Test-mechanism: handicap chase debutants whose mark dropped after a heavy hurdle beating

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-cc72aea8 (Rival: the debutant lift is handicap marks after heavy hurdle beatings) passed discovery as a niche: A/E 1.25 on 225 bets (p 0.155), same-race control 0.98. O-40119b5d
2. AskedAmong handicap chase debutants beaten heavily over hurdles, does the lift sit with horses whose official rating fell since the last run? answered
3. MechanismIf the lift comes from a lenient mark, it needs the mark to have actually fallen since the heavy hurdle beating. Debutants whose rating fell carry a mark set on form in a discipline that did not suit them. Those whose mark held do not get that benefit.
4. Why the market misses itPunters read the distance beaten in the form. The mark change and its meaning for a horse switching to fences are not obvious on the racecard.
5. Predictedpositive, A/E 1.25 to 1.50 on 100+ bets, clearly above debutants whose mark did not fall
6. Falsified ifA/E below 1.10, or no higher than the debutants whose official_rating >= last_official_rating
7. Rule sketchrace_type = 'Chase' AND COALESCE(chase_runs, 0) = 0 AND last_race_type = 'Hurdle' AND last_beaten_lengths >= 20 AND is_handicap = 1 AND official_rating IS NOT NULL AND last_official_rating IS NOT NULL AND official_rating < last_official_rating
8. question idQ-719ea36f

The question's answer so far: No. The mark-drop debutant cut was killed at IRE A/E 0.425 (GB 1.50, a single-country effect). The broader mark test found cut and uncut marks level (1.069 vs 1.075), so the rating change is not the mechanism.

Mechanism and rationale

If the lift comes from a lenient mark, it needs the mark to have actually fallen since the heavy hurdle beating. Debutants whose rating fell carry a mark set on form in a discipline that did not suit them. Those whose mark held do not get that benefit.

The rule

race_type = 'Chase' AND COALESCE(chase_runs, 0) = 0 AND last_race_type = 'Hurdle' AND last_beaten_lengths >= 20 AND is_handicap = 1 AND official_rating IS NOT NULL AND last_official_rating IS NOT NULL AND official_rating < last_official_rating

Reads: race_type, chase_runs, last_race_type, last_beaten_lengths, is_handicap, official_rating, last_official_rating

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

ab7bbd0592cbe1faab87c992de6463d093cb4c156e7cb4d7628d277a7fed7a29

Frozen

not frozen

The definition can still change.

Confounders

• Short window: last_official_rating only exists from November 2025 (official_rating from October 2025), so the sample is small and the result could be noise
• Country: mark drops are bigger and handicap debutants are mostly GB, and earlier cuts were lifted in GB only, so a dropped-mark lift could just be the GB effect again (refinement 3 checks this)
• Class drop: horses that are dropped in the weights often also drop in class, so the lift may come from easier races and not the mark. race_class is only available for IRE from July 2026, so this cannot be checked cleanly here
• Long shots: a few winners at 16/1 or bigger can inflate A/E (refinement 1 checks this)
• Time decay: the handicap debutant lift fell from 1.45 to 1.02 after late April 2026, so the result may depend on when the sample falls

Control

{
  "description": "The test is set by the mechanism itself: handicap chase debutants after a hurdle beating of 20 lengths or more whose mark did NOT fall (official_rating >= last_official_rating). If a lenient mark is the cause, the dropped group should clearly beat this control (by about 0.15 A/E or more) and the control should sit near 1.0. If both groups score about the same, the lift is a chase-debut or heavy-beating effect and not a mark effect. The other runners in the same races should come out near 1.0 (0.95 to 1.05). A same-race lift would point to a race-level property such as field or class, not to the selected horses.",
  "rule": "race_type = 'Chase' AND COALESCE(chase_runs, 0) = 0 AND last_race_type = 'Hurdle' AND last_beaten_lengths >= 20 AND is_handicap = 1 AND official_rating IS NOT NULL AND last_official_rating IS NOT NULL AND official_rating >= last_official_rating"
}

Discovery result

pass broad rule (niche) · 17 Oct, 03:10

Bets

164

20 winners

A/E

1.33

95%: 0.81 to 2.06

SP return

+14.4%

level stakes

Betfair SP return

+41.6%

162 bets with a BSP

Strike rate 12.2%. p = 0.13. Same-race control: the other runners in these races had A/E 0.97, 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. 4 of 5 survived.

• perturbation: Beaten-lengths cut lowered to 15pass

  218 bets · A/E 1.15 (0.74 to 1.69) · control 0.99 · 17 Oct, 03:18

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271171.21

  after 2026-04-271011.08

• perturbation: Beaten-lengths cut raised to 25pass

  129 bets · A/E 1.30 (0.71 to 2.19) · control 0.97 · 17 Oct, 03:18

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: >= 20 -> 16 (-20%)pass

  203 bets · A/E 1.14 (0.71 to 1.72) · control 0.99 · 17 Oct, 03:10

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271081.23

  after 2026-04-27951.04

• neighbour: >= 20 -> 24 (+20%)pass

  134 bets · A/E 1.22 (0.67 to 2.05) · control 0.98 · 17 Oct, 03:10

  Breakdown

Lineage

Agent runs

• designer v1ok · 40s · 17 Oct, 03:07

  Output (schema-checked)

  {
    "rule": "race_type = 'Chase' AND COALESCE(chase_runs, 0) = 0 AND last_race_type = 'Hurdle' AND last_beaten_lengths >= 20 AND is_handicap = 1 AND official_rating IS NOT NULL AND last_official_rating IS NOT NULL AND official_rating < last_official_rating",
    "refinements": [
      {
        "label": "Under 16/1 (long-shot confounder)",
        "rule": "race_type = 'Chase' AND COALESCE(chase_runs, 0) = 0 AND last_race_type = 'Hurdle' AND last_beaten_lengths >= 20 AND is_handicap = 1 AND official_rating IS NOT NULL AND last_official_rating IS NOT NULL AND official_rating < last_official_rating AND sp_decimal < 17"
      },
      {
        "label": "Dose: mark dropped 5 lb or more",
        "rule": "race_type = 'Chase' AND COALESCE(chase_runs, 0) = 0 AND last_race_type = 'Hurdle' AND last_beaten_lengths >= 20 AND is_handicap = 1 AND official_rating IS NOT NULL AND last_official_rating IS NOT NULL AND official_rating <= last_official_rating - 5"
      },
      {
        "label": "GB only (country confound check)",
        "rule": "race_type = 'Chase' AND COALESCE(chase_runs, 0) = 0 AND last_race_type = 'Hurdle' AND last_beaten_lengths >= 20 AND is_handicap = 1 AND official_rating IS NOT NULL AND last_official_rating IS NOT NULL AND official_rating < last_official_rating AND region = 'GB'"
      }
    ],
    "control": {
      "description": "The test is set by the mechanism itself: handicap chase debutants after a hurdle beating of 20 lengths or more whose mark did NOT fall (official_rating >= last_official_rating). If a lenient mark is the cause, the dropped group should clearly beat this control (by about 0.15 A/E or more) and the control should sit near 1.0. If both groups score about the same, the lift is a chase-debut or heavy-beating effect and not a mark effect. The other runners in the same races should come out near 1.0 (0.95 to 1.05). A same-race lift would point to a race-level property such as field or class, not to the selected horses.",
      "rule": "race_type = 'Chase' AND COALESCE(chase_runs, 0) = 0 AND last_race_type = 'Hurdle' AND last_beaten_lengths >= 20 AND is_handicap = 1 AND official_rating IS NOT NULL AND last_official_rating IS NOT NULL AND official_rating >= last_official_rating"
    },
    "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": [
      "Short window: last_official_rating only exists from November 2025 (official_rating from October 2025), so the sample is small and the result could be noise",
      "Country: mark drops are bigger and handicap debutants are mostly GB, and earlier cuts were lifted in GB only, so a dropped-mark lift could just be the GB effect again (refinement 3 checks this)",
      "Class drop: horses that are dropped in the weights often also drop in class, so the lift may come from easier races and not the mark. race_class is only available for IRE from July 2026, so this cannot be checked cleanly here",
      "Long shots: a few winners at 16/1 or bigger can inflate A/E (refinement 1 checks this)",
      "Time decay: the handicap debutant lift fell from 1.45 to 1.02 after late April 2026, so the result may depend on when the sample falls"
    ],
    "mechanism": "A heavy hurdle beating (20 lengths or more) leads the handicapper to lower the mark. That lower mark rests on form in a discipline that did not suit the horse. On its chase debut the horse can run to its real level off a lenient mark. The mechanism needs the mark to have actually fallen: debutants whose mark held or rose get no weight benefit and should not show the lift.",
    "market_miss": "Punters price the debutant off the distance it was beaten in the form. The fact that its mark has fallen, and what that means for a horse switching to fences, is not obvious on the racecard, so the price does not allow for the weight edge.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.25 to 1.50 on roughly 100 or more bets, clearly above the debutants whose mark did not fall",
      "falsified_if": "A/E below 1.10, or A/E no higher than the control group of debutants whose official_rating >= last_official_rating, or a lift that disappears under 16/1"
    },
    "notes": "I checked the columns: official_rating and last_official_rating are pre_off, and the rule can only use data from November 2025. The draft rule was sound and is kept as it is. Because the window is short, the sample may fall below 150 bets. Under lab policy a thin result in the right direction goes on to falsification and forward tracking. The control rule is the one the falsification condition names, so it must be measured alongside the main rule."
  }

Timeline

1. Job 2619 finished

   17 Oct, 03:18 · engine:compute-1

2. H-cc8aceed Falsification → KILLED falsified: IRE A/E 0.425

   17 Oct, 03:18 · engine:falsification

3. falsification result for H-cc8aceed (partition: GB only, under 16/1): pass

   17 Oct, 03:18 · engine:evaluator

4. compute-1 took job 2619 (attack, attempt 1)

   17 Oct, 03:18 · compute-1

5. Job 2618 finished

   17 Oct, 03:18 · engine:compute-1

6. falsification result for H-cc8aceed (control: Control: chase-experienced horses after a heavy hurdle beating with a dropped mark): fail

   17 Oct, 03:18 · engine:evaluator

7. compute-1 took job 2618 (attack, attempt 1)

   17 Oct, 03:18 · compute-1

8. Job 2617 finished

   17 Oct, 03:18 · engine:compute-1

9. falsification result for H-cc8aceed (control: Control: same dropped mark after a heavy hurdle beating, but staying over hurdles): pass

   17 Oct, 03:18 · engine:evaluator

10. compute-1 took job 2617 (attack, attempt 1)

    17 Oct, 03:18 · compute-1

11. Job 2616 finished

    17 Oct, 03:18 · engine:compute-1

12. falsification result for H-cc8aceed (perturbation: Beaten-lengths cut raised to 25): pass

    17 Oct, 03:18 · engine:evaluator

13. compute-1 took job 2616 (attack, attempt 1)

    17 Oct, 03:18 · compute-1

14. Job 2615 finished

    17 Oct, 03:18 · engine:compute-1

15. falsification result for H-cc8aceed (perturbation: Beaten-lengths cut lowered to 15): pass

    17 Oct, 03:18 · engine:evaluator

16. compute-1 took job 2615 (attack, attempt 1)

    17 Oct, 03:18 · compute-1

17. Job 2598 finished

    17 Oct, 03:15 · engine:ai-1

18. Queued attack on the compute lane: attack: GB only, under 16/1

    17 Oct, 03:15 · engine:scheduler

19. Queued attack on the compute lane: attack: Control: chase-experienced horses after a heavy hurdle beating with a dropped mark

    17 Oct, 03:15 · engine:scheduler

20. Queued attack on the compute lane: attack: Control: same dropped mark after a heavy hurdle beating, but staying over hurdles

    17 Oct, 03:15 · engine:scheduler