Sharpen: one-off chase error that ended the last run in a fall or unseat

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-514cfd70 (Test-mechanism: one-off chase error that ended the last run (non-compl) passed discovery as a niche: A/E 1.39 on 97 bets (p 0.099), same-race control 0.95. O-50cc9a22
   • H-514cfd70 (Test-mechanism: one-off chase error that ended the last run (non-compl) survived falsification: 4 of 5 attacks left it standing; it is frozen for its one vault query. O-bd96cc21
2. AskedInside the one-off-error non-finisher pool, is the lift carried by fallers and unseaters (the error itself ended the run) rather than by pull-ups after an error? testing
3. MechanismWhen the error itself ends the run, the horse shows no evidence of lost ability. After two clean runs, its jumping is not a habit, so its earlier level should come back.
4. Why the market misses itF or UR on the form line is read as a jumping risk and lengthens the price. The clean remarks from the two previous runs say it was a one-off.
5. Predictedpositive, A/E 1.35 to 1.70 on 50 to 80 bets
6. Falsified ifA/E at or below 1.15, or pull-ups after an error scoring the same or higher than fallers and unseaters
7. Rule sketchrace_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND last_fell_or_unseated = 1

Mechanism and rationale

When the error itself ends the run, the horse shows no evidence of lost ability. After two clean runs, its jumping is not a habit, so its earlier level should come back.

The rule

race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND last_fell_or_unseated = 1

Reads: race_type, remarked_last3, jumping_errors_last3, last_jumping_error, last_fell_or_unseated

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

1f32be84e1b8f7dd103b4e0297ba9de1b7f4cc089292ba321aca424fb91e347e

Frozen

not frozen

The definition can still change.

Confounders

• Small sample: the parent non-finisher pool is only 97 bets, so this subset will probably fall short of 150 and go on to falsification and forward tracking as a thin result
• The Irish subset (0.43 on 23 in the parent pool) pulling the pooled figure down, or the GB lift being a regional effect rather than a faller/unseat effect
• Jockey change after an unseat: a stronger rider booked next time could be the real cause
• Festival and big-field falls, where traffic causes the fall and does not reflect the horse's jumping
• Price band: the thread-wide pattern is that runners at 16/1 or bigger never beat par, so the price mix could create or hide the effect
• Calendar: the late-period skew after 27 April 2026 seen across this thread
• Unseats caused by the rider rather than the horse's jumping error being mixed in with real falls

Control

{
  "description": "Main comparison: chase runners whose one-off jumping error ended their last run in a pull-up rather than a fall or unseat, with the same two clean earlier runs. If the lift is about the error ending the run before any ability was shown, fallers and unseaters should beat the pull-ups clearly; pull-ups may show lost ability or a hidden problem and should sit near par or below. The other runners in the same races should be near par (about 0.95 to 1.0), so the lift is not a feature of those races.",
  "rule": "race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND last_pulled_up = 1"
}

Discovery result

fail broad rule · 17 Oct, 02:09

Bets

37

3 winners

A/E

0.81

95%: 0.17 to 2.36

SP return

-32.4%

level stakes

Betfair SP return

-22.6%

37 bets with a BSP

Strike rate 8.1%. p = 0.72. Same-race control: the other runners in these races had A/E 1.02, so the selection did not beat its own races. Zone: discovery. Evaluator eval-1.0, data 6a4e98be18e83933

Falsification

No attacks yet. Only promising ideas are attacked.

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 · 37s · 17 Oct, 02:01

  Output (schema-checked)

  {
    "rule": "race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND last_fell_or_unseated = 1",
    "refinements": [
      {
        "label": "GB only (IRE error non-finishers sat at 0.43)",
        "rule": "race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND last_fell_or_unseated = 1 AND region = 'GB'"
      },
      {
        "label": "Under 16/1 (outsiders at 16/1+ never beat par in this thread)",
        "rule": "race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND last_fell_or_unseated = 1 AND sp_decimal < 17"
      }
    ],
    "control": {
      "description": "Main comparison: chase runners whose one-off jumping error ended their last run in a pull-up rather than a fall or unseat, with the same two clean earlier runs. If the lift is about the error ending the run before any ability was shown, fallers and unseaters should beat the pull-ups clearly; pull-ups may show lost ability or a hidden problem and should sit near par or below. The other runners in the same races should be near par (about 0.95 to 1.0), so the lift is not a feature of those races.",
      "rule": "race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND last_pulled_up = 1"
    },
    "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": [
      "Small sample: the parent non-finisher pool is only 97 bets, so this subset will probably fall short of 150 and go on to falsification and forward tracking as a thin result",
      "The Irish subset (0.43 on 23 in the parent pool) pulling the pooled figure down, or the GB lift being a regional effect rather than a faller/unseat effect",
      "Jockey change after an unseat: a stronger rider booked next time could be the real cause",
      "Festival and big-field falls, where traffic causes the fall and does not reflect the horse's jumping",
      "Price band: the thread-wide pattern is that runners at 16/1 or bigger never beat par, so the price mix could create or hide the effect",
      "Calendar: the late-period skew after 27 April 2026 seen across this thread",
      "Unseats caused by the rider rather than the horse's jumping error being mixed in with real falls"
    ],
    "mechanism": "If a single jumping error ends the run in a fall or unseat, the horse has given no evidence that its ability has dropped. The run simply stopped. Because its two earlier runs carried clean remarks, the error looks like a one-off and not a habit, so its earlier level should return. A pull-up after an error is different: it may point to injury, lost confidence or poor form, and so should not share the lift.",
    "market_miss": "An F or UR on the form line is read as a jumping risk. It also leaves no finishing position or beaten distance to rate, so the price drifts. The market gives too much weight to that last figure and too little to the clean remarks from the two runs before it, which show the error was a one-off.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.35 to 1.70 on roughly 50 to 80 bets",
      "falsified_if": "A/E at or below 1.15 on the broad rule, or the pull-up-after-error control scoring the same as or higher than the fallers and unseaters"
    },
    "notes": "The rule sketch is kept unchanged. Every column it uses is pre-off and has appeared in earlier rules in this thread. The refinements are taken only from the reasoning and the thread's findings: GB-only and under 16/1. The predicted sample (50 to 80) is below the 150-bet discovery floor, so a positive thin result should go to falsification and forward tracking, not be killed for size. The forward bar has been raised to A/E 1.05 so that the live record must still show an edge."
  }

Timeline

1. Job 2406 finished

   17 Oct, 02:09 · engine:compute-1

2. H-5840be9c Discovery → KILLED fails: 37 bets, needs 150; A/E 0.809, needs 1.05; p 0.7159, needs 0.05 or lower; periods [], each needs 1.0; A/E 0.809 against same-race control 1.024, needs 0.03 above (not a niche pass: A/E 0.809 below 1.05; p 0.7159182599778604 above 0.5; does not beat the same-race control 1.024)

   17 Oct, 02:09 · engine:evaluator

3. discovery result for H-5840be9c (broad rule): fail

   17 Oct, 02:09 · engine:evaluator

4. H-5840be9c Designed → RESEARCHdiscovery measuring on the discovery zone

   17 Oct, 02:09 · engine:orchestrator

5. compute-1 took job 2406 (evaluate, attempt 1)

   17 Oct, 02:09 · compute-1

6. Job 2404 finished

   17 Oct, 02:01 · engine:ai-1

7. Queued evaluate on the compute lane: discovery run of the broad rule

   17 Oct, 02:01 · engine:scheduler

8. H-5840be9c Proposed → RESEARCHdesigned pre-spec written and hashed

   17 Oct, 02:01 · agent:designer@v1

9. ai-1 took job 2404 (design, attempt 1)

   17 Oct, 02:01 · ai-1

10. Queued design on the ai lane: write the pre-spec

    17 Oct, 02:00 · engine:scheduler

11. New idea H-5840be9c: Sharpen: one-off chase error that ended the last run in a fall or unseat (sharpen)

    17 Oct, 02:00 · agent:investigator@v1

Every change is an event with an actor: an agent and its version, an engine, a human, or the importer. Current stage: Killed.