Generalise: hurdle winners lightly raced on the Flat beat their prices in Irish staying Flat handicaps all year

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-7f585029 (Rival: the festival lift is a large-field effect, not a course effect) was predicted positive but came out the other way: A/E 0.79 on 107 bets (p 0.396 for the opposite direction). O-c18c7bad
   • H-cf4b198a (Sharpen: proven hurdlers lightly raced on the Flat in Irish summer sta) passed discovery as a niche: A/E 1.26 on 50 bets (p 0.341), same-race control 0.94. O-d3f516ff
2. AskedIs the lift strongest where the Flat mark has the least information, i.e. horses lightly raced on the Flat but proven over hurdles? answered
3. MechanismA Flat mark set on a few Flat runs knows nothing of a horse's proven hurdle ability. The gap exists whenever such a horse runs in a staying Flat handicap, not only at the summer festivals.
4. Why the market misses itThe Flat form book shows a few ordinary Flat runs. The hurdle win sits under a different code in the form and is easy to discount on Flat terms.
5. Predictedpositive, A/E 1.08 to 1.15 on 120+ bets, with the lift not confined to 16/1 or bigger
6. Falsified ifA/E at or below 1.00 on 100+ bets, or a lift only at 16/1 or bigger
7. Rule sketch

Mechanism and rationale

A Flat mark set on a few Flat runs knows nothing of a horse's proven hurdle ability. The gap exists whenever such a horse runs in a staying Flat handicap, not only at the summer festivals.

The rule

region = 'IRE' AND race_type = 'Flat' AND is_handicap = 1 AND distance_yards >= 2640 AND hurdle_wins >= 1 AND flat_runs <= 6

Reads: region, race_type, is_handicap, distance_yards, hurdle_wins, flat_runs

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

1fb157aaeeadaaf3ca77768db75187491d191b2a2b559d339a1bf5d8d9c469f0

Frozen

not frozen

The definition can still change.

Confounders

• Summer overlap: the all-year rule contains the June-September slice that already scored 1.26 on 50, and that slice was the reason this was proposed. A broad pass driven by summer is not generalisation, so the outside-summer refinement has to point the same way.
• Trainer concentration: a few big dual-purpose yards (e.g. Mullins, Elliott, de Bromhead-type operations) may supply most bets. A yard's placing skill would then pass as an information gap. Check the share of the top 3 trainers.
• Price band: the lift could be short-priced plot horses or a few big BSP winners. The hypothesis is falsified if the lift exists only at 16/1 or bigger.
• Leopardstown/Curragh autumn and Galway/Listowel festival staying handicaps aimed at by dual-purpose yards. Check that no single course carries the result.
• Data-window artefact: flat_runs and hurdle_wins count only from July 2020, so older horses with long pre-2020 Flat records can look lightly raced. Check the age profile.
• Rival explanation: lightly raced Flat stayers improve with racing whatever their hurdle record (see the control rule).

Discovery result

pass broad rule (niche) · 17 Oct, 03:29

Bets

99

12 winners

A/E

1.26

95%: 0.65 to 2.19

SP return

-28.5%

level stakes

Betfair SP return

-14.3%

94 bets with a BSP

Strike rate 12.1%. p = 0.25. Same-race control: the other runners in these races had A/E 0.95, 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. 3 of 6 survived.

• period: Outside June-September onlyinfo

  49 bets · A/E 1.25 (0.46 to 2.72) · control 0.96 · 17 Oct, 03:30

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-2729–

  after 2026-04-2720–

• perturbation: Loosened thresholds: 13f+ and up to 8 Flat runspass

  140 bets · A/E 1.36 (0.81 to 2.16) · control 0.93 · 17 Oct, 03:30

  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.

Lineage

Agent runs

• designer v1ok · 34s · 17 Oct, 03:23

  Output (schema-checked)

  {
    "rule": "region = 'IRE' AND race_type = 'Flat' AND is_handicap = 1 AND distance_yards >= 2640 AND hurdle_wins >= 1 AND flat_runs <= 6",
    "refinements": [
      {
        "label": "Outside summer only (the actual question: does the gap exist away from June-September?)",
        "rule": "region = 'IRE' AND race_type = 'Flat' AND is_handicap = 1 AND distance_yards >= 2640 AND hurdle_wins >= 1 AND flat_runs <= 6 AND month NOT BETWEEN 6 AND 9"
      },
      {
        "label": "14/1 or shorter (the summer sharpen's winners came at short prices; an information gap should show where the market is engaged)",
        "rule": "region = 'IRE' AND race_type = 'Flat' AND is_handicap = 1 AND distance_yards >= 2640 AND hurdle_wins >= 1 AND flat_runs <= 6 AND sp_decimal <= 15"
      }
    ],
    "control": {
      "description": "Same-race control: the other runners in the same Irish staying Flat handicaps. Expect them near 0.95-1.00 (the summer sharpen's race-mates scored 0.94), so a selection A/E of 1.08+ against a control near 1.0 points to the horse profile, not the race. Comparison group: lightly raced Flat staying handicappers with no hurdle win (the rival H-6234a360). If they score as well as the selection, the lift comes from being lightly raced on the Flat, not from hidden hurdle ability.",
      "rule": "region = 'IRE' AND race_type = 'Flat' AND is_handicap = 1 AND distance_yards >= 2640 AND hurdle_wins = 0 AND flat_runs <= 6"
    },
    "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": [
      "Summer overlap: the all-year rule contains the June-September slice that already scored 1.26 on 50, and that slice was the reason this was proposed. A broad pass driven by summer is not generalisation, so the outside-summer refinement has to point the same way.",
      "Trainer concentration: a few big dual-purpose yards (e.g. Mullins, Elliott, de Bromhead-type operations) may supply most bets. A yard's placing skill would then pass as an information gap. Check the share of the top 3 trainers.",
      "Price band: the lift could be short-priced plot horses or a few big BSP winners. The hypothesis is falsified if the lift exists only at 16/1 or bigger.",
      "Leopardstown/Curragh autumn and Galway/Listowel festival staying handicaps aimed at by dual-purpose yards. Check that no single course carries the result.",
      "Data-window artefact: flat_runs and hurdle_wins count only from July 2020, so older horses with long pre-2020 Flat records can look lightly raced. Check the age profile.",
      "Rival explanation: lightly raced Flat stayers improve with racing whatever their hurdle record (see the control rule)."
    ],
    "mechanism": "A Flat handicap mark set on six or fewer Flat runs is calibrated only on Flat evidence and says nothing of stamina and ability proven by winning over hurdles. Whenever such a horse runs in an Irish staying (14f+) Flat handicap, its true ability is likely above what its Flat mark implies. Nothing in this mechanism depends on the season, so the lift should hold all year, not only at the summer festivals.",
    "market_miss": "Flat punters and Flat-focused pricing read the Flat form lines: a few ordinary runs, often over the wrong trip or on the way back from a break. The hurdle win sits under a different code in the form and is easy to discount on Flat terms, so the horse is priced on its unexposed-but-moderate Flat profile and not on its proven staying ability.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.08 to 1.15 on 120+ bets all year, with the outside-June-September subset also above 1.0 and the lift not confined to 16/1 or bigger",
      "falsified_if": "A/E at or below 1.00 on 100+ bets, or a lift that exists only at 16/1 or bigger (14/1-or-shorter subset at or below 1.00), or an all-year lift that disappears outside June-September (outside-summer A/E at or below 1.00)"
    },
    "notes": "I kept the rule sketch: all its columns are pre-off and it adds no slicing beyond the mechanism. Refinement 1 answers question Q-1f69cedd directly, because the broad rule includes the summer slice that motivated it. The control rule doubles as a check on rival H-6234a360. sp_decimal <= 15 means 14/1 or shorter."
  }

Timeline

1. Job 2686 finished

   17 Oct, 03:30 · engine:compute-1

2. H-7460c63b Falsification → KILLED falsified: 2 of 5 attacks broke it; fewer than two periods with enough bets to compare (thin: the forward record will judge); GB: too few bets to judge

   17 Oct, 03:30 · engine:falsification

3. falsification result for H-7460c63b (partition: Young horses only (age 6 or under) to remove the data-window artefact): pass

   17 Oct, 03:30 · engine:evaluator

4. compute-1 took job 2686 (attack, attempt 1)

   17 Oct, 03:30 · compute-1

5. Job 2685 finished

   17 Oct, 03:30 · engine:compute-1

6. falsification result for H-7460c63b (partition: Excluding the big festival and dual-purpose-target courses): pass

   17 Oct, 03:30 · engine:evaluator

7. compute-1 took job 2685 (attack, attempt 1)

   17 Oct, 03:30 · compute-1

8. Job 2684 finished

   17 Oct, 03:30 · engine:compute-1

9. falsification result for H-7460c63b (control: Control: hurdle winners already exposed on the Flat): fail

   17 Oct, 03:30 · engine:evaluator

10. compute-1 took job 2684 (attack, attempt 1)

    17 Oct, 03:30 · compute-1

11. Job 2683 finished

    17 Oct, 03:30 · engine:compute-1

12. falsification result for H-7460c63b (control: Control: hurdle runs but no hurdle win): fail

    17 Oct, 03:30 · engine:evaluator

13. compute-1 took job 2683 (attack, attempt 1)

    17 Oct, 03:30 · compute-1

14. Job 2682 finished

    17 Oct, 03:30 · engine:compute-1

15. falsification result for H-7460c63b (perturbation: Loosened thresholds: 13f+ and up to 8 Flat runs): pass

    17 Oct, 03:30 · engine:evaluator

16. compute-1 took job 2682 (attack, attempt 1)

    17 Oct, 03:30 · compute-1

17. Job 2681 finished

    17 Oct, 03:30 · engine:compute-1

18. falsification result for H-7460c63b (period: Outside June-September only): info

    17 Oct, 03:30 · engine:evaluator

19. compute-1 took job 2681 (attack, attempt 1)

    17 Oct, 03:30 · compute-1