is scratch. - Live trading remains OFF. No broker, no real money. No compile claimed (no MetaEditor here).