Every programme, its status, and the finding that put it there. Updated 2 September 2026, 22:40. Every verdict below was pre-registered before its data was seen; commits are in the estate repos.
Programmes closed on evidence
6
market making ×2, racing ×3, stats engine
Measuring
1
Pinnacle vs Smarkets, gate fixed, first verdict Monday
Prediction tally
13 / 10
right / wrong, on the record before each result
Real money lost, all programmes
$29.77
everything else DRY or paper. RULES 7.7 stands.
| Programme | Status | Finding | Ref |
|---|
| Smarkets market making | CLOSED | Structurally impossible on three independent grounds: ladder arithmetic, adverse selection on realised fills matching P&L, and commission structure. Terminated 1 Sept. | SMK_TERMINAL_FINDING_20260901 |
| Polymarket sports maker | DRY | Adverse selection confirmed by markout: both PHI and QUEUE bracket ends negative. Sharper anchors now fed from the paid Odds API; that improves fair values, it does not repair the structure. | MAKER_MARKOUT_RESULTS_20260901 |
| Racing: residual signal | CLOSED | Eight declaration fields, none correlates with the market's error. p-values textbook-uniform. 80% power at 0.20 z. Positive control fired at 38.6 sd. | RESIDUAL_SIGNAL_RESULTS_20260902 |
| Racing: price dynamics | CLOSED | Market log-loss improves 2.1% across the entire final hour (football 0.4%). No early window. Drift does not predict the closing error. | PRICE_DYNAMICS_RESULTS_20260902 |
| Racing / football: one-sided betting | CLOSED | Lay-the-outsider sits on zero with tight CI in both sports; the ladder is fair. Back-the-favourite resolved on 48,592 historical bets: +0.45%, no edge. | ONESIDED_RESULTS, CLOSING_LINE_RESULTS |
| Value betting, historical | CLOSED | +5.99% then +8.33% on held-out data, both artefacts: market-max prices were never simultaneous (41.46% sub-1.00 overrounds). Simultaneous capture drops that to 5.56%. Dataset burned for this hypothesis. | REPLICATION_RESULTS_20260902 |
| Stats engine (football) | CLOSED | Poisson attack/defence model captures 77.7% of Pinnacle's information and sits 2.2% behind it. Gap is flat across every observable situation; lineups and injuries would close ~4.5% of it. Model built, measured, and beaten by the price. | STATS_ENGINE_RESULTS, GAP_DECOMPOSITION |
| Soft-book value, live | CLOSED | Five major UK books on EPL and Championship, 393 simultaneous lines: zero beat Pinnacle by 1%. Median line 2.35% below fair. | this page, 2 Sept |
| Pinnacle vs Smarkets | MEASURING | Smarkets showed more dispersion vs Pinnacle than soft books did (15 of 87 lines positive after commission at 20h+). Rig live: hourly Pinnacle, 60s Smarkets with size, joined within 90s. Gate: net edge >2%, ≤2h to kickoff, ≥£50 stake, ≥1 per match-day over 5 match-days, positive closing-line value. Registered prediction: NO DIFFERENCE. | OPERATING_SMARKETS_VALUE, live panel |
| Survivor colonies (open, BTC) | DRY | Negative in paper at last review. Nothing from this week bears on their markets; no mentor boost posted for want of evidence. | elder_log.md |
| # | Question | Answer |
|---|
| 1 | Can we make markets? | No. Both venues, independently. |
| 2 | Does the racing price leave anything on the table? | No. Not in static fields, not in movement, not one-sided. |
| 3 | Can historical odds prove value betting? | No — the free data's best-price column is a maximum across time and cannot measure it. |
| 4 | Can a stats engine beat the price? | It gets 77.7% of the way and stops 2.2% short, uniformly. Public data does not close the rest. |
| 5 | Is there anything left? | One thing: whether the exchange is looser than Pinnacle where it counts. Measuring now, verdict Monday. |
| What | Fix |
|---|
| Smarkets trader at 4 cycles/hour | ANALYZE on the tape restored 144/hour; query from >4 min to 16 ms. No hardware bought. |
| Backup livelock | VACUUM INTO replaces .backup; busy_timeout added; two databases added to the set. |
| Outcome capture | Settlement sweep built and wired; result detail decays within days so it runs every 10 min. 780+ markets labelled. |
| Collector compliance | Throttled 9.6 to 2.0 req/min; trades and volumes off; the maker programme they fed is closed. |
| Collector football tracking | Substring league filter admitted Egypt as major; event listing read one page and never reached Friday fixtures. Both fixed; EPL was tracked 11 times in three weeks before, 11 major matches tonight. |
| Odds API | Paid 20K plan; maker and survivor caps raised from 6 and 15 to 100 and 50 per day. |