Test-mechanism: checked, recovered and still within 2 lengths carries the H19 lift

Reasoning

Why this hypothesis exists: what the lab saw, the question it asked, the mechanism it proposed, what it predicted, and the result that would prove it wrong. Written before the test was run.

1. Observed
   • H-48a2eca6 (Test-mechanism: the edge sits with horses the market did not shorten a) passed discovery but was falsified: 2 of 5 attacks broke it; IRE: too few bets to judge. O-0bee0352
   • H-b595a70a (Sharpen: troubled narrow loser back within 35 days) passed discovery as a niche: A/E 1.19 on 200 bets (p 0.134), same-race control 0.96. O-0fc94a65
   • H-4882f53a (Generalise: checked mid-run and beaten 2 to 5 lengths still beats its ) was predicted positive but came out the other way: A/E 0.8 on 263 bets (p 0.148 for the opposite direction). O-fd5d7674
2. AskedAmong H19 runners, do those noted as recovering or finishing strongly after the check beat their price by more than those that were not? answered
3. MechanismIf the edge is ability lost to the check, a horse that recovered and ran on to finish within 2 lengths showed the most spare ability. One that was checked and never picked up may simply have been beaten.
4. Why the market misses itThe form figure and the beaten distance look the same for both kinds of run. The recovery is only in the in-running comment, which the price barely uses: the price-movement split showed no allowance for the trouble at all.
5. Predictedpositive, A/E 1.30+ on 80+ bets, above H19 runners without a recovery note
6. Falsified if

Mechanism and rationale

If the edge is ability lost to the check, a horse that recovered and ran on to finish within 2 lengths showed the most spare ability. One that was checked and never picked up may simply have been beaten.

The rule

last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND last_recovered = 1

Reads: last_momentum_interrupted, last_position, last_beaten_lengths, last_recovered

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

8f345d41fa64064489400a0aaf9ba850ab76fd1e647428a14a5e4abc199d4f29

Frozen

not frozen

The definition can still change.

Confounders

• Comment-writing style differs by country (GB vs IRE) and by race reader, so last_recovered may reflect who wrote the comment more than how the horse ran
• Overlap with late-headway closers: 'recovered' comments often come with 'made headway'. Closers without trouble were priced correctly (0.969 on 5,540), but the overlap still needs checking
• Small sample: the recovered subset of H19's 242 bets is likely under 150, so the discovery floor may not be met even if the effect is real
• Price band: in other families a single short-price band has carried apparent lifts, so check whether the lift is spread across prices
• Period stability: H19 rose from 1.07 to 1.26 across periods, so a subset could ride one period
• SP vs exchange: H19's edge was mostly at BSP, so an SP-measured A/E may understate it or behave differently

Control

{
  "description": "The key comparison is the other half of H19: runners checked last time and beaten 2 lengths or less whose comment carries no recovery note. If the mechanism holds, they should score clearly below the recovered group (near par, about 1.0 to 1.1), while the recovered group sits at 1.30 or more. The same-race other runners should sit near 1.0, which shows the lift is not a property of the races. If the unrecovered half scores as well as or better than the recovered half, the recovery note carries no information and the H19 lift belongs to the check flag alone.",
  "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND (last_recovered = 0 OR last_recovered IS NULL)"
}

Discovery result

pass broad rule (niche) · 16 Oct, 14:52

Bets

31

9 winners

A/E

1.59

95%: 0.73 to 3.02

SP return

+12.3%

level stakes

Betfair SP return

+3.9%

30 bets with a BSP

Strike rate 29.0%. p = 0.12. Same-race control: the other runners in these races had A/E 0.87, so the selection beat its own races. Zone: discovery. Evaluator eval-1.0, data 6a4e98be18e83933

Falsification

Each attack is a deliberate attempt to kill the idea. 0 of 5 survived.

• perturbation: Margin widened to 2.5 lengthsinfo

  38 bets · A/E 1.42 (0.65 to 2.70) · control 0.92 · 16 Oct, 14:55

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-2711–

  after 2026-04-2727–

• perturbation: Margin tightened to 1.6 lengthsinfo

  26 bets · A/E 1.74 (0.75 to 3.42) · control 0.84 · 16 Oct, 14:55

  Breakdown

Sealed holdout (vault)

Not queried. Only a frozen idea can spend its one query against the 2020-25 holdout.

Forward record

No forward record. Shadow tracking starts after the vault and counts only races run after the freeze.

Other experiments

• neighbour: > 1 -> 0 (-20%)pass

  47 bets · A/E 1.06 (0.48 to 2.01) · control 0.98 · 16 Oct, 14:55

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-2711–

  after 2026-04-2736–

• neighbour: > 1 -> 2 (+20%)pass

  21 bets · A/E 1.47 (0.54 to 3.21) · control 0.88 · 16 Oct, 14:55

  Breakdown

  bets

Lineage

Agent runs

• designer v1ok · 24s · 16 Oct, 14:51

  Output (schema-checked)

  {
    "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND last_recovered = 1",
    "refinements": [
      {
        "label": "14/1 or shorter (price-band confounder; edges in this family have died at big prices)",
        "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND last_recovered = 1 AND sp_decimal <= 15"
      }
    ],
    "control": {
      "description": "The key comparison is the other half of H19: runners checked last time and beaten 2 lengths or less whose comment carries no recovery note. If the mechanism holds, they should score clearly below the recovered group (near par, about 1.0 to 1.1), while the recovered group sits at 1.30 or more. The same-race other runners should sit near 1.0, which shows the lift is not a property of the races. If the unrecovered half scores as well as or better than the recovered half, the recovery note carries no information and the H19 lift belongs to the check flag alone.",
      "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND (last_recovered = 0 OR last_recovered IS NULL)"
    },
    "pass_criteria": {
      "min_bets": 150,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "Comment-writing style differs by country (GB vs IRE) and by race reader, so last_recovered may reflect who wrote the comment more than how the horse ran",
      "Overlap with late-headway closers: 'recovered' comments often come with 'made headway'. Closers without trouble were priced correctly (0.969 on 5,540), but the overlap still needs checking",
      "Small sample: the recovered subset of H19's 242 bets is likely under 150, so the discovery floor may not be met even if the effect is real",
      "Price band: in other families a single short-price band has carried apparent lifts, so check whether the lift is spread across prices",
      "Period stability: H19 rose from 1.07 to 1.26 across periods, so a subset could ride one period",
      "SP vs exchange: H19's edge was mostly at BSP, so an SP-measured A/E may understate it or behave differently"
    ],
    "mechanism": "If H19's edge is ability lost to the check, then a horse that was checked, recovered and still finished within 2 lengths showed the most spare ability. A horse that was checked and never picked up again may simply have been beaten on merit. So the recovered subset should carry most of the H19 lift, and the unrecovered subset should sit closer to par.",
    "market_miss": "The form figure and the beaten distance are identical for recovered and unrecovered losers. The recovery shows only in the in-running comment, which prices barely use. The price-movement split found no allowance for the trouble either way (the half whose price did not shorten scored 1.236), so the market seems blind to the finer detail of the trouble as well.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.30+ on roughly 80+ bets, clearly above H19 runners without a recovery note",
      "falsified_if": "Falsified if A/E is below 1.10, or at or below the A/E of the control group (H19 runners with last_recovered = 0 or NULL)."
    },
    "notes": "The draft rule is sound: every column is pre-off and it adds only the one condition the mechanism needs, so it is kept as written. The single refinement targets the price-band confounder. A thin positive result still goes on to falsification and forward tracking. The verdict should depend on the gap between the recovered and unrecovered halves of H19, not on the recovered half's A/E alone."
  }

Timeline

1. Job 1868 finished

   16 Oct, 14:55 · engine:compute-1

2. hypothesis note H-d0a5f0ee

   16 Oct, 14:55 · engine:perturb

3. perturbation result for H-d0a5f0ee (neighbour: drop: last_position > 1): pass

   16 Oct, 14:55 · engine:perturb

4. perturbation result for H-d0a5f0ee (neighbour: drop: last_momentum_interrupted = 1): fail

   16 Oct, 14:55 · engine:perturb

5. perturbation result for H-d0a5f0ee (neighbour: <= 2 -> 3 (+20%)): pass

   16 Oct, 14:55 · engine:perturb

6. perturbation result for H-d0a5f0ee (neighbour: <= 2 -> 1 (-20%)): pass

   16 Oct, 14:55 · engine:perturb

7. perturbation result for H-d0a5f0ee (neighbour: > 1 -> 2 (+20%)): pass

   16 Oct, 14:55 · engine:perturb

8. perturbation result for H-d0a5f0ee (neighbour: > 1 -> 0 (-20%)): pass

   16 Oct, 14:55 · engine:perturb

9. compute-1 took job 1868 (mutate, attempt 1)

   16 Oct, 14:55 · compute-1

10. Job 1882 finished

    16 Oct, 14:55 · engine:compute-1

11. H-d0a5f0ee Falsification → KILLED falsified: 1 of 1 attacks broke it; fewer than two periods with enough bets to compare (thin: the forward record will judge); GB: too few bets to judge; IRE: too few bets to judge

    16 Oct, 14:55 · engine:falsification

12. falsification result for H-d0a5f0ee (price_band: Not carried by short prices: SP above 4.0): info

    16 Oct, 14:55 · engine:evaluator

13. compute-1 took job 1882 (attack, attempt 1)

    16 Oct, 14:55 · compute-1

14. Job 1881 finished

    16 Oct, 14:55 · engine:compute-1

15. falsification result for H-d0a5f0ee (partition: Excluding horses that also made late headway): info

    16 Oct, 14:55 · engine:evaluator