Generalise: one-off chase error, well beaten, not dropping in class (GB), with or without a close run

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-8443fa31 (Rival: close-earlier-run chase edge is class relief, tested on GB runn) survived falsification: 6 of 6 attacks left it standing; it is frozen for its one vault query. O-54398e1f
2. AskedWithin the one-off chase-error pool (GB), is the edge carried by not dropping in class rather than by a close earlier run, so that class droppers are the weak group whatever their earlier form? answered
3. MechanismAfter a run ruined by one jumping error, connections who keep the horse at the same grade or step it up are showing they still trust it: the mistake, not the ability, cost the run. Connections who drop the horse in class are often admitting a decline, and a class drop alone does not restore a chaser that is losing its form. The trainer's placing decision carries information about the horse's well-being that the form figure cannot.
4. Why the market misses itPunters and odds compilers reward class droppers and mark down a horse whose figure reads P/U/F or 10L+ beaten. A horse running at the same or a higher grade has no class-drop flag on the card, so only the bad figure is priced. The dropper gets backed for the drop, which the close-run droppers' 0.997 on 40 suggests is already priced.
5. Predictedpositive, A/E 1.20 to 1.40 on about 240 to 260 GB bets
6. Falsified ifA/E of 1.10 or lower on the no-drop pool, or the class-dropper complement (race_class > last_race_class) scoring as high as the no-drop pool. Either would mean class placing carries no information here.

Mechanism and rationale

After a run ruined by one jumping error, connections who keep the horse at the same grade or step it up are showing they still trust it: the mistake, not the ability, cost the run. Connections who drop the horse in class are often admitting a decline, and a class drop alone does not restore a chaser that is losing its form. The trainer's placing decision carries information about the horse's well-being that the form figure cannot.

The rule

region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class <= last_race_class

Reads: region, race_type, remarked_last3, jumping_errors_last3, last_jumping_error, last_position, last_beaten_lengths, race_class, last_race_class

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

4a0258b1227d0c56811a5adc322ecaea5e0db50a58794de0a5c0588ab2ae5ef3

Frozen

17 Oct, 01:23

Nothing in the definition can change after this.

Confounders

• Heavy overlap with the five frozen one-off chase-error subsets (especially H-8443fa31). This reframes the same runners, so it is not independent evidence, and the five subsets count as one piece of evidence.
• About 27 class risers (A/E about 2.4 in the close-run pool) may carry most of the lift. The strictly-same-class part scored only 1.169 on 93, so the risers-only and same-class refinements must be read side by side.
• Handicap marks fall after a bad run, so a horse at the same class may be running off a lower mark. That is mark relief, not trust shown by the trainer.
• GB only, because IRE race_class starts in 2026-07. The result cannot be generalised to IRE.
• Across the thread, profit has shown at BSP but not at SP, and the no-close-run part returned only +3% at BSP. An A/E edge may not survive the overround.
• race_class compares today only with the error run, not with the earlier runs in the last three. A 'non-drop' can still be a drop from the horse's usual level.
• A class rise can go with a step up to graded or better-quality races, where the market may price differently. That is a race-quality effect separate from the trainer's intent.

Discovery result

pass broad rule · 17 Oct, 01:19

Bets

253

38 winners

A/E

1.36

95%: 0.96 to 1.86

SP return

+20.1%

level stakes

Betfair SP return

+43.5%

253 bets with a BSP

Strike rate 15.0%. p = 0.041. Same-race control: the other runners in these races had A/E 0.96, 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. 5 of 6 survived.

• perturbation: Beaten-lengths cut lowered to 8Lpass

  279 bets · A/E 1.19 (0.85 to 1.63) · control 0.98 · 17 Oct, 01:23

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271541.02

  after 2026-04-271251.40

• perturbation: Beaten-lengths cut raised to 12.5Lpass

  240 bets · A/E 1.30 (0.90 to 1.82) · control 0.97 · 17 Oct, 01:23

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': 1.05, 'max_p': 0.2}info

  0 bets · · 17 Oct, 01:23

  Breakdown

  betsA/E0.71.01.3

  forward0–

Other experiments

• neighbour: > 10 -> 8 (-20%)pass

  279 bets · A/E 1.19 (0.85 to 1.63) · control 0.98 · 17 Oct, 01:19

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271541.02

  after 2026-04-271251.40

• neighbour: > 10 -> 12 (+20%)pass

  241 bets · A/E 1.34 (0.93 to 1.86) · control 0.96 · 17 Oct, 01:19

  Breakdown

Lineage

Agent runs

• designer v1ok · 39s · 17 Oct, 01:19

  Output (schema-checked)

  {
    "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class <= last_race_class",
    "refinements": [
      {
        "label": "class risers only",
        "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class < last_race_class"
      },
      {
        "label": "same class only",
        "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class = last_race_class"
      },
      {
        "label": "under 16/1 (chase-thread price band where the edge was strongest)",
        "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class <= last_race_class AND sp_decimal < 17"
      }
    ],
    "control": {
      "description": "The same-race other runners should score about 1.0, which shows the lift belongs to these horses and not to the races they run in. The key control is the class-dropper complement: the same one-off-error, well-beaten GB chasers who drop in class today (race_class > last_race_class). The mechanism predicts they score about 1.0 (the close-run droppers scored 0.997 on 40). If droppers score as high as the no-drop pool, class placing carries no information.",
      "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class > last_race_class"
    },
    "pass_criteria": {
      "min_bets": 150,
      "min_ae": 1.1,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1.05,
      "max_p": 0.2
    },
    "confounders": [
      "Heavy overlap with the five frozen one-off chase-error subsets (especially H-8443fa31). This reframes the same runners, so it is not independent evidence, and the five subsets count as one piece of evidence.",
      "About 27 class risers (A/E about 2.4 in the close-run pool) may carry most of the lift. The strictly-same-class part scored only 1.169 on 93, so the risers-only and same-class refinements must be read side by side.",
      "Handicap marks fall after a bad run, so a horse at the same class may be running off a lower mark. That is mark relief, not trust shown by the trainer.",
      "GB only, because IRE race_class starts in 2026-07. The result cannot be generalised to IRE.",
      "Across the thread, profit has shown at BSP but not at SP, and the no-close-run part returned only +3% at BSP. An A/E edge may not survive the overround.",
      "race_class compares today only with the error run, not with the earlier runs in the last three. A 'non-drop' can still be a drop from the horse's usual level.",
      "A class rise can go with a step up to graded or better-quality races, where the market may price differently. That is a race-quality effect separate from the trainer's intent."
    ],
    "mechanism": "After a run ruined by one jumping error (one error in three remarked runs, the last of them, and the horse pulled up, fell or was beaten more than 10L), the trainer's placing is a signal about well-being. Keeping the horse at the same grade or stepping it up shows the trainer believes the mistake, not lost ability, cost the run. Dropping it in class often admits a decline that a drop alone does not fix. So the no-drop group should hold more horses whose true ability is intact, and the bad figure understates them.",
    "market_miss": "The card flags class droppers, and punters and odds compilers reward them. They also mark down a P/U/F or 10L+ beaten figure. A horse at the same or a higher grade carries no drop flag, so the bad figure is the only thing priced, and the trainer's implicit vote of confidence is ignored. Droppers are backed for the drop, which the close-run droppers' 0.997 on 40 suggests is already in the price.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.20 to 1.40 on about 240 to 260 GB bets; class-dropper control about 1.0",
      "falsified_if": "No-drop pool A/E of 1.10 or lower, or the class-dropper complement (race_class > last_race_class) scoring at or above the no-drop pool's A/E. Either would mean class placing carries no information here."
    },
    "notes": "The rule sketch is kept unchanged. It uses only pre-off columns and race_class is available for GB from 2020-07. The sketch drops the close-earlier-run condition on purpose, to answer Q-b17a1795: is non-drop, rather than a close run, the condition that carries the edge? The pass A/E is raised to 1.10 to match the falsification threshold. The risers and same-class refinements test whether the lift is just the roughly 27 risers. Read any pass alongside H-8443fa31 because the two overlap heavily."
  }

Timeline

1. Job 2225 finished

   17 Oct, 01:23 · engine:compute-1

2. forward result for H-c2928b4e (forward vs bar {'min_bets': 100, 'min_ae': 1.05, 'max_p': 0.2}): info

   17 Oct, 01:23 · engine:evaluator

3. compute-1 took job 2225 (forward_check, attempt 1)

   17 Oct, 01:23 · compute-1

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

   17 Oct, 01:23 · engine:scheduler

5. Job 2223 finished

   17 Oct, 01:23 · engine:compute-1

6. H-c2928b4e Frozen → SHADOW vault skipped: remarked_last3, jumping_errors_last3, last_jumping_error, last_beaten_lengths do not exist before the discovery zone; track it forward instead

   17 Oct, 01:23 · engine:orchestrator

7. compute-1 took job 2223 (vault, attempt 1)

   17 Oct, 01:23 · compute-1

8. Job 2222 finished

   17 Oct, 01:23 · engine:compute-1

9. Queued vault on the compute lane: frozen: one sealed vault query

   17 Oct, 01:23 · engine:scheduler

10. H-c2928b4e Falsification → FROZEN survived falsification

    17 Oct, 01:23 · engine:falsification

11. falsification result for H-c2928b4e (price_band: 14/1 or shorter only (drop the 3-win 16/1+ tail)): pass

    17 Oct, 01:23 · engine:evaluator

12. compute-1 took job 2222 (attack, attempt 1)

    17 Oct, 01:23 · compute-1

13. Job 2221 finished

    17 Oct, 01:23 · engine:compute-1

14. falsification result for H-c2928b4e (partition: Non-handicaps only (no mark relief possible)): info

    17 Oct, 01:23 · engine:evaluator

15. compute-1 took job 2221 (attack, attempt 1)

    17 Oct, 01:23 · compute-1

16. Job 2220 finished

    17 Oct, 01:23 · engine:compute-1

17. falsification result for H-c2928b4e (partition: Completed runs only (beaten over 10L, no P/U/F)): pass

    17 Oct, 01:23 · engine:evaluator

18. compute-1 took job 2220 (attack, attempt 1)

    17 Oct, 01:23 · compute-1

19. Job 2219 finished

    17 Oct, 01:23 · engine:compute-1

20. falsification result for H-c2928b4e (control: Control: well-beaten GB non-droppers with no jumping error): pass

    17 Oct, 01:23 · engine:evaluator