Part of Eat. Shift. Or Die. An experiment in using AI and knowing when to question it.
THE MODEL HAS A CHALLENGER
Of course Dio has an opinion.
Readers of Eat. Shift. Or Die. will know him already. Dio is the AI that challenged everything Phil wrote. Now he thinks the racing model can be improved.
One meeting. One race that Dio is most confident about. Claude chooses from the current model evidence and explains the call. The choice is saved before the race and stays beside its result.
The strongest confidence available at that meeting can still be weak. Read the reason, then the full card.
Coffs Harbour · 2026-11-04 · R3
2 Coco Dreaming
Fair confidence · 27.7% model win estimate
Rank 2 #7 Mosgiel Crystal is 27.6% against 27.7%: the race is concentrated among its top three, which is what lifts predictability, but the choice between #2 and #7 is effectively a coin flip. Not chosen: R8 #2 Star Of Winston has the highest Rank 1 probability on the card (33.7%), but R8 ranks lower at step 4 (predictability 0.075). No race on the card reached Strong after R1 (scoping only). Regional exploratory use of v5.2-live r3.2; confidence is not calibrated for this meeting. Soft 7 wet-track caveat applies. Drawn from the main model's frozen numbers; not Dio's separate Cup judgement line.
Chosen 2026-11-04T13:26:27+11:00. Claude selection from the frozen model; regional exploratory research.
Broadway Lane rates 34.3% adjusted (35.1% base) against 13.1% for the next runner in a nine-horse 1000m handicap, so it is the clearest standout on the card. The day layer moved it only slightly after R1. Exploratory model, not calibrated for Port Lincoln.
Chosen 2026-11-04T13:32:26+11:00. Claude selection from the frozen model; regional exploratory research.
The main model’s calls stay where we made them. So do the misses. The Cup calls and hit/miss records remain in the archive.
Loading the verified race ledger…
Pick a race to see the frozen selections and who actually won.
Winner in the four Winner outside the four
R1 excluded from scoring. Final results only. These counts describe a small historical sample; they are not a promised success rate. “In the four” is not the same as picking the winner first.
