Status: production (auto-copy) · Source: audit · Current version: 2
· Legacy tag: audit_version=v3
The production money-maker. A single deep-research pass per market (one model) that estimates a true probability and only deploys capital behind risk discipline — the guardrails the resolved track record demanded after the single-model approach blew up in June.
How it works
- Pull a category's active markets; screen out illiquid / wide-spread / already-priced ones.
- Per candidate: read the resolution rules (definitional traps), pull price history + live orderbook, do primary-source news diligence, commit a probability.
- Reject loudly (logged), then size the survivors.
What makes it "Keel" (the discipline)
- Conviction-weighted sizing, EV-agnostic: HIGH ≤15% / MEDIUM ≤7% / LOW ≤3% of capital. (The resolved record showed the model's own EV% was anti-predictive.)
- Entry-band screen: favor 60–90¢ favorites; restrict 35–60¢ coin-flips (HIGH + ≥15¢ edge only); default-reject <35¢ tails. (Coin-flips/tails bled; favorites hit ~90%.)
- 15% per-cluster cap so correlated double-bets can't sink a run. (The old 35% cap let a ~30% double-Kash-Patel loss through.)
- Cash is a position — holding cash in an efficient week is correct.
Why it earned auto-copy
Derived from 21 real resolved trades: HIGH conviction hit 80% / +36% ROI; the fixes above target the exact leaks (MEDIUM tier, coin-flips, correlated bets) that turned April–May's +$3.9k into June's −$3.4k.
Version history
- v1 — the original v3-prompt methodology (
mispricing-audit-v3.md). - v2 (Jul 2026) — Stage 2.5: resolver profile + risk pricing. Every
pick's probability must forecast the resolver's verdict at resolution time
(exact source, mechanism, definitional boundaries), with each resolver risk
subtracted as an explicit point deduction shown in the rationale
(
Resolver: …; priced risks: −X pts (…)); deductions >15 pts → reject. Motivated by the Fable 5 AI run's one loss (thesis argued model quality; the market paid a leaderboard snapshot that a mid-month launch flipped) and by Preseen's published methodology, which prices resolver-conflation risk explicitly. Rows stampstrategy_version=2so the record separates pre/post.