Confidently
wrong.

28 February journal; 7 March project README and audit log.
The journal records Flemington and Randwick spreadsheets. The README sets a $100 starting balance and $1,000 target by March 2027. The project is called Confidently Wrong 2026.

bankroll_history.csv; 7 and 14 March result files.
Balances shown are the archive’s entries, not independently verified account statements. The separate performance summary was still stuck on the earlier $52.63 figure. This ledger is one early experiment, not a complete account of Phil’s racing spend.

21 March daily journal.
Kev records confusing the race identity and failing to challenge implausibly uniform data. His explanation blamed test data from the provider; that cause has not been independently established.

Hong Kong report and QA report, written 13 April UTC / 14 April Sydney.
The narrative lists LUCKY ARCHER in slots 10–14. QA says initial parsing counted 448 horses, later claims 120, and assigns 92% overall confidence. The number is the agent’s self-assessment, not a measured accuracy rate.

18 April retrospective; 30 May setup review.
The later review clarifies zero-for-ten refers to abbreviated cold batch top picks, not every full live analysis. The April aggregate contains inconsistent score labels, so this story does not pool its win rates.

19 April Sunshine Coast retrospective.
The Irish was flagged in commentary but not promoted into the ranked four, then won. The note made numerical promotion of live signals central to the next version.

25 April retrospective.
Records three trifecta hits among ten logged boxes. Several races were unlogged and dividends were incomplete. Its positive-day claim is not treated here as a reconciled profit figure.

2 May pre-retrospective summary.
Records the notification failure, disabling 14 remaining fires, switching to the active conversation and continuing in calibration mode after Phil stopped active betting. The header and body disagree about whether he stopped after Race 3 or Race 4, so no exact stopping time is asserted.

22 August research thread and Kev’s final sourcing report.
The thread records the failed launch, five concurrent workers exhausting the token allowance, bounded recovery, an initial verifier pass and four material misses found afterwards. Ten tasks ultimately completed, including correction and final acceptance.

The consequential error.
Date-keyed pages had been treated as proof of preserved pre-race data. They were not. A provider’s horse ID had also been confused with a verified official identity. These were corrected in the final report.

Codex experiment report, reproduction and invalidated gains.
All 717 handoff checksums passed. BLS v5.1 was reproducible; later simplified weights were not reproducible from the preserved clean matrix.

Two invalidated gains.
Published jockey/trainer aggregates appeared to improve log loss by 0.02451 but lacked point-in-time integrity. Reversed context summaries appeared to improve it by 0.16519 and were target-label dependent. Neither entered the frozen candidate.

Codex experiment report, 22 August.
Rolling validation improved from 2.04319 to 2.02019. Gains across the three chronological folds were +0.03108, +0.03264 and −0.00153. The exposed 337-race block was not an untouched test.

Locked day-layer audit.
610 races across 83 meeting clusters. The quality-adjusted layer worsened log loss by 0.00350; the inherited fixed-count rule worsened it by 0.00749. The specified Step 4A rule was deleted under the predeclared decision. This does not disprove every possible use of within-day information.

25 August blocked-opening receipt.
All three independently audited dated outcome archives were unavailable. No canonical final comparison or final model verdict was produced. Three morning operational misses were retained; no successful morning manifest existed.

Frozen prospective records.
94 races, 1,135 runners, 100,000 finishing-order simulations per race per model: 18.8 million orders across two models. These are simulated possibilities, not observed trials. The blockage says nothing conclusive about which model was better.

Codex history, 29 August.
Phil’s messages arrived at 12:16 and 12:17 AEST. He later supplied final results and explicitly asked for official confirmation afterwards. The quotations above are verbatim.

Claude end-of-day retrospective and controlling audit.
Claude’s local Race 10 shortlist omitted Lisztomania; the controlling final-field card included it. The primary card was preserved. The audit also recorded a missed pre-jump review rather than manufacturing one after the race.

29 August Merrylands R3 coverage receipt.
Nine runners in the field, eight backed with the favourite excluded, four $1 markets per backed runner. It enumerated 3,024 ordered top-four outcomes; 2,453 (81.1177%) exceeded the $32 stake.

The important distinction.
Multiplicative normalisation of each quoted market gave a price-based expected return of $26.15, or −$5.85 net. Prices were held fixed; late deductions and dead heats were not modelled. Neither outcome counts nor this price diagnostic prove an executable edge.

29 August end-of-day audit against archived Australian Turf Club results.
Ten races completed, nine scored. Race 2 was excluded because four official starters were absent from the locked prediction. Three top-pick wins and seven winners in four are both measured over the nine eligible races.

Probability score.
Mean winner log loss 2.248171. Equal-chance benchmark is the mean of log(field size) over the same nine races, approximately ${uniform.toFixed(4)}. Claude’s retrospective instead quoted 2.2034; this page recomputes it from the official starter counts for the nine eligible races. A lower loss is better. This is one meeting, not an estimate of long-run betting returns.

Correcting the track input did not settle the argument.
The Good 4 diagnostic was very slightly better on scoring rules over Races 4–10, but winner-in-four fell from five of seven to four of seven. A sensible correction did not deliver a clear practical gain.