Sharpen: fell-going-well chasers with no other recent jumping errors

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
   • H85-fell-when-going-well (H85: fell when going well) was GREEN in the earlier research: A/E 1.13 on 405 bets (periods 1.15 / 1.11); now PROMISING. O-add18290
2. AskedIs the edge concentrated in horses whose fall was a one-off, rather than horses with repeated jumping errors? answered
3. MechanismA chaser that otherwise jumps cleanly and fell while travelling well lost a run to bad luck. A habitual faller's falls predict more falls.
4. Why the market misses itA habit shows over several runs and gets priced. A single F next to otherwise clean form is read as failure, and the market does not see the good run behind it.
5. Predictedpositive, A/E 1.25 to 1.40 on 100+ bets
6. Falsified ifA/E at or below the parent system's 1.30, or below 1.10.
7. Rule sketchrace_type = 'Chase' AND last_fell_going_well = 1 AND jumping_errors_last3 <= 1
8. question idQ-3bc191f4
9. operatorsharpen
10. agentinvestigator@v1

The question's answer so far: No. Restricting to horses with at most one jumping error in the last three runs lowered A/E to 1.22 on 139, against 1.30 for the parent on 165, so a one-off fall does not concentrate the edge.

Mechanism and rationale

A chaser that otherwise jumps cleanly and fell while travelling well lost a run to bad luck. A habitual faller's falls predict more falls.

The rule

race_type = 'Chase' AND last_fell_going_well = 1 AND jumping_errors_last3 <= 1

Reads: race_type, last_fell_going_well, jumping_errors_last3

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

ed8228d942824fdb46dbad58fd8877ec47d795ed5bd42b52cd853d74cb5f0fcd

Frozen

not frozen

The definition can still change.

Confounders

• jumping_errors_last3 probably counts the fall itself. If so, <= 1 means 'no error apart from the fall'; if not, it lets through one other error. The meaning of the threshold rests on a column definition we have not confirmed, and that should be checked when reading the result.
• Small sample: this is a subset of the parent's 165 chase bets, and the pre-off columns only start on 2025-11-01, so discovery covers one season. It will very likely miss the 150-bet floor, so it goes to falsification and forward tracking.
• Horses with few chase runs have a 'clean' last-3 count by default, not as evidence that they jump well (the refinement addresses this).
• Overlap with the parent: any lift may simply be inherited from the fell-going-well chase pool. The habitual-error control and the comparison against the parent's ~1.30 are what separate the two.
• Bounce-back of well-beaten chasers: fancied chasers that were well beaten with no error already score about 1.11, so part of any lift is generic.
• Calendar/late-period skew and GB vs IRE differences seen in sibling chase-error cuts; the IRE non-finisher subset was well below par.

Discovery result

pass broad rule (niche) · 17 Oct, 02:10

Bets

139

22 winners

A/E

1.22

95%: 0.76 to 1.84

SP return

+49.7%

level stakes

Betfair SP return

+125.1%

139 bets with a BSP

Strike rate 15.8%. p = 0.21. 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.

• price_band: 14/1 or shorter only: is the edge just a few 16/1+ shock winners?pass

  102 bets · A/E 1.05 (0.62 to 1.67) · control 0.99 · 17 Oct, 02:11

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-27721.26

  after 2026-04-2730–

• control: Habitual-error control: fell going well with 2+ recent jumping errorsinfo

  26 bets · A/E 1.71 (0.63 to 3.73) · control 0.90 · 17 Oct, 02:11

  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 · 46s · 17 Oct, 02:08

  Output (schema-checked)

  {
    "rule": "race_type = 'Chase' AND last_fell_going_well = 1 AND jumping_errors_last3 <= 1",
    "refinements": [
      {
        "label": "established chaser (clean record means something)",
        "rule": "race_type = 'Chase' AND last_fell_going_well = 1 AND jumping_errors_last3 <= 1 AND chase_runs >= 3"
      }
    ],
    "control": {
      "description": "Two comparisons. (1) Same-race control: the other runners in the same chases should be near par (A/E about 0.95 to 1.00). That shows the lift does not come from the kind of race these horses run in. (2) The key contrast for this sharpen is the habitual-error group: chasers that fell going well last time and have 2 or more jumping errors in their last 3 runs. The mechanism predicts this group sits clearly below the clean group, near par or under. If the habitual group shows the same lift or a bigger one, the 'one-off fall' explanation is wrong and the edge belongs to the parent condition (fell going well in a chase), not to clean jumping.",
      "rule": "race_type = 'Chase' AND last_fell_going_well = 1 AND jumping_errors_last3 >= 2"
    },
    "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": [
      "jumping_errors_last3 probably counts the fall itself. If so, <= 1 means 'no error apart from the fall'; if not, it lets through one other error. The meaning of the threshold rests on a column definition we have not confirmed, and that should be checked when reading the result.",
      "Small sample: this is a subset of the parent's 165 chase bets, and the pre-off columns only start on 2025-11-01, so discovery covers one season. It will very likely miss the 150-bet floor, so it goes to falsification and forward tracking.",
      "Horses with few chase runs have a 'clean' last-3 count by default, not as evidence that they jump well (the refinement addresses this).",
      "Overlap with the parent: any lift may simply be inherited from the fell-going-well chase pool. The habitual-error control and the comparison against the parent's ~1.30 are what separate the two.",
      "Bounce-back of well-beaten chasers: fancied chasers that were well beaten with no error already score about 1.11, so part of any lift is generic.",
      "Calendar/late-period skew and GB vs IRE differences seen in sibling chase-error cuts; the IRE non-finisher subset was well below par.",
      "The profit has so far appeared only at BSP; at SP ROI was negative for the parent."
    ],
    "mechanism": "A chaser that fell or unseated while travelling well, with no other jumping errors in its recent runs, lost that run to bad luck: the comment shows it was going well when the race ended for it. Its ability and jumping are intact, so the next run should go back to that standard. A horse with repeated errors is different: its fall reflects a real jumping weakness that predicts more errors, so the 'going well' reading tells you less about its next run.",
    "market_miss": "Punters and odds-compilers read form figures. A lone F sits in the form string as a failure, with no finishing position to show the horse was travelling well. A habitual faller's problem shows over several runs and is already priced in, but a single fall next to otherwise clean form gets the same penalty without the context behind it, so the clean one-off faller goes off bigger than its chance. The bias should be strongest on exchange prices (BSP), where the parent's profit appeared.",
    "prediction": {
      "direction": "positive",
      "size": "A/E about 1.25 to 1.40 on roughly 100+ bets; the habitual-error control (>=2 errors) clearly lower, near or below 1.0",
      "falsified_if": "A/E below 1.10, or not above the parent's ~1.30 (the sharpen adds nothing), or the habitual-error control (jumping_errors_last3 >= 2) does as well as or better than the clean group."
    },
    "notes": "This is a sharpen of the fell-going-well chase system (A/E 1.297 on 165). The columns only exist from 2025-11-01, so discovery covers a single season and the sample is a subset of the parent. Expect it to fall short of 150 bets. A thin positive result goes to falsification and forward tracking rather than being killed. The forward bar is set at A/E 1.05, stricter than the 1.0 floor, because the parent already runs at about 1.30, so a sharpen that only matches par adds nothing. When reading the result, check whether jumping_errors_last3 counts the last-run fall. If it does not, the rule allows one other error, and a stricter reading (= 0) would be the true 'clean' test; that cut has not been pre-registered here. Report SP and BSP ROI separately."
  }

Timeline

1. Job 2437 finished

   17 Oct, 02:11 · engine:compute-1

2. H-6d4d8b5d Falsification → KILLED falsified: IRE A/E 0.781; fewer than two periods with enough bets to compare (thin: the forward record will judge)

   17 Oct, 02:11 · engine:falsification

3. falsification result for H-6d4d8b5d (partition: Handicap chases only): pass

   17 Oct, 02:11 · engine:evaluator

4. compute-1 took job 2437 (attack, attempt 1)

   17 Oct, 02:11 · compute-1

5. Job 2436 finished

   17 Oct, 02:11 · engine:compute-1

6. falsification result for H-6d4d8b5d (partition: Established chasers only (3+ chase runs, so a clean record means something)): pass

   17 Oct, 02:11 · engine:evaluator

7. compute-1 took job 2436 (attack, attempt 1)

   17 Oct, 02:11 · compute-1

8. Job 2435 finished

   17 Oct, 02:11 · engine:compute-1

9. falsification result for H-6d4d8b5d (control: Clean chase fallers NOT going well: is 'going well' the ingredient?): pass

   17 Oct, 02:11 · engine:evaluator

10. compute-1 took job 2435 (attack, attempt 1)

    17 Oct, 02:11 · compute-1

11. Job 2434 finished

    17 Oct, 02:11 · engine:compute-1

12. falsification result for H-6d4d8b5d (control: Habitual-error control: fell going well with 2+ recent jumping errors): info

    17 Oct, 02:11 · engine:evaluator

13. compute-1 took job 2434 (attack, attempt 1)

    17 Oct, 02:11 · compute-1

14. Job 2433 finished

    17 Oct, 02:11 · engine:compute-1

15. falsification result for H-6d4d8b5d (price_band: 14/1 or shorter only: is the edge just a few 16/1+ shock winners?): pass

    17 Oct, 02:11 · engine:evaluator

16. compute-1 took job 2433 (attack, attempt 1)

    17 Oct, 02:11 · compute-1

17. Job 2422 finished

    17 Oct, 02:11 · engine:ai-1

18. Queued attack on the compute lane: attack: Handicap chases only

    17 Oct, 02:11 · engine:scheduler

19. Queued attack on the compute lane: attack: Established chasers only (3+ chase runs, so a clean record means something)

    17 Oct, 02:11 · engine:scheduler