How the numbers are computed

The design law

• Deterministic dollars. Your trades are matched into round-trips (FIFO lot matching). Every figure is computed from those round-trips by fixed arithmetic.
• Past tense only. Figures are historical attribution — what a pattern cost in the period analyzed. They are not predictions, and adopting a rule carries no promised result.
• Language models narrate; they never compute. When AI is used (reading a statement into rows, phrasing findings), it is handed the already-computed numbers and is forbidden from inventing any. Extraction output is shown to you as rows you can review.
• No directives. The product never says buy or sell. A code-level tripwire test sweeps generated copy for directive language on every change.

The detectors

Each flagged "leak" has a fixed definition and a dollar counterfactual computed from your history:

• Disposition effect — losers held materially longer than winners while the average loss exceeds the average win; costed as the excess of each loss beyond your average win.
• Overtrading / cost drag — churn measured by short holds and estimated fees + slippage against gross P&L.
• Oversizing / tail risk — your worst loss versus your median loss; costed as the excess beyond 3× median.
• Revenge trading — oversized entries within days of a loss; costed as the realized P&L of those entries.
• Inverted payoff — average loss larger than average win; costed as the gap across your losses.
• Price-path checks (when price data is available) — what a fixed stop would have changed on losers, and how far winners ran after you sold. Measured on real OHLC paths, not modeled.

The headline "disciplined P&L" applies two rules together — a size cap at your median notional and a fixed stop — so overlapping leak estimates are never summed into one inflated number.

Forward validation — we test our own claims

Every flagged finding is stored with the window it was computed on. Once at least five newer trades exist, the same detector re-runs on only those trades and records whether the pattern was detected again or not detected forward — with both window sizes shown. We say "not detected forward," not "fixed": a five-trade window is evidence, not proof, and we won't claim causation our data can't support.

The discipline score & benchmarks

• The score starts at 85 and subtracts fixed penalties per detected leak by severity, floored at 5. It is a habit summary, not a performance prediction.
• Percentiles only from real data. Comparative claims ("top X%") appear only once enough traders are in the cohort (50+), and the label names that cohort. Until then we show your score alone. Published research on retail trading behavior (e.g. Barber & Odean, 2000) motivates which leaks we detect; it cannot place you on a percentile, so we don't pretend it can.
• Quarterly summaries bucket round-trips by exit date; the quarterly counterfactual uses your full-history median size, so quarter figures stay consistent with the all-time headline.

Limitations, stated plainly

• Partial history skews results — the coach sees what you give it (or what your connected broker reports).
• Fill timestamps are date-granular: same-day sequencing inside a day is not observable.
• Small samples are noisy; detectors require minimum trade counts before claiming anything.
• The sample audit on the landing page is synthetic — a believable demo trader, clearly labeled.