How the gates separate. In numbers.
Equal notional per signal, fill at next session open. WINNER / EARLY / ENTRY / PASS side by side — if ENTRY doesn't beat PASS, we have no product.
EARLY (+0.5% avg · 68% win) beats WINNER (-1.4% · 50%): coil > chase on this book.
ENTRY/WINNER averages +1.0pp over PASS on forward returns — gates are separating.
Why the published number moved.
Published P&L figures moved as methodology was corrected, not as quiet rewrites of settled trades. Settled lines are the same book; the headline definition changed. Sequential full-capital compound is an artifact. Current public headline is mean trade return; see also 10%-sized book path.
| When | Figure | What it was | Why it changed | Status |
|---|---|---|---|---|
| Early public track (pre-audit) | WR ~13% / various equity claims | Sparse closed sample, mixed labels, no CI gates. | Book was still filling; raw hit rates on tiny n were noise. | SUPERSEDED |
| First equity headline | ≈ −33.6% ($1k → ~$664 style path) | Full-capital sequential compound of per-trade returns (or fixed tickets treated as 100% of a small book). | With ~30–40 concurrent holds, sequential stacking multiplies a ~0.5% mean drag into a large drawdown. Sizing artifact, not signal quality. | ARTIFACT |
| Sequential compound (positive path) | +32.4% sequential | Same full-capital sequential method on the closed book when mean trade flipped positive. | Same bug upside: overlapping holds treated as free redeploys. Published only as a labeled overstatement after audit. | ARTIFACT |
| Concurrent equal-weight (v4.8.0) | +2.6% concurrent portfolio | Daily equal-weight of open names, then compound — handles overlap. | Correct for “always fully invested in the open book.” Still not a 10%-per-name desk. | SUPERSEDED |
| Mean trade primary (v4.8.2 · current) | Mean trade +0.37% · WR 64% (n=116) · 10% sized book ≈ −0.1% | Headline = equal-weight average of closed-line returns. Sized path = 10% equity per name, max 10 slots. Sequential kept only as artifact. | Answers “is the average signal any good?” without inventing portfolio leverage. CIs and n-gates on cohort tables. | CURRENT |
The headline was a sizing artifact — not a signal verdict.
Sizing artifact: full-capital sequential compound is 14.7% while mean trade is only 0.24% (n=120). That gap is fixed-dollar tickets treated as 100% of a $1k book — not signal quality. At 10% equity per name (max 10 slots) the same history is -1.5%. Concurrent equal-weight (always fully invested in open names) is -0.7%. Headline = mean trade.
Fixed-dollar tickets on a small book, compounded as if each trade redeployed 100% of equity, produce wild drawdowns (or gains) from a half-percent mean. That is arithmetic, not edge. The public headline is now mean trade return; the 10%-of-equity path shows what a real desk with position limits would have seen.
- PRIMARY — Mean trade: equal weight per closed line. Answers “is the average signal any good?” Independent of book size.
- SIZED BOOK — Each trade deploys 10% of then-current equity, max 10 concurrent. Answers “what would a real desk with position limits see?”
- CONCURRENT EW — Each calendar day, equal-weight the open names’ daily share of return, then compound. Fully invested when names are open; handles overlap.
- ARTIFACT — Full-capital sequential: multiply (1+r) for every trade in series. Equivalent to sizing each ticket at 100% of book. Invalid with overlap or fixed-dollar tickets on a small book — produces the old −33% / +32% headline swings.
- Fill check: printed return_pct vs (exit−entry)/entry on sample.
- Fill check on sample — no mismatches above 6 bps.
| Ticker | Signal | Entry | Exit | Printed | Calc | Match |
|---|---|---|---|---|---|---|
| UNCY | winner | 2026-06-23 | 2026-06-28 | −11.10% | −11.10% | ok |
| MKC | winner | 2026-06-23 | 2026-06-28 | +7.75% | +7.75% | ok |
| DRI | winner | 2026-06-23 | 2026-06-28 | +1.49% | +1.49% | ok |
| MU | winner | 2026-06-23 | 2026-06-28 | +7.66% | +7.66% | ok |
| GIS | early | 2026-06-23 | 2026-06-28 | +4.59% | +4.59% | ok |
| KMI | buy | 2026-06-23 | 2026-06-28 | +2.00% | +2.00% | ok |
| TFC | buy | 2026-06-23 | 2026-06-28 | +1.65% | +1.65% | ok |
| MMM | buy | 2026-06-23 | 2026-06-28 | +1.34% | +1.34% | ok |
| ABT | buy | 2026-06-23 | 2026-06-28 | +3.97% | +3.97% | ok |
| WBD | buy | 2026-06-23 | 2026-06-28 | −0.52% | −0.52% | ok |
| RSG | buy | 2026-06-23 | 2026-06-28 | +3.49% | +3.49% | ok |
| WM | buy | 2026-06-23 | 2026-06-28 | +3.06% | +3.06% | ok |
120 settled trades · concurrent equal-weight (overlap-aware) · fills at next-session open. Forward book; sealed boards on /verify for immutability.
Do the gates actually separate?
| Verdict | Band | N | Win rate | Avg fwd return |
|---|---|---|---|---|
| WINNER | ★ winner / extended | 22 n=22 — early, wide error bars | 95% CI 31–69% | −1.4% CI −5.0 to +2.2% |
| EARLY | ◆ early coil | 31 | 95% CI 50–81% | +0.5% CI −0.6 to +1.7% |
| ENTRY | entry gate (incl. legacy buy) | 67 | 95% CI 51–73% | +0.7% CI −0.3 to +1.6% |
| PASS | failed gates · sealed shadow | 100 n=100 · board→board sealed control | 95% CI 6–17% | −0.3% CI −0.5 to −0.1% |
Methodology on the engine page → · raw rows via /api/engine/payload. PASS is a control arm — if N is 0 / pending, the paper sampler hasn't booked failed gates yet (starts on the next paper settle after the control-arm deploy). Error bars: hit rates carry 95% Wilson intervals; mean returns carry 95% t-intervals off the raw settled trades. Cohorts under n=10 make no claim, under n=30 they say "early" — a win rate without an interval is marketing, not measurement.