Test-mechanism: recent winners beaten only 8 to 19 lengths should show a smaller rebound

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-363890c3 (Sharpen: recent winners beaten 20+ lengths next time over jumps) survived falsification: 5 of 5 attacks left it standing; it is frozen for its one vault query. O-1a2e3a25
   • H-807e3cb2 (Sharpen: winners beaten 20+ lengths on their first run after winning) passed discovery: A/E 1.25 on 502 bets (p 0.049), same-race control 0.98. O-512837ca
   • H-b0d006d8 (Rival: the recent-winner rebound is a longshot artefact and vanishes a) passed discovery: A/E 1.26 on 447 bets (p 0.03), same-race control 0.97. O-ef2e2d1f
2. AskedIs there a dose-response in the margin: do recent winners beaten 8 to 19 lengths show a smaller lift than those beaten 20+? answered
3. MechanismA 20+ length beating looks to punters like the horse has been found out by its new mark. A moderate beating looks less damning, so the market should shade the price less and there is less mispricing to recover.
4. Why the market misses itThe form figure and the large margin dominate how the horse reads on paper. A smaller margin reads as an ordinary defeat that is already roughly priced.
5. Predictedpositive, A/E 1.00 to 1.12 on 800+ bets, clearly below the 20+ group's 1.34
6. Falsified ifA/E at or above 1.25, which would mean the margin is not the cause and recent winners are underbet after any defeat

Mechanism and rationale

A 20+ length beating looks to punters like the horse has been found out by its new mark. A moderate beating looks less damning, so the market should shade the price less and there is less mispricing to recover.

The rule

is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND best_last3_beaten = 0 AND last_position > 1 AND last_beaten_lengths >= 8 AND last_beaten_lengths < 20

Reads: is_handicap, race_type, best_last3_beaten, last_position, last_beaten_lengths

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

3bf97035e475259d5310bed13352e78dffbaf334da3043f59d27f525ae807773

Frozen

17 Oct, 04:31

Nothing in the definition can change after this.

Confounders

• Recent winners may be underbet after any defeat, whatever the margin. That would lift both bands equally, so it is checked against the 20+ control.
• The mark rises after the win. A moderately beaten horse may still be too high in the weights, which would push A/E down for reasons unrelated to how the market reads the margin.
• IRE weighting: the sister niche's lift was IRE-heavy, so a lift here could just reflect the region mix (see the GB refinement).
• Longshot share: a lift that sits mainly at 16/1 and above would come from a few wins, not from mispricing (see the 14/1 refinement).
• best_last3_beaten only has data from 2025-11-01, so the window is short and spans few seasons, and the result may reflect one jumps season.
• Race class and field size can differ between the 8-19 and 20+ bands. Bigger beatings come more often in bigger or higher-class fields.

Control

{
  "description": "Other runners in the same races should show about 1.00 or a little below. If they do, any lift is not a property of the races. The key comparison for this question is the sister group: the same recent-winner jumps handicappers beaten 20+ lengths last time. It is measured over the same window and should come out near the earlier 1.34. The mechanism is supported only if this rule's A/E sits clearly below that group, by at least about 0.15. If both groups are similar, the margin is not the cause.",
  "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND best_last3_beaten = 0 AND last_position > 1 AND last_beaten_lengths >= 20"
}

Discovery result

pass broad rule (niche) · 17 Oct, 04:29

Bets

828

121 winners

A/E

1.12

95%: 0.93 to 1.34

SP return

-10.6%

level stakes

Betfair SP return

+6.6%

824 bets with a BSP

Strike rate 14.6%. p = 0.12. Same-race control: the other runners in these races had A/E 0.99, 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. 6 of 6 survived.

• perturbation: Perturb band wider: 6 to 21 lengthspass

  1,094 bets · A/E 1.10 (0.94 to 1.29) · control 0.99 · 17 Oct, 04:31

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-275971.13

  after 2026-04-274971.07

• perturbation: Perturb band narrower: 10 to 16 lengthspass

  487 bets · A/E 1.25 (0.98 to 1.57) · control 0.97 · 17 Oct, 04:31

Sealed holdout (vault)

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

Forward record

• forward vs bar {'min_bets': 100, 'min_ae': 1.0, 'max_p': 0.2}info

  0 bets · · 17 Oct, 04:31

  Breakdown

  betsA/E0.71.01.3

  forward0–

Other experiments

• neighbour: > 1 -> 0 (-20%)pass

  828 bets · A/E 1.12 (0.93 to 1.34) · control 0.99 · 17 Oct, 04:36

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-274541.10

  after 2026-04-273741.14

• neighbour: > 1 -> 2 (+20%)pass

  742 bets · A/E 1.14 (0.93 to 1.37) · control 0.99 · 17 Oct, 04:36

  Breakdown

Lineage

Agent runs

• designer v1ok · 32s · 17 Oct, 04:27

  Output (schema-checked)

  {
    "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND best_last3_beaten = 0 AND last_position > 1 AND last_beaten_lengths >= 8 AND last_beaten_lengths < 20",
    "refinements": [
      {
        "label": "14/1 or shorter (the earlier findings show the rebound holds below 16/1, while longshots inflate the broad A/E)",
        "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND best_last3_beaten = 0 AND last_position > 1 AND last_beaten_lengths >= 8 AND last_beaten_lengths < 20 AND sp_decimal <= 15"
      },
      {
        "label": "GB only (checks the IRE-weighting confounder that drove the sister niche)",
        "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND best_last3_beaten = 0 AND last_position > 1 AND last_beaten_lengths >= 8 AND last_beaten_lengths < 20 AND region = 'GB'"
      },
      {
        "label": "Dose check: upper half of the band, 14 to 19 lengths (if margin drives the rebound, this half should beat the 8-13 half)",
        "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND best_last3_beaten = 0 AND last_position > 1 AND last_beaten_lengths >= 14 AND last_beaten_lengths < 20"
      }
    ],
    "control": {
      "description": "Other runners in the same races should show about 1.00 or a little below. If they do, any lift is not a property of the races. The key comparison for this question is the sister group: the same recent-winner jumps handicappers beaten 20+ lengths last time. It is measured over the same window and should come out near the earlier 1.34. The mechanism is supported only if this rule's A/E sits clearly below that group, by at least about 0.15. If both groups are similar, the margin is not the cause.",
      "rule": "is_handicap = 1 AND race_type IN ('Hurdle', 'Chase') AND best_last3_beaten = 0 AND last_position > 1 AND last_beaten_lengths >= 20"
    },
    "pass_criteria": {
      "min_bets": 300,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "Recent winners may be underbet after any defeat, whatever the margin. That would lift both bands equally, so it is checked against the 20+ control.",
      "The mark rises after the win. A moderately beaten horse may still be too high in the weights, which would push A/E down for reasons unrelated to how the market reads the margin.",
      "IRE weighting: the sister niche's lift was IRE-heavy, so a lift here could just reflect the region mix (see the GB refinement).",
      "Longshot share: a lift that sits mainly at 16/1 and above would come from a few wins, not from mispricing (see the 14/1 refinement).",
      "best_last3_beaten only has data from 2025-11-01, so the window is short and spans few seasons, and the result may reflect one jumps season.",
      "Race class and field size can differ between the 8-19 and 20+ bands. Bigger beatings come more often in bigger or higher-class fields."
    ],
    "mechanism": "Punters read a beating of 20 lengths or more after a recent win as proof that the horse has been found out by its new mark, so they push its price out too far. A beating of 8 to 19 lengths reads as an ordinary defeat and gets only a small markdown. If the size of the margin is what drives the mispricing, these moderately beaten recent winners should show a smaller rebound than the heavily beaten ones: a small positive A/E or none, clearly below 1.34.",
    "market_miss": "The form figure and a large beaten margin dominate how a horse reads on paper. A large margin looks damning and gets overreacted to. A moderate margin looks unremarkable and is already roughly priced, so there is less error to recover.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.00 to 1.12 on 800+ bets, clearly below the 20+ control (about 1.34 expected) by at least 0.15",
      "falsified_if": "A/E at or above 1.25, or within 0.10 of the 20+ control group measured over the same window. Either result means the margin is not the cause and recent winners are underbet after any defeat."
    },
    "notes": "This is a test-mechanism spec, and the mechanism predicts a weak result. If the rule fails the 1.05 discovery bar, that supports the hypothesis that margin drives the rebound, provided the 20+ control holds up over the same window. A rule passing at 1.25 or more falsifies the mechanism, even though it would look like a new edge. I added last_position > 1 so that last time's winners and non-finishers are clearly excluded. best_last3_beaten = 0 means a win in the two runs before the last one. min_bets is raised to 300 because the prediction expects 800+ bets."
  }

Timeline

1. Job 2963 finished

   17 Oct, 04:36 · engine:compute-1

2. hypothesis note H-32b815ff

   17 Oct, 04:36 · engine:perturb

3. perturbation result for H-32b815ff (neighbour: < 20 -> 24 (+20%)): pass

   17 Oct, 04:36 · engine:perturb

4. perturbation result for H-32b815ff (neighbour: < 20 -> 16 (-20%)): pass

   17 Oct, 04:36 · engine:perturb

5. perturbation result for H-32b815ff (neighbour: >= 8 -> 10 (+20%)): pass

   17 Oct, 04:36 · engine:perturb

6. perturbation result for H-32b815ff (neighbour: >= 8 -> 6 (-20%)): pass

   17 Oct, 04:36 · engine:perturb

7. perturbation result for H-32b815ff (neighbour: > 1 -> 2 (+20%)): pass

   17 Oct, 04:36 · engine:perturb

8. perturbation result for H-32b815ff (neighbour: > 1 -> 0 (-20%)): pass

   17 Oct, 04:36 · engine:perturb

9. compute-1 took job 2963 (mutate, attempt 1)

   17 Oct, 04:36 · compute-1

10. Queued mutate on the compute lane: parameter stability: do neighbouring thresholds hold?

    17 Oct, 04:31 · engine:scheduler

11. Job 2957 finished

    17 Oct, 04:31 · engine:compute-1

12. forward result for H-32b815ff (forward vs bar {'min_bets': 100, 'min_ae': 1.0, 'max_p': 0.2}): info

    17 Oct, 04:31 · engine:evaluator

13. compute-1 took job 2957 (forward_check, attempt 1)

    17 Oct, 04:31 · compute-1

14. Queued forward_check on the compute lane: daily forward check of a shadow system

    17 Oct, 04:31 · engine:scheduler

15. Job 2953 finished

    17 Oct, 04:31 · engine:compute-1

16. H-32b815ff Frozen → SHADOW vault skipped: best_last3_beaten, last_beaten_lengths do not exist before the discovery zone; track it forward instead

    17 Oct, 04:31 · engine:orchestrator