Going-well departure with weak visible form: the hidden form is largest where the form book looks worst

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-9f384f7c (Fell or unseated while going well: hidden form at normal prices) passed discovery as a niche: A/E 1.1 on 245 bets (p 0.291), same-race control 0.98. O-cae9b321
   • H-9f384f7c (Fell or unseated while going well: hidden form at normal prices): the country 'GB' carries the effect (A/E 1.26) while the rest score IRE 0.8. O-dd41f483
2. AskedDo unseated runners that were in contention when they departed beat their price at under 16/1, where the plain BD/UR rule does not? answered
3. MechanismIf the horse was travelling well when it departed, that run was its best recent effort, but the result records F or U. When its finished runs were well beaten, the market prices it on those runs, and the gap between true and visible ability is largest.
4. Why the market misses itPunters and compilers read the beaten distances. A well-beaten recent record plus a non-finish looks like a horse out of form, although the run it did not finish showed improvement.
5. Predictedpositive, A/E 1.12 to 1.25 on 100+ bets
6. Falsified ifA/E below 1.05, or no higher than the parent's 1.099
7. Rule sketchlast_fell_going_well = 1 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND best_last3_beaten > 5

Mechanism and rationale

If the horse was travelling well when it departed, that run was its best recent effort, but the result records F or U. When its finished runs were well beaten, the market prices it on those runs, and the gap between true and visible ability is largest.

The rule

last_fell_going_well = 1 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND best_last3_beaten > 5

Reads: last_fell_going_well, race_type, sp_decimal, best_last3_beaten

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

4bfefda6c03cf3b6ce990ab6b846a42501206e96af8443ea88ff7352f73b109f

Frozen

not frozen

The definition can still change.

Confounders

• Thin sample: last_fell_going_well and best_last3_beaten both start in 2025-11, and the parent was only 245 bets, so this subset will probably fall short of 150 bets in discovery and depend on forward tracking
• best_last3_beaten is NULL when there is no finished run in the last three, so those horses drop out silently. Horses whose only recent runs were departures are excluded, which is a different population from the parent
• Class or rating drops after poor runs: well-beaten horses may get easier races or lower marks, so a lift may come from the handicapper or the class drop and not from hidden form in the departure
• GB-only concentration: the parent's lift sat in GB (1.27) against Ireland (0.80). That may reflect differences in how comments flag 'going well', and a split result here should not be read as support for the mechanism
• Comment-coding noise: 'going well' is read from race comments and may be applied unevenly across courses, commentators or regions
• Mid-band dependence: the parent's excess came from about 137 bets in the mid-short price bands, so a few winners can swing the A/E

Discovery result

pass broad rule (niche) · 16 Oct, 13:35

Bets

94

16 winners

A/E

1.35

95%: 0.77 to 2.19

SP return

+2.1%

level stakes

Betfair SP return

+22.2%

94 bets with a BSP

Strike rate 17.0%. p = 0.15. 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 5 survived.

• perturbation: Weak-form cut loosened: best of last three beaten more than 4 lengthspass

  102 bets · A/E 1.23 (0.70 to 1.99) · control 0.97 · 16 Oct, 13:40

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-27591.46

  after 2026-04-27430.83

• perturbation: Weak-form cut tightened: best of last three beaten more than 6 lengthspass

  86 bets · A/E 1.36 (0.75 to 2.29) · control 0.95 · 16 Oct, 13:40

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: <= 15 -> 12 (-20%)pass

  74 bets · A/E 1.39 (0.78 to 2.29) · control 0.93 · 16 Oct, 13:35

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-27451.55

  after 2026-04-2729–

• neighbour: <= 15 -> 18 (+20%)pass

  101 bets · A/E 1.48 (0.88 to 2.34) · control 0.93 · 16 Oct, 13:35

  Breakdown

Lineage

Agent runs

• designer v1ok · 41s · 16 Oct, 13:33

  Output (schema-checked)

  {
    "rule": "last_fell_going_well = 1 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND best_last3_beaten > 5",
    "refinements": [
      {
        "label": "dose-response: visible form clearly poor (best of last three beaten more than 10 lengths)",
        "rule": "last_fell_going_well = 1 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND best_last3_beaten > 10"
      }
    ],
    "control": {
      "description": "Two comparisons. (1) The other runners in the same races should come out near 1.0 (the parent's same-race control was 0.983). If they also score well above 1.0, the lift belongs to the races and not to these horses. (2) The contrast group is going-well departures in the same price band and code whose visible form is already good (best of last three beaten 5 lengths or less). The mechanism says the market is fooled only when the beaten distances look bad, so this group should score at or below the parent's 1.099, and the main rule should beat it clearly. If the contrast group scores as high as or higher than the main rule, the weak-form sharpening adds nothing and the hypothesis fails.",
      "rule": "last_fell_going_well = 1 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND best_last3_beaten <= 5"
    },
    "pass_criteria": {
      "min_bets": 150,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1.05,
      "max_p": 0.2
    },
    "confounders": [
      "Thin sample: last_fell_going_well and best_last3_beaten both start in 2025-11, and the parent was only 245 bets, so this subset will probably fall short of 150 bets in discovery and depend on forward tracking",
      "best_last3_beaten is NULL when there is no finished run in the last three, so those horses drop out silently. Horses whose only recent runs were departures are excluded, which is a different population from the parent",
      "Class or rating drops after poor runs: well-beaten horses may get easier races or lower marks, so a lift may come from the handicapper or the class drop and not from hidden form in the departure",
      "GB-only concentration: the parent's lift sat in GB (1.27) against Ireland (0.80). That may reflect differences in how comments flag 'going well', and a split result here should not be read as support for the mechanism",
      "Comment-coding noise: 'going well' is read from race comments and may be applied unevenly across courses, commentators or regions",
      "Mid-band dependence: the parent's excess came from about 137 bets in the mid-short price bands, so a few winners can swing the A/E"
    ],
    "mechanism": "A horse that was travelling well when it fell or unseated ran its best recent race in that run, but the form book records only F or U. When its finished runs in the last three were all beaten more than 5 lengths, the visible record holds nothing but poor efforts and a non-finish. The departure then hides the only evidence of improvement, so the gap between true and visible ability should be largest in exactly these horses.",
    "market_miss": "Punters and compilers price from beaten distances and finishing positions. A string of well-beaten runs plus a non-finish reads as a horse out of form. The in-running comment 'going well when fell/unseated' gets little weight because it is not in the summary form figures, and it gets least weight where the form figures are worst. Where the horse has a recent close finish, the market already has visible evidence of ability, so there is little left to misprice.",
    "prediction": {
      "direction": "positive",
      "size": "A/E about 1.12 to 1.25, likely on fewer than 150 discovery bets, and clearly above the good-visible-form contrast group",
      "falsified_if": "A/E below 1.05, or no higher than the parent's 1.099, or no higher than the contrast group (going-well departures with best_last3_beaten <= 5), or the same-race other runners scoring as high as the selections"
    },
    "notes": "I kept the rule sketch unchanged. All its columns are pre_off. best_last3_beaten > 5 needs at least one finished run in the last three (NULL drops out), and every finished run in that window must be beaten more than 5 lengths. I kept the 14/1 cap from the parent because edges die at big prices and the parent's niche was the mid-short band. I did not add a GB-only refinement: the findings say the regional split is a coverage lead, not something to sharpen on. The only refinement is the dose-response test (>10 lengths) that the mechanism predicts. Both key columns start in 2025-11, so expect a thin discovery result to go on to falsification and forward tracking, not a clean pass. I set the forward bar's min A/E to 1.05, above the 1.0 floor, because the parent already sits at about 1.10 and a sharpening that only matches its prices adds nothing."
  }

Timeline

1. Job 1658 finished

   16 Oct, 13:40 · engine:compute-1

2. H-e7675af7 Falsification → KILLED falsified: IRE A/E 0.96; driven by one price band (up to 7/1): the rest score 0.77; 2 of 5 attacks broke it; fewer than two periods with enough bets to compare (thin: the forward record will judge)

   16 Oct, 13:40 · engine:falsification

3. falsification result for H-e7675af7 (partition: Handicaps only: where class and mark drops after poor runs would operate): pass

   16 Oct, 13:40 · engine:evaluator

4. compute-1 took job 1658 (attack, attempt 1)

   16 Oct, 13:40 · compute-1

5. Job 1657 finished

   16 Oct, 13:40 · engine:compute-1

6. falsification result for H-e7675af7 (control: Control without the going-well flag: weak-form fallers and unseaters that were not going well): fail

   16 Oct, 13:40 · engine:evaluator

7. compute-1 took job 1657 (attack, attempt 1)

   16 Oct, 13:40 · compute-1

8. Job 1656 finished

   16 Oct, 13:40 · engine:compute-1

9. falsification result for H-e7675af7 (control: Contrast: going-well departures whose visible form is already good (beaten 5 lengths or less)): fail

   16 Oct, 13:40 · engine:evaluator

10. compute-1 took job 1656 (attack, attempt 1)

    16 Oct, 13:40 · compute-1

11. Job 1655 finished

    16 Oct, 13:40 · engine:compute-1

12. falsification result for H-e7675af7 (perturbation: Weak-form cut tightened: best of last three beaten more than 6 lengths): pass

    16 Oct, 13:40 · engine:evaluator

13. compute-1 took job 1655 (attack, attempt 1)

    16 Oct, 13:40 · compute-1

14. Job 1654 finished

    16 Oct, 13:40 · engine:compute-1

15. falsification result for H-e7675af7 (perturbation: Weak-form cut loosened: best of last three beaten more than 4 lengths): pass

    16 Oct, 13:40 · engine:evaluator

16. compute-1 took job 1654 (attack, attempt 1)

    16 Oct, 13:40 · compute-1

17. Job 1643 finished

    16 Oct, 13:39 · engine:ai-1

18. Queued attack on the compute lane: attack: Handicaps only: where class and mark drops after poor runs would operate

    16 Oct, 13:39 · engine:scheduler