Evidence

The evidence room

Every figure, with its sample, source and status

The homepage tells the story. This page proves it. The ledger audits it.

Everything below is drawn from real captures and governed research artefacts. Each chart carries its sample size, its source, the date it was computed and an honest research status. Where the data does not exist, the gap is published rather than papered over.

The largest single analysis behind the programme benchmarked 531,115 runner-race records against the market's own implied probability — and returned a well-powered null: no structural predictor beat the price. That null is published, not hidden. The grading instrument has since assessed 275 races prospectively.

The market before the off

How much movement was still left in the market?

At 60 minutes before the race, the favourite's implied chance was on average 4.48 percentage points away from where it would stand one minute before the off. By three minutes out, 1.43. The field average moves roughly half as far at every stage.

Prices can move in either direction. This measures the SIZE of the remaining change — not whether the horse shortened or drifted.

FavouriteField average
0.0 1.3 2.6 4.0 5.3 percentage points still to move 60 30 15 10 5 3 60 minutes before the off 3 minutes before the off At 60 minutes, the favourite was on average 4.48 percentage points of implied win probability away from where it would stand one minute before the off. At 30 minutes, the favourite was on average 4.09 percentage points of implied win probability away from where it would stand one minute before the off. At 15 minutes, the favourite was on average 3.39 percentage points of implied win probability away from where it would stand one minute before the off. At 10 minutes, the favourite was on average 2.61 percentage points of implied win probability away from where it would stand one minute before the off. At 5 minutes, the favourite was on average 1.94 percentage points of implied win probability away from where it would stand one minute before the off. At 3 minutes, the favourite was on average 1.43 percentage points of implied win probability away from where it would stand one minute before the off. At 60 minutes, the field average was on average 2.16 percentage points of implied win probability away from where it would stand one minute before the off. At 30 minutes, the field average was on average 1.97 percentage points of implied win probability away from where it would stand one minute before the off. At 15 minutes, the field average was on average 1.59 percentage points of implied win probability away from where it would stand one minute before the off. At 10 minutes, the field average was on average 1.23 percentage points of implied win probability away from where it would stand one minute before the off. At 5 minutes, the field average was on average 0.89 percentage points of implied win probability away from where it would stand one minute before the off. At 3 minutes, the field average was on average 0.69 percentage points of implied win probability away from where it would stand one minute before the off.

Favourites move roughly twice as much as the field average.

  • Samplen = 784 · 7,077 runner observations per checkpoint
  • SourceT-gate exchange receipts, 2026-07-16 to 2026-08-20
  • As at2026-08-20
  • StatusHISTORICAL

The market record behind the research

Elite Pass captures the exchange at fixed checkpoints before every race it follows, and keeps the full board each time — not just the favourite. Across 32 racing days that is 7,261 captures over 1,180 races. An eighth checkpoint at two minutes was added recently and is excluded here, because it does not span the window.

  • Samplen = 1,180 · 7,261 captures · 64,967 runner price observations
  • SourceT-gate exchange receipts, 2026-07-16 to 2026-08-20
  • As at2026-08-20
  • StatusHISTORICAL
The grading instrument

What the instrument does with a card

Across seven card-days the grading instrument assessed 275 races. It issued 72 AMBER — its lower tier, meaning the evidence converges but not completely — and declined to grade the other 203. It has never issued a GREEN. GREEN is currently unreachable because one required component is not frozen, and the instrument will not award a grade whose inputs are incomplete. A grade is evidence convergence, not a prediction: it applies to one runner and it is not a bet.

  • Samplen = 275 · 7 card-days
  • SourceDAILY_GRADES_2026-08-* (frozen TARGET_GRADE_1.0)
  • As at2026-08-21
  • StatusPROSPECTIVE
EP Rank research state

EP Rank: no fitted model

This is the chart we would most like to show you, and the one we will not fabricate. EP Rank needs a fitted weighting between its component views. The two evidence lanes that would justify one currently stand at zero of the required sample, so the scoring function refuses to return a value at all. When it fits, it will be published here with its sample size like everything else.

  • Sampleno admissible sample
  • Sourceelite_pass_structure.py (WEIGHTS = None; score() raises NotFittedError)
  • As at2026-08-21
  • StatusDATA_GAP

The scoring function is deliberately locked: its weight vector is unset, and calling it raises an error rather than returning a plausible-looking number. The two evidence lanes that would justify a weighting currently stand at zero of their required samples. When a fitted model exists, its output will be published here with its sample size like everything else.

Research findings — full frames

The public findings on the Research page, with their complete samples, sources and statuses. The favourite analysis is documented internally in FAVOURITE_RESEARCH_REPORT_1.0 (corpus, eligibility, joint-favourite handling, exclusions, confidence intervals).

Across 60,671 races, favourites won 35.0% — and failed to win 65.0%.

The favourite is usually the horse the market considers most likely to win. Most likely does not mean certain: across every race in our historical corpus with a complete Betfair book and a single clear favourite, the favourite lost nearly twice as often as it won.

  • Samplen = 60,671 · unique BSP favourites; 300 joint-favourite races reported separately
  • SourceFAVOURITE_RESEARCH_REPORT_1.0 (RBD historical cache, recomputed 2026-08-21)
  • As at2026-08-21
  • StatusHISTORICAL

Shorter favourites win more often — almost exactly as often as the market expected.

Are all favourites equal? No — a 6/4 favourite and a 5/1 favourite are different propositions, and the win rates show it. But at every price band, the actual win rate lands within a percentage point of what the market's own prices implied. The shorter the favourite, the more often it wins. The market knows that too.

  • Samplen = 60,671 · five BSP bands
  • SourceFAVOURITE_RESEARCH_REPORT_1.0 (RBD historical cache, recomputed 2026-08-21)
  • As at2026-08-21
  • StatusHISTORICAL

Favourites are more vulnerable in big fields — and the market prices that in too.

In small fields the favourite wins nearly half the time; with sixteen or more runners, barely a fifth. That is a real, pre-race, structural difference — and at every field size the market's expectation tracks it. Field size alone tells you where favourites fail more often. It does not, on its own, tell you anything the price hasn't already said.

  • Samplen = 60,671 · pre-defined field-size bands
  • SourceFAVOURITE_RESEARCH_REPORT_1.0 (RBD historical cache, recomputed 2026-08-21)
  • As at2026-08-21
  • StatusHISTORICAL

Good form finds better horses. The market already knows they're better.

Runners our morning rating ranks top win far more often when the market also likes them — a gradient from 41.8% down to 11.0%. Convincing, until you ask what the price alone predicted: 40.7% down to 12.0%. The gradient is the price. In a separate half-million-runner benchmark, no structural predictor we tested beat the market either.

  • Samplen = 2,487 · races with a unique top-rated runner; gradient measured against BSP
  • SourceRBD_MARKET_CONDITIONAL_ADJUDICATION_2026-08-19 (governed adjudication)
  • As at2026-08-19
  • StatusHISTORICAL
Provenance

Where every number comes from

Each figure published anywhere on this site carries its own definition, source, date and governance status. Nothing is a rounded-up marketing number.

FigureWhat it countsSourceAs atStatus
n = 0EP Rank: no fitted model
This is the chart we would most like to show you, and the one we will not fabricate. EP Rank needs a fitted weighting between its component views. The two evidence lanes that would justify one currently stand at zero of the required sample, so the scoring function refuses to return a value at all. When it fits, it will be published here with its sample size like everything else.
elite_pass_structure.py (WEIGHTS = None; score() raises NotFittedError)2026-08-21DATA_GAP
n = 60,671Favourites are more vulnerable in big fields — and the market prices that in too.
In small fields the favourite wins nearly half the time; with sixteen or more runners, barely a fifth. That is a real, pre-race, structural difference — and at every field size the market's expectation tracks it. Field size alone tells you where favourites fail more often. It does not, on its own, tell you anything the price hasn't already said.
FAVOURITE_RESEARCH_REPORT_1.0 (RBD historical cache, recomputed 2026-08-21)2026-08-21HISTORICAL
n = 60,671Shorter favourites win more often — almost exactly as often as the market expected.
Are all favourites equal? No — a 6/4 favourite and a 5/1 favourite are different propositions, and the win rates show it. But at every price band, the actual win rate lands within a percentage point of what the market's own prices implied. The shorter the favourite, the more often it wins. The market knows that too.
FAVOURITE_RESEARCH_REPORT_1.0 (RBD historical cache, recomputed 2026-08-21)2026-08-21HISTORICAL
n = 60,671Across 60,671 races, favourites won 35.0% — and failed to win 65.0%.
The favourite is usually the horse the market considers most likely to win. Most likely does not mean certain: across every race in our historical corpus with a complete Betfair book and a single clear favourite, the favourite lost nearly twice as often as it won.
FAVOURITE_RESEARCH_REPORT_1.0 (RBD historical cache, recomputed 2026-08-21)2026-08-21HISTORICAL
n = 2,487Good form finds better horses. The market already knows they're better.
Runners our morning rating ranks top win far more often when the market also likes them — a gradient from 41.8% down to 11.0%. Convincing, until you ask what the price alone predicted: 40.7% down to 12.0%. The gradient is the price. In a separate half-million-runner benchmark, no structural predictor we tested beat the market either.
RBD_MARKET_CONDITIONAL_ADJUDICATION_2026-08-19 (governed adjudication)2026-08-19HISTORICAL
n = 275What the instrument does with a card
Across seven card-days the grading instrument assessed 275 races. It issued 72 AMBER — its lower tier, meaning the evidence converges but not completely — and declined to grade the other 203. It has never issued a GREEN. GREEN is currently unreachable because one required component is not frozen, and the instrument will not award a grade whose inputs are incomplete. A grade is evidence convergence, not a prediction: it applies to one runner and it is not a bet.
DAILY_GRADES_2026-08-* (frozen TARGET_GRADE_1.0)2026-08-21PROSPECTIVE
n = 784How much movement was still left in the market?
At 60 minutes before the race, the favourite's implied chance was on average 4.48 percentage points away from where it would stand one minute before the off. By three minutes out, 1.43. The field average moves roughly half as far at every stage.
T-gate exchange receipts, 2026-07-16 to 2026-08-202026-08-20HISTORICAL
n = 1,180The market record behind the research
Elite Pass captures the exchange at fixed checkpoints before every race it follows, and keeps the full board each time — not just the favourite. Across 32 racing days that is 7,261 captures over 1,180 races. An eighth checkpoint at two minutes was added recently and is excluded here, because it does not span the window.
T-gate exchange receipts, 2026-07-16 to 2026-08-202026-08-20HISTORICAL
71,626Races in the historical research corpus
Distinct historical races held in the research corpus behind the programme's large-n studies: 71,626 races across 78 British and Irish courses, 2002-03-27 to 2026-06-28, keyed date|track|time. An inventory figure — what the archive holds. The convenience-corpus caveats travel with it: several exports truncated at 60,000 rows and the top five courses supply around 44% of complete-field races, so it is not a representative sweep of two decades of racing.
RBD historical corpus (elite_pass_cache, 27 RBD_Data CSVs)2026-08-21RECOMPUTED
633,474Historical runner-race records analysed
Runner-race records carried through the completed place-market study (PLACE_ACCUMULATOR_VERDICT) — the largest single analysis the programme has run. 633,474 rows, each a unique (date, track, horse) key, drawn from a research corpus of 71,626 races across 78 British and Irish courses, 2002-2026. The corpus is a convenience corpus, weighted toward all-weather and major courses — it is not a representative sweep of two decades of racing, and the figure is a count of records analysed, not of races.
place_dataset.csv -> PLACE_ACCUMULATOR_VERDICT (completed study)2026-08-21RECOMPUTED
7,261Pre-race market captures
Timestamped pre-off exchange captures across 1,180 races and 64,967 runner-price observations, 32 racing days, 2026-07-16 to 2026-08-20. Each capture keeps the full back/lay board, not one price. Genuine exchange observations — never Industry SP, forecast or tissue prices. Excludes the recently added two-minute checkpoint, which does not span the window.
T-gate capture receipts (tgates_*.jsonl)2026-08-20RECOMPUTED
1,180Races with a timestamped market receipt
Distinct races for which Elite Pass holds its own timestamped pre-off exchange capture — a real back/lay board recorded at fixed checkpoints before the off. 32 racing days, 2026-07-16 to 2026-08-20, complete days only. These are genuine exchange observations, not Industry SP, forecast or tissue prices.
T-gate capture receipts (tgates_*.jsonl)2026-08-20RECOMPUTED
275Races assessed by the grading instrument
Races the frozen grading instrument has assessed across 7 card-days (6, 7, 16, 17, 19, 20 and 21 August 2026). It issued 72 AMBER and 203 NO_GRADE, and has never issued a GREEN. A grade applies to ONE runner in a race; it is not a ranking of the field, not a probability and not a bet.
DAILY_GRADES_2026-08-{06,07,16,17,19,20,21}.json2026-08-21RECOMPUTED
531,115Runners benchmarked against the market
The benchmark subset of the same completed study: runner-race records carrying a genuine Betfair exchange place price, each scored against the market's own implied probability rather than against nothing. The study was well-powered (minimum detectable effects 0.18-0.43pp) and returned a null: no structural predictor beat the price. That null is published, not hidden.
PLACE_ACCUMULATOR_VERDICT (benchmark subset, n = 531,115)2026-08-21GOVERNED_ARTEFACT
271Racing systems & theories analysed
Racing systems, theories and analytical approaches examined by the founder over the life of the programme, including work predating and sitting outside this repository. Asserted by the operator; not derivable from the published artefacts.
OPERATOR_RESEARCH_REGISTER2026-08-21OPERATOR_CONFIRMED
0Systems yet to clear our validation bar
Tested ideas that have reached a validated state — the only state that would grant decision authority. Zero, all-time. Confirmed five independent ways: no promotion git tags exist; all promotion-docket entries sit at the entry stage with no preregistration filed; the global decision-authority constant reads LOCKED_NO_VALIDATED_EDGE; the hypothesis registers carry no validated row; and research capital remains GBP 0.00.
PROMOTION_PROTOCOL + governance/docket.json + repository sweep2026-08-21RECOMPUTED