Test-mechanism: more bends, more lift; wide-trip narrow defeats over 10f+ last time

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-e6917785 (Test-mechanism: the wide-trip lift needs a last run with bends (7f+)) passed discovery: A/E 1.26 on 643 bets (p 0.004), same-race control 0.94. O-6497b473
2. AskedDoes the wide-trip lift grow with the number of bends in the last run (10f+ against 7 to 9f), or on sharper turning tracks? answered
3. MechanismThe ground lost to a wide trip grows with each bend taken wide. Middle-distance and staying races take in more turn than 7 to 8f races, so the hidden cost, and so the later rebound, should be larger.
4. Why the market misses itBeaten lengths are read the same way whatever the trip. Nobody adjusts the margin for how many turns the horse covered wide.
5. Predictedpositive, A/E 1.28 to 1.40 on about 300 bets, above the 7 to 9f remainder
6. Falsified ifA/E at or below 1.259 (the 7f+ parent), or the 7 to 9f remainder scoring as high or higher
7. Rule sketchrace_type = 'Flat' AND last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND sp_decimal <= 15 AND last_distance_yards >= 2200
8. question idQ-673da5a8
9. operatorsharpen

The question's answer so far: No dose-response on trip. Wide-trip narrow defeats over 10f+ score 1.17 on 206 (p 0.17), below the 7f+ cut's 1.26, so more bends do not mean more lift. The sharp-track half of the question is still untested.

Mechanism and rationale

The ground lost to a wide trip grows with each bend taken wide. Middle-distance and staying races take in more turn than 7 to 8f races, so the hidden cost, and so the later rebound, should be larger.

The rule

race_type = 'Flat' AND last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND sp_decimal <= 15 AND last_distance_yards >= 2200

Reads: race_type, last_raced_wide, last_position, last_beaten_lengths, sp_decimal, last_distance_yards

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

38e75a34666aadb40f8f3b48c33e81d9adbaa3d78738715fc59ce10c2e61c622

Frozen

not frozen

The definition can still change.

Confounders

• Staying races are a different population: older horses, more handicaps and more exposed form. The lift could belong to that population and have nothing to do with bends. Check it with the same-race control and with the share of handicaps.
• IRE share: earlier wide-trip results relied on IRE. The GB-only refinement checks this.
• Fewer bets: the 10f+ cut is a subset of the 7f+ parent (643 bets), so the A/E will be noisy and the CI wide.
• Pace of longer races: slower early pace may make a wide trip cost less, or make the finish more tactical. Beaten lengths may then mean something different, apart from the ground lost.
• Nested sample: the 10f+ cut is part of the parent that has already been seen. A high score could partly be the parent's own selection effect, so the remainder comparison matters more than the raw A/E.
• Track configuration: whether a 10f+ trip was run on a round course with several bends, or on a straight or near-straight course, is not measured here. Trip length is only a proxy for the number of bends.

Control

Discovery result

fail broad rule · 17 Oct, 04:41

Bets

206

42 winners

A/E

1.17

95%: 0.84 to 1.58

SP return

-6.0%

level stakes

Betfair SP return

+11.8%

206 bets with a BSP

Strike rate 20.4%. p = 0.17. 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

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 · 28s · 17 Oct, 04:40

  Output (schema-checked)

  {
    "rule": "race_type = 'Flat' AND last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND sp_decimal <= 15 AND last_distance_yards >= 2200",
    "refinements": [
      {
        "label": "dose-response: 12f+ last run (most bends)",
        "rule": "race_type = 'Flat' AND last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND sp_decimal <= 15 AND last_distance_yards >= 2640"
      },
      {
        "label": "GB only (IRE-dependence confounder)",
        "rule": "race_type = 'Flat' AND last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND sp_decimal <= 15 AND last_distance_yards >= 2200 AND region = 'GB'"
      }
    ],
    "control": {
      "description": "The lab's same-race comparison (other runners in the same races) should score about par (around 0.94 to 1.0, as with the 7f+ parent's control at 0.94). If they score high too, the lift belongs to the race type (staying races) and not to the wide-trip excuse. The key comparison is the 7f to under 10f remainder of the parent rule, using the same filters. The mechanism needs that remainder to score clearly below the 10f+ cut. If it scores as high or higher, the bends story fails.",
      "rule": "race_type = 'Flat' AND last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND sp_decimal <= 15 AND last_distance_yards >= 1540 AND last_distance_yards < 2200"
    },
    "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": [
      "Staying races are a different population: older horses, more handicaps and more exposed form. The lift could belong to that population and have nothing to do with bends. Check it with the same-race control and with the share of handicaps.",
      "IRE share: earlier wide-trip results relied on IRE. The GB-only refinement checks this.",
      "Fewer bets: the 10f+ cut is a subset of the 7f+ parent (643 bets), so the A/E will be noisy and the CI wide.",
      "Pace of longer races: slower early pace may make a wide trip cost less, or make the finish more tactical. Beaten lengths may then mean something different, apart from the ground lost.",
      "Nested sample: the 10f+ cut is part of the parent that has already been seen. A high score could partly be the parent's own selection effect, so the remainder comparison matters more than the raw A/E.",
      "Track configuration: whether a 10f+ trip was run on a round course with several bends, or on a straight or near-straight course, is not measured here. Trip length is only a proxy for the number of bends."
    ],
    "mechanism": "The ground a horse loses by racing wide grows with each bend it takes wide. A Flat run over 10f or more (2200 yards or more) usually covers more turn than a 7 to 9f run. So a narrow defeat (beaten 5 lengths or less, not the winner) after a wide trip at 10f+ hides more lost ground. Its true merit is understated by more, and the rebound next time should be bigger than in the 7 to 9f remainder.",
    "market_miss": "Punters and form readers read beaten lengths the same way whatever the trip and the course shape. A wide trip is noted only as a binary comment. Nobody scales the excuse by how many bends were covered wide, so the market under-rates the hidden ability most at the longer trips.",
    "prediction": {
      "direction": "positive",
      "size": "A/E about 1.28 to 1.40 on roughly 300 bets, clearly above the 7f-to-under-10f remainder",
      "falsified_if": "The broad rule's A/E is at or below 1.259 (the 7f+ parent's score), or the 7f-to-under-10f control remainder scores as high as the 10f+ cut or higher. Either result means there is no gradient with the number of bends."
    },
    "notes": "I kept the draft rule unchanged. All its columns are pre-off: last_raced_wide, last_position, last_beaten_lengths and last_distance_yards describe the previous run, and sp_decimal is the price at the off. A cap of sp_decimal <= 15 means 14/1 or shorter, which matches the parent. 2200 yards is 10f, and the 1540-yard bound in the control is 7f, to match the parent's 7f+ cut. The refinements test only the dose-response (12f+) and the IRE confounder. The pass criteria are tightened to A/E 1.10, because this is a sharpen of a parent already scoring 1.259 and a smaller lift would add nothing."
  }

Timeline

1. Job 2992 finished

   17 Oct, 04:41 · engine:compute-1

2. H-233f3726 Discovery → KILLED fails: p 0.1729, needs 0.05 or lower; periods [0.9, 1.33], each needs 1.0 (not a niche pass: a period points the other way [0.9, 1.33])

   17 Oct, 04:41 · engine:evaluator

3. discovery result for H-233f3726 (broad rule): fail

   17 Oct, 04:41 · engine:evaluator

4. H-233f3726 Designed → RESEARCHdiscovery measuring on the discovery zone

   17 Oct, 04:41 · engine:orchestrator

5. compute-1 took job 2992 (evaluate, attempt 1)

   17 Oct, 04:41 · compute-1

6. Job 2990 finished

   17 Oct, 04:40 · engine:ai-1

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

   17 Oct, 04:40 · engine:scheduler

8. H-233f3726 Proposed → RESEARCHdesigned pre-spec written and hashed

   17 Oct, 04:40 · agent:designer@v1

9. ai-1 took job 2990 (design, attempt 1)

   17 Oct, 04:40 · ai-1

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

    17 Oct, 04:39 · engine:scheduler

11. New idea H-233f3726: Test-mechanism: more bends, more lift; wide-trip narrow defeats over 10f+ last time (sharpen)

    17 Oct, 04:39 · 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.