Where angles are proven, not claimed. Every study below is real flat-bet ROI on 892,795 settled finishes at actual final-tote payoffs — the audited ROI every competitor hides behind hit-rates. This is the harness our autonomous research agent runs on each angle it proposes: keep only what beats the base rate and holds across time.
The favorite-longshot bias, on our own data: every band loses to takeout, and the loss worsens monotonically as the price lengthens — odds-on are "least bad," longshots are graveyards. The real lesson: you cannot beat this by betting a price band. The edge has to come from selection — the Value engine, the intel moat, and validated angles — never from price alone.
| Price band | Bets | Win% | ROI |
|---|---|---|---|
| 1· under even (odds-on) | 2,776 | 54.0% | -21.9% |
| 2· even–5/2 | 65,598 | 34.0% | -22.5% |
| 3· 5/2–5/1 (mid) | 254,060 | 18.9% | -25.6% |
| 4· 5/1–10/1 | 239,035 | 9.8% | -29.8% |
| 5· 10/1+ | 331,326 | 4.3% | -39.4% |
Each angle backtested with A/E (lift over base rate) and a first-half/second-half stability split. KEEP = positive ROI, stable, n≥300.
| Angle | Bets | Win% | ROI | A/E | ROI h1/h2 | Verdict |
|---|---|---|---|---|---|---|
| Flat-bet odds-on favourites | 2,776 | 54.0% | -21.9% | 4.41 | -17.4/-25.7 | DROP |
| Flat-bet mid-price band (5/2–5/1) | 254,060 | 18.9% | -25.6% | 1.54 | -24.0/-27.4 | DROP |
| Flat-bet longshots 10/1+ | 331,326 | 4.3% | -39.4% | 0.35 | -36.6/-42.0 | DROP |
The research agent proposes selection angles and self-validates them. Connection angles (hot trainers/jockeys) are selected on the first half of history and proven out-of-sample on the second — the leakage-safe way to tell a real edge from a dredged one. 0 angle(s) survived → promoted to the Angle Library.
| Proposed angle | Bets | Win% | ROI | A/E | split | Verdict |
|---|---|---|---|---|---|---|
| Hot trainers (h1-selected) — OOS on h2 | 5,485 | 21.1% | -30.9% | 1.77 | OOS h2 | DROP |
| Hot trainers × mid-price (convergence) — OOS | 2,291 | 20.8% | -30.7% | 1.75 | OOS h2 | DROP |
| Profitable trainers h1 (by ROI) — OOS on h2 | 4,839 | 11.9% | -31.9% | 1.0 | OOS h2 | DROP |
| Hot jockeys (h1-selected) — OOS on h2 | 13,649 | 15.6% | -29.4% | 1.32 | OOS h2 | DROP |
| First-time Lasix | 42,561 | 10.4% | -31.6% | 0.85 | -31.2/-32.0 | DROP |
| Blinkers ON (first time) | 12,855 | 11.0% | -30.2% | 0.9 | -30.5/-27.3 | DROP |
| Inside post (1-3) in small field (≤7) | 158,922 | 16.1% | -27.1% | 1.32 | -24.9/-29.5 | DROP |
| Far outside post (10+) | 72,334 | 6.9% | -51.0% | 0.57 | -46.4/-54.9 | DROP |
The research agent runs as a standing loop: each cycle it re-validates the Angle Library, promotes new out-of-sample survivors, and retires members that decay (two straight failures). Current library: 0 angle(s) — and if the market gives up nothing, it stays empty and says so. That honesty is the product.
| Cycle | Evaluated | Survivors | Promoted | Retired | Library |
|---|---|---|---|---|---|
| 2026-06-07 | 11 | 0 | — | — | 0 |
| 2026-06-07 | 11 | 0 | — | — | 0 |
| 2026-06-07 | 6 | 0 | — | — | 0 |
| 2026-06-07 | 4 | 0 | — | — | 0 |
| 2026-06-07 | 11 | 0 | — | — | 0 |
The honest answer to "everything's priced in." We compare angle vs non-angle horses at the same price (uplift / matched control) — positive, consistent within-price lift = an edge the market hasn't priced; ~0 = correlational. Candidate causal edge: Blinkers OFF. Caveat: matched on the morning line, not the closing line — a strong candidate to confirm against final odds (A1), not yet a proven beatable edge.
| Angle | Bets | Price-adj lift (pts) | Verdict |
|---|---|---|---|
| Blinkers OFF | 6,809 | +2.38 | CAUSAL EDGE |
| Inside post, small field | 158,922 | +0.93 | PRICED IN |
| Bute+Lasix | 42,670 | +0.54 | PRICED IN |
| Blinkers ON (first time) | 12,855 | -0.09 | PRICED IN |
| First-time Lasix | 42,561 | -0.22 | PRICED IN |
| Far outside post | 72,334 | -2.22 | NEGATIVE |
The agent sets its own agenda — agenda generated by LLM (self-directed) — back-tests each hypothesis out-of-sample, promotes survivors, and picks the next focus. Honest: most novel ideas die against an efficient market. Next focus: Short-priced Blinkers-Off bounce (equipment removal signals trainer confidence i.
| Self-proposed hypothesis | Bets | ROI | A/E | Verdict |
|---|---|---|---|---|
| Short-priced Blinkers-Off bounce (equipment removal signals | 241 | -27.8% | 2.47 | DROP |
| Longshot First-Time Lasix in large field (public ignores, ov | 7,167 | -38.8% | 0.27 | DROP |
| Extreme outside post in small field (post 7+ forces wide tri | 37,819 | -28.7% | 1.02 | DROP |
| Both-medication horses at short price in mid-size field (BL | 2,888 | -17.7% | 2.09 | DROP |
| Blinkers-On longshot in medium field (public ignores equipme | 3,236 | -31.8% | 0.33 | DROP |
| Inside post (1-3) in very large field at mid-long odds (rail | 14,591 | -33.8% | 0.64 | DROP |