No cherry-picked winner — the whole run is on file. You don’t read this research, you watch it: below, the lab replays as one living graph — every candidate a node coloured by its verdict, every mutation an edge from parent to child, survivors lighting up and the rest fading to grey. Scrub the timeline, or click any node for its dossier.
Real rows from the box — including the families we had to bury.
How to read it: OOS = out-of-sample (data it never trained on) · PF = profit factor (above 1 = net winner) · Retention = how much of the edge survived out-of-sample · Verdict = the gate’s call.
| Candidate | Lane | Retention | Verdict |
|---|---|---|---|
| gap-play-004Top of the board — the first the lab would take to forward testing. | TSLA · 1D | 165% | PASS |
| supertrend-flip-009 | XAGUSD · 4h | 296% | PASS |
| donchian-breakout-021Part of the gold-daily donchian edge that cleared the deflated bar under four exit regimes. | XAUUSD · 1D | 217% | PASS |
| donchian-breakout-003 | XAUUSD · 4h | 244% | PASS |
| fvg-retrace-016 | XAGUSD · 2h | 228% | PASS |
| squeeze-break-001 | XAUUSD · 4h | 74% | KILL |
| capitulation-reversal-002Out-of-sample said no, after real costs. Cause of death on file — re-tests refused. | BTCUSD · 4h | 68% | KILL |
In plain words: a “deflated” t-stat checks whether a strong result is just chance from trying 251 ideas — above ~3 means probably real, and it can’t be gamed. Here, one donchian edge cleared the deflated significance bar under four exit regimes — counted, honestly, as ONE edge. And it's still only “promising.”
In plain words: the lab killed its OWN best result — the consistency gate caught that the edge was recency-concentrated, so it buried it, then demoted five more. The gates only ever get harder.
Real rows from the demo lab in the box — open lab/leaderboard.md and check every number yourself. That's the pitch: if you trust the folder, you'll trust the lab.
One self-contained offline file — every candidate, gate, dossier and correlation, with copy-paste Pine. The real board, generated from the 251-candidate demo lab you get.
The whole search in three views — every candidate, the verdict funnel it went through, and which edge lives on which market.
Click any survivor and the lab opens its full file — the evidence a quant desk builds AFTER a backtest survives and BEFORE risking real size.
Beyond single strategies — where edges can live, how survivors fit together, and how they hold up forward in real time.
The lab doesn't stop at 251. Copy a mission to your AI agent and the library keeps growing — every mission carries the honesty rules.
What died, why, and what each wave taught — the binding memory that makes the next run smarter.
Yours in the box — open it offline, no server. Every number reads from the same files as your own runs.
Nobody shows you the hundreds that died finding it. We built this lab backwards from one belief — a strategy is only as trustworthy as the graveyard behind it. So this one ships its graveyard. Every kill, every weak hand, every honest “not yet” — on disk, in plain files, signed by the evidence.
On purpose. Everything this lab refuses to be is exactly why you can trust what it tells you.
No entries texted to your phone. The lab hands you hypotheses with evidence trails — you stay the portfolio manager.
The ceiling verdict is PASS = “promising, validate further.” Anyone promising more than that is selling you the dream.
Re-optimizing on the engine is how you overfit — refused by design. Mutations are coarse, ≤2 parameters, judged on plateaus.
A backtest, even across many honest windows, is still a backtest. The lab names its own next gates: Monte-Carlo, significance, cross-market, small size.
Most products would never ship their graveyard. This one is the graveyard.