# Learning-loop review — 2026-07-21

6 closed trades reviewed, 2 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

# AI Review

## 1. Is overall expectancy holding?

Overall expectancy is **-0.15% per trade** across 6 closed trades, with a Wilson 95% lower bound of 18.8% on a 50% raw win rate. This sample is far too small to mean anything — the anti-self-deception rule requires ~20 trades minimum before any bucket (including "overall") can influence decisions. 6 trades is a coin-flip-indistinguishable result; no conclusion about real edge can be drawn yet.

## 2. Which bucket is the biggest drag?

**No bucket in this report has ≥20 trades.** The largest bucket anywhere (overall, by strategy, by exit reason) tops out at 6 trades, and most cuts (RSI band, MA50 distance, relative volume, regime, symbol, strategy+symbol) fragment that same handful of trades into groups of 1–3. Per rule 5, every one of these is a hypothesis, not a result. I'll stop the drag-attribution analysis here as instructed.

## 3. Entry problem, exit problem, or regime problem?

Not answerable yet — this diagnosis requires a bucket that clears the 20-trade bar, which none does. (For color only, not as a basis for conclusions: MAE losers avg -0.95% vs MAE winners avg -0.23%, and MFE losers avg only 0.06%, which on its face looks more like a tight-stop/entry-timing pattern than a give-it-back exit problem — but with only 3 stop-outs and 3 EOD-flatten exits, this is noise until the sample grows.)

## 4. Proposed parameter change

**NO CHANGE.** No bucket clears the 20-trade minimum (largest is 6 trades, all from orb-v1, which is itself listed as "backtest-only, not live-wired" in the live params — so this is paper/backtest data, not even the live book). Proposing a parameter tweak off 6 trades would violate rule 1 and rule 5 outright, and there'd be no way to attribute a later result to the change per rule 2.

What to wait for: accumulate at least ~20 closed trades in a single bucket (start with "by strategy" — orb-v1 is closest, but the strategy is currently marked backtest-only) before any parameter is touched. Given orb-v1 is producing ~6 trades in this window, that's roughly another 3 comparable periods before the review can act. Continue running `pnpm learn` daily as-is; it will simply keep reporting "insufficient sample" until a bucket crosses the threshold.

## 5. Notable open positions

Both open positions (BTC/USD, ETH/USD, crypto-trend-v1) are only 0.2 days old — too fresh to be notable in terms of aging or drawdown. Worth flagging as situational color only: neither has a calendar time-stop (max hold is open-ended, relying solely on SL/TP), which is a structural note for later review of crypto-trend-v1's exit rules, but this is not evidence for any parameter change today since these trades haven't closed.