Warrior_EA/docs/MIND.md

The Mind - the trader behind the signal

Mind/ turns Warrior_EA from "a signal that fires" into "a signal that fires and a trader who looks at it, sizes it, writes it down and reviews the notebook". The validated dip-buy book is untouched: the Mind is off the trade path in MIND_OBSERVE (the default) and can only take risk off in the acting modes.

Loop

step	what	where
LOOK	every modality reads the closed signal bar into one SMarketContext	Modality*.mqh
RECALL	the reviewer asks how trades in similar conditions have gone	JournalReviewer.mqh
SIZE	P(win) -> fraction of base risk, never above 1 (AFML ch. 10)	BetSizer.mqh
WRITE	on close, the trade + context + outcome (R, MAE, MFE) go to a CSV and back to the reviewer	TradeBook.mqh
ORCHESTRATE	the only class that wires the four	WarriorMind.mqh

Modalities: price (z, ATR-distances, regime, volatility percentile), volume (tick volume with the hour-of-day removed, and effort-vs-result), Wyckoff (phase, last event, spring, the reader's verdict), breadth (share of the other indices also dipping), calendar.

Modes (MindMode)

• MIND_OBSERVE (default) - trades exactly as without it. Verified 2026-09-30: SP500 2023-01..09, same two trades and the same final balance (100,239.27) with the Mind off and observing.
• MIND_SIZE - a setup the journal rates below the book's average gets m(p)/m(p0) of the base risk (floor MindMinScale). Needs MindMinTrades closed trades first.
• MIND_SIZE_GATE - as SIZE, and a setup the journal rates a coin flip (p <= 0.5) is skipped.

Reading the result

Common\Files\Warrior_EA\Mind\book_<SYMBOL>_<TF>.csv - one row per closed trade, 48 columns. python research/read_book.py pools every book and ranks context fields by rank correlation with R, with the multiple-testing count printed under the table. At the end of each run the Experts log prints the reviewer's band table, the per-trade Sharpe, PSR(0) and an approximate DSR.

How to use it honestly. A field that looks good on the first backtest is a hypothesis. Run the next period (or another index) with it fixed before believing it; review_p vs realised R in read_book.py is the check that the journal's own forecast has any skill.

Design rules (why it is shaped like this)

• One column contract. ENUM_CTX + CtxName() in MarketContext.mqh; modalities write it, the book persists it, the reviewer buckets it. No other file spells a column.
• Open/closed. A new way of looking = a new CModality subclass registered in CWarriorMind::Init(). Nothing else changes.
• Authored rules are priors, not gates. CWyckoffReader::Bias() encodes the Wyckoff books' reading, but it only reaches the trade through a journal bucket, so the trades decide whether "against the structure" is a worse place to buy a dip. In the first test a dip the reader scored -0.75 ("distribution phase E") paid +0.50R - one trade, not a verdict, but the reason for this rule.
• Causal. Modalities read closed bars only; a trade enters the reviewer when it closes. In the tester the book is truncated at Init so one run never informs the next.

DRY changes made with it

• System/RegimeMath.mqh - the efficiency-ratio / variance-ratio arithmetic that lived inside CWarriorSignal, now shared with the price modality (signal behaviour unchanged).
• Mind/VolumeFeed.mqh - the one definition of relative volume. SignalNeural and the management net's crossing features now use it. This changes their feature vectors (names renamed so an old model file cannot load): retrain, and do not compare against nets trained before 2026-09-30.
• System/FeatureScale.mqh - FeatSquash, shared by both.
• mql5_patches/ holds older snapshots that no build includes (FracDiff.mqh has its own VolumeAt with the old idea). Left alone; delete or migrate when someone needs them.

Known limits

• The volume level-detrend (Detrended) is UNTESTED as a predictor; the book records rvol_tod (detrended) and rvol_tod_raw, so the first journal is its test.
• The reviewer is additive over four dimensions (volatility, volume, Wyckoff verdict, breadth): it cannot see interactions and says so in its header. With ~650 trades in five years there is not data for more.
• Skipped setups (vol-gated dips, vetoes) are not booked, so the Mind cannot yet learn from the trades it did not take.
• The DSR in the run summary uses the large-sample trial variance 1/(T-1); set MindTrials to the number of configurations really tried.