Like-for-like: ground-on-bends test inside the full parent base (sharp last-time tracks)

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-4e7cc5f4 (Wide trip after a turf sprint (no bends) still pays: the tag is not ab) passed discovery as a niche: A/E 1.18 on 129 bets (p 0.261), same-race control 0.97. O-9cf7e027
   • H-5b9fa0d3 (Wide trip last time on all-weather today still pays) passed discovery: A/E 1.23 on 462 bets (p 0.029), same-race control 0.95. O-bfd4e54e
2. AskedWithin the full parent rule (tilt exclusion and figures kept), does the wide-trip A/E differ between sharp and galloping last-time tracks, and between low and high draws today? answered
3. MechanismIf going wide costs extra ground on tight bends, then a close finish after a wide trip at a sharp track hides more ability than one at a galloping track. Earlier tests dropped the parent's exclusion of held-up or midfield runners from tilted races. Those runners were flattered by the pace shape, not by the wide trip, and may have masked the comparison.
4. Why the market misses itThe form book shows only a close-up placing. The wide note sits in running comments, and the price does not account for the course shape it happened on.
5. Predictedpositive, A/E 1.30 or more on about 500 bets, at least 0.08 above the same full base with last_track_sharpness = 0
6. Falsified ifA/E within 0.05 of, or below, the galloping-track version of the full parent base. That would confirm that track shape does not matter even like-for-like, and the ground-on-bends explanation should be retired.

Mechanism and rationale

If going wide costs extra ground on tight bends, then a close finish after a wide trip at a sharp track hides more ability than one at a galloping track. Earlier tests dropped the parent's exclusion of held-up or midfield runners from tilted races. Those runners were flattered by the pace shape, not by the wide trip, and may have masked the comparison.

The rule

last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND last_slow_start = 0 AND NOT coalesce(last_run_style IN ('midfield', 'held_up') AND last_race_tilt >= 0.2 AND (last_position - 1.0) / max(last_field_size - 1, 1) <= 0.4, 0) AND l_top_speed_pct IS NOT NULL AND last_track_sharpness >= 2

Reads: last_raced_wide, last_position, last_beaten_lengths, last_slow_start, last_run_style, last_race_tilt, last_field_size, l_top_speed_pct, last_track_sharpness

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

d2406436578692eb37cd8d471680119b53baa69ee46ed2b14098f9796a8e3ecf

Frozen

16 Oct, 12:35

Nothing in the definition can change after this.

Confounders

• Prior evidence already points the other way: galloping 1.23 on 270 against sharp 1.18 on 612, measured without the tilt exclusion. A like-for-like null is the base expectation.
• Sharp tracks are mostly all-weather ovals, so the split is partly by surface. The sharp-turf refinement separates track shape from surface, but its sample is small.
• last_raced_wide has only about one year of data, so there are only two periods, and the post-April-2026 concentration seen in sibling subsets may drive both arms.
• Requiring l_top_speed_pct limits the sample to RaceiQ-covered courses, which changes the course mix in each arm.
• The 16/1+ longshot lift can inflate either arm through a few winners. The 14/1 refinement checks this.
• Galloping control sample may be thin (under 200), so the sharp-galloping gap will have a wide CI.

Control

{
  "description": "The key comparison is the same full parent base with last_track_sharpness = 0 (galloping last-time tracks), keeping the tilt exclusion and figure requirement, so that only track shape changes. The mechanism predicts sharp >= galloping + 0.08. Other runners in the same races should be about A/E 0.95-1.0, which would show the lift is not a property of the races.",
  "rule": "last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND last_slow_start = 0 AND NOT coalesce(last_run_style IN ('midfield', 'held_up') AND last_race_tilt >= 0.2 AND (last_position - 1.0) / max(last_field_size - 1, 1) <= 0.4, 0) AND l_top_speed_pct IS NOT NULL AND last_track_sharpness = 0"
}

Discovery result

pass broad rule · 16 Oct, 12:29

Bets

571

102 winners

A/E

1.21

95%: 0.99 to 1.47

SP return

+29.7%

level stakes

Betfair SP return

+75.6%

570 bets with a BSP

Strike rate 17.9%. p = 0.033. 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.

• control: Like-for-like galloping control (last_track_sharpness = 0)pass

  238 bets · A/E 1.30 (0.97 to 1.70) · control 0.93 · 16 Oct, 12:35

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-27750.91

  after 2026-04-271631.49

• control: Same sharp-track close finish but NOT widefail

  10,420 bets · A/E 0.98 (0.93 to 1.03) · control 1.01 · 16 Oct, 12:35

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': 1.0, 'max_p': 0.2}info

  0 bets · · 16 Oct, 12:35

  Breakdown

  betsA/E0.71.01.3

  forward0–

Other experiments

• neighbour: > 1 -> 0 (-20%)pass

  802 bets · A/E 1.12 (0.94 to 1.32) · control 0.97 · 16 Oct, 12:35

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-273941.15

  after 2026-04-274081.09

• neighbour: > 1 -> 2 (+20%)pass

  405 bets · A/E 1.34 (1.04 to 1.71) · control 0.95 · 16 Oct, 12:35

  Breakdown

Lineage

Agent runs

• designer v1ok · 36s · 16 Oct, 12:26

  Output (schema-checked)

  {
    "rule": "last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND last_slow_start = 0 AND NOT coalesce(last_run_style IN ('midfield', 'held_up') AND last_race_tilt >= 0.2 AND (last_position - 1.0) / max(last_field_size - 1, 1) <= 0.4, 0) AND l_top_speed_pct IS NOT NULL AND last_track_sharpness >= 2",
    "refinements": [
      {
        "label": "sharp-turf-last-time",
        "rule": "last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND last_slow_start = 0 AND NOT coalesce(last_run_style IN ('midfield', 'held_up') AND last_race_tilt >= 0.2 AND (last_position - 1.0) / max(last_field_size - 1, 1) <= 0.4, 0) AND l_top_speed_pct IS NOT NULL AND last_track_sharpness >= 2 AND lower(last_surface) = 'turf'"
      },
      {
        "label": "14-1-or-shorter",
        "rule": "last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND last_slow_start = 0 AND NOT coalesce(last_run_style IN ('midfield', 'held_up') AND last_race_tilt >= 0.2 AND (last_position - 1.0) / max(last_field_size - 1, 1) <= 0.4, 0) AND l_top_speed_pct IS NOT NULL AND last_track_sharpness >= 2 AND sp_decimal <= 15"
      }
    ],
    "control": {
      "description": "The key comparison is the same full parent base with last_track_sharpness = 0 (galloping last-time tracks), keeping the tilt exclusion and figure requirement, so that only track shape changes. The mechanism predicts sharp >= galloping + 0.08. Other runners in the same races should be about A/E 0.95-1.0, which would show the lift is not a property of the races.",
      "rule": "last_raced_wide = 1 AND last_position > 1 AND last_beaten_lengths <= 5 AND last_slow_start = 0 AND NOT coalesce(last_run_style IN ('midfield', 'held_up') AND last_race_tilt >= 0.2 AND (last_position - 1.0) / max(last_field_size - 1, 1) <= 0.4, 0) AND l_top_speed_pct IS NOT NULL AND last_track_sharpness = 0"
    },
    "pass_criteria": {
      "min_bets": 150,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "Prior evidence already points the other way: galloping 1.23 on 270 against sharp 1.18 on 612, measured without the tilt exclusion. A like-for-like null is the base expectation.",
      "Sharp tracks are mostly all-weather ovals, so the split is partly by surface. The sharp-turf refinement separates track shape from surface, but its sample is small.",
      "last_raced_wide has only about one year of data, so there are only two periods, and the post-April-2026 concentration seen in sibling subsets may drive both arms.",
      "Requiring l_top_speed_pct limits the sample to RaceiQ-covered courses, which changes the course mix in each arm.",
      "The 16/1+ longshot lift can inflate either arm through a few winners. The 14/1 refinement checks this.",
      "Galloping control sample may be thin (under 200), so the sharp-galloping gap will have a wide CI."
    ],
    "mechanism": "If a wide trip costs extra ground mainly on bends, a close finish (beaten 5L or less, not first, no slow start) after a wide trip at a sharp, tight-turning track hides more ability than the same finish at a galloping track. Held-up or midfield runners who finished in the top 40% of tilted races are excluded, as in the parent, because their close finish reflects the pace shape and not ability hidden by the trip.",
    "market_miss": "The form figures show only a close-up placing. The wide trip appears only in running comments, and the price does not adjust for how much the course's bends magnified the ground lost.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.30 or more on about 500 bets, at least 0.08 above the same full base restricted to last_track_sharpness = 0",
      "falsified_if": "The sharp-track A/E is within 0.05 of, or below, the galloping-track version of the full parent base (the control rule). That would confirm track shape does not matter even like-for-like, and the ground-on-bends explanation should be retired."
    },
    "notes": "The rule sketch is kept as given. It holds the full parent base (tilt exclusion plus figure requirement) and changes only last-time track sharpness, as the earlier finding on refinements required. All columns are pre-off last-race fields. The comparison against the galloping control is the main test. Passing the absolute bar alone does not confirm the mechanism, because the parent base already shows a lift. The refinements are suggested by the reasoning and the confounders: sharp turf separates track shape from all-weather surface, and the 14/1 cap guards against the longshot lift."
  }

Timeline

1. Job 1388 finished

   16 Oct, 12:35 · engine:compute-1

2. forward result for H-1eed6021 (forward vs bar {'min_bets': 100, 'min_ae': 1.0, 'max_p': 0.2}): info

   16 Oct, 12:35 · engine:evaluator

3. compute-1 took job 1388 (forward_check, attempt 1)

   16 Oct, 12:35 · compute-1

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

   16 Oct, 12:35 · engine:scheduler

5. Job 1360 finished

   16 Oct, 12:35 · engine:compute-1

6. hypothesis note H-1eed6021

   16 Oct, 12:35 · engine:perturb

7. perturbation result for H-1eed6021 (neighbour: >= 0.2 -> 0.24 (+20%)): pass

   16 Oct, 12:35 · engine:perturb

8. perturbation result for H-1eed6021 (neighbour: >= 0.2 -> 0.16 (-20%)): pass

   16 Oct, 12:35 · engine:perturb

9. perturbation result for H-1eed6021 (neighbour: <= 5 -> 6 (+20%)): pass

   16 Oct, 12:35 · engine:perturb

10. perturbation result for H-1eed6021 (neighbour: <= 5 -> 4 (-20%)): pass

    16 Oct, 12:35 · engine:perturb

11. perturbation result for H-1eed6021 (neighbour: > 1 -> 2 (+20%)): pass

    16 Oct, 12:35 · engine:perturb

12. perturbation result for H-1eed6021 (neighbour: > 1 -> 0 (-20%)): pass

    16 Oct, 12:35 · engine:perturb

13. compute-1 took job 1360 (mutate, attempt 1)

    16 Oct, 12:35 · compute-1

14. Job 1383 finished

    16 Oct, 12:35 · engine:compute-1

15. H-1eed6021 Frozen → SHADOW vault skipped: last_raced_wide, last_beaten_lengths, last_slow_start, last_run_style, last_race_tilt do not exist before the discovery zone; track it forward instead

    16 Oct, 12:35 · engine:orchestrator

16. compute-1 took job 1383 (vault, attempt 1)

    16 Oct, 12:35 · compute-1