Elite Racing Group Ledger

Ledger opened 19 August 2026 · 1 day settled

Every morning's selections are hashed and published before the first race; every evening they are settled against the exchange. No tips are sold. No stakes are placed. The record exists so that any claim made later can be checked against something that was fixed beforehand.

19 August 2026

FREEZE & SETTLEMENT

File SHA-256 — this is the one you can check

c8a54ce9 e6400743 388541e9 557606dc
db34a418 372026d1 7bbb54e6 2da35cbe

Content digest recorded at freeze, inside the file

f2a6f869 460b85ee de85059f 26b81d57
0ce0e058 840e09d8 9a32ce9d 96a8a59b

Frozen at10:39 BST

First race13:50

Races36

Runners407

Source filee7c7e837

Outcome readfalse

Check it yourself

Download the freeze file and hash it. It should reproduce the file digest above, byte for byte.

shasum -a 256 PROSPECTIVE_FREEZE_2026-08-19.json

The second digest is the instrument’s own content seal, written into the file by the freeze code before racing. A file cannot contain its own hash, so the two necessarily differ — the content seal is not reproducible with shasum alone, and is shown for completeness rather than as something you can independently verify here.

Download the freeze file · settlement file

Hash verified at settlement33 races settledZero outcome fields at freeze

Hypothesis	Races	Wins	Expected	Endpoint
H1 · all unique RBD1	33	6	8.01	−2.01
H2 · RBD1 = BSP favourite	17	—	—	−2.31
H3 · RBD1 ≠ BSP favourite	16	—	—	+0.29

A dash means the settlement artefact does not record that field, not that the value is zero. H2, H3 stored the endpoint only.

Running total

ACCRUAL

Nothing here is a result yet, and saying so is the point. The sample needed to detect an effect worth detecting is 2,616 races. Until then these are entries in a register, not evidence of anything.

Accrued33 / 2,616

Progress1.3%

Days settled1

Targets issued0

Selections0

Capital£0.00

What has been ruled out

NULL REGISTER

Most of this project's output is negative results, published here because nobody else in the sport publishes them. Each line is a question that was asked properly and answered no.

Does the morning rating add anything once the market price is held constant? z = −0.42 2,487 races with a unique top-rated horse. Residual −0.35pp against race-normalised BSP. Adding the rating to a market-only model made out-of-sample log-loss worse.

Is a horse the market "accepts" a more reliable anchor because our model also picked it? no effect Accepted runners win 30.2% against 5.5% for rejected ones — a six-fold gradient. The identical gradient appears for horses the model never touched (30.7% / 5.6% expected). The anchor is the price.

Does the 8am rating anticipate where the price moves during the day? z = −0.03 Tested against a live positive control: eventual winners shorten at z = +5.57, so the test can see real information. It would have caught a 3.85% average price move. The rating moves it by 0.05%.

Do horses whose price collapses pre-race underperform their final price? z = −1.15 Raw steam-fade looked real at −0.91pp. It vanished once compared within the same price band — it was band composition, not overreaction.

Is there value in the place market where the win market is efficient? refused Measured directly: residuals −0.74 to +2.80pp, all inside the confidence interval. Harville conversion permanently refused — it overstates favourites.

Does the draw carry information on its own? killed Powered held-out null at n = 477. Closed, and not reopened.

Three findings that looked significant were killed by the project's own checks before they were reported: a power-recalibration artefact, a monotone field-size artefact, and a market-rank cell that failed period replication (−10.01pp in test, +0.03pp in training). They are logged rather than deleted.

The reference market

WHY THIS IS HARD

Betfair's starting price is the benchmark every result is measured against. Across 25,140 runners it calibrates at a slope of 0.9778 with an expected calibration error of 0.0034 — no price band is off by more than 3.6 percentage points. It is a professionalised counterparty, and beating it is the whole difficulty.

To detect a miscalibration of…	Races required
30% relative	181
10% relative	1,625
5% relative	6,499

At α = 0.05 and power 0.80, against a market favourite rate of p ≈ 0.326. Sample size is the bottleneck, not ideas.

Standing terms

GOVERNANCE

These are enforced in code, not by intention. The alert layer machine-blocks betting language; the freeze refuses to run if a single outcome field is populated on the morning card; a missed market checkpoint is recorded as a gap and never backfilled from a later price.

Capital £0.00 No staking No validated edge No tips sold Retrospective results are hypothesis-generating only No claim of profitability is made anywhere on this page