Pace victim: led or raced prominently in a race that collapsed, beaten

Mechanism and rationale

When a race's pace collapses (last_race_tilt >= 0.2, where closers were favoured), the front-runners are physically exhausted by the early speed and drop back whatever their ability. This is the mirror image of the flattered closer, which is already a strong negative (0.76 at its best). If the hold-up horses that finished close were flattered, the leaders who weakened were understated.

The rule

last_run_style IN ('led', 'prominent') AND last_race_tilt >= 0.2 AND last_position > 1 AND last_field_size >= 8

Reads: last_run_style, last_race_tilt, last_position, last_field_size

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

3038c0424b796d4c79e97cf40b7fd33bae3b7fe8749b23c34bc56636a51057a5

Frozen

15 Oct, 23:24

Nothing in the definition can change after this.

Confounders

• Self-inclusion in the tilt: a beaten leader pushes its own race's tilt towards 'closers favoured', so a high tilt can partly be the horse's own failure. last_field_size >= 8 dilutes this; the neutral-tilt control checks it.
• Genuine non-stayers: 'led, weakened' can be a trip problem; refinement 1 removes horses stepping up in trip, and the lab should split by trip change.
• Soft-ground front-runner that weakened, already a negative inside H120: split by going_group and last_going_group.
• Front-runners as a trait group (lone-front-runner pace maps were 0.97): the neutral-tilt control separates trait from race shape.
• Track pace bias: some courses favour closers by layout; split by last_track_pace_bias and walk any course pattern forward.
• Today's pace: a leader meeting another collapse is not helped; split by n_front_last / n_led_last in today's race.
• Short window: last_race_tilt, last_run_style and last_beaten_lengths exist only from 2025-11-01, so there is about a year of data; the period test is two halves, not three periods.

Discovery result

pass broad rule (niche) · 15 Oct, 23:23

Bets

2,674

311 winners

A/E

1.07

95%: 0.96 to 1.20

SP return

-20.6%

level stakes

Betfair SP return

-1.8%

2,660 bets with a BSP

Strike rate 11.6%. p = 0.12. Same-race control: the other runners in these races had A/E 0.99, so the selection beat its own races. Zone: discovery. Evaluator eval-1.0, data ef1163efa87ef7d2

Falsification

Each attack is a deliberate attempt to kill the idea. 6 of 6 survived.

• perturbation: Stronger collapse: tilt threshold raised 25% to 0.25pass

  1,935 bets · A/E 1.09 (0.95 to 1.24) · control 0.99 · 15 Oct, 23:24

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-277961.20

  after 2026-04-271,1391.01

• perturbation: Milder collapse: tilt threshold lowered 20% to 0.16pass

  3,360 bets · A/E 1.08 (0.97 to 1.19) · control 0.99 · 15 Oct, 23:24

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': 150, 'min_ae': 1.03, 'max_p': 0.2}info

  0 bets · · 15 Oct, 23:24

  Breakdown

  betsA/E0.71.01.3

  forward0–

• forward vs bar {'min_bets': 150, 'min_ae': 1.03, 'max_p': 0.2}info

  0 bets · · 16 Oct, 23:51

  Breakdown

  betsA/E0.71.01.3

  forward0–

Other experiments

• neighbour: >= 0.2 -> 0.16 (-20%)pass

  3,360 bets · A/E 1.08 (0.97 to 1.19) · control 0.99 · 16 Oct, 11:02

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271,4171.15

  after 2026-04-271,9431.02

• neighbour: >= 0.2 -> 0.24 (+20%)pass

  2,042 bets · A/E 1.07 (0.94 to 1.22) · control 0.99 · 16 Oct, 11:02

  Breakdown

Lineage

No parent and no descendants yet. When an agent mutates, generalises or splits this idea, the new hypotheses appear here.

Agent runs

• designer v1ok · 43s · 15 Oct, 23:22

  Output (schema-checked)

  {
    "rule": "last_run_style IN ('led', 'prominent') AND last_race_tilt >= 0.2 AND last_position > 1 AND last_field_size >= 8",
    "refinements": [
      {
        "label": "Not up in trip today (stamina doubt not being asked again)",
        "rule": "last_run_style IN ('led', 'prominent') AND last_race_tilt >= 0.2 AND last_position > 1 AND last_field_size >= 8 AND distance_yards <= last_distance_yards"
      },
      {
        "label": "Looked ugly: beaten 3 lengths or more (the result the form book misreads)",
        "rule": "last_run_style IN ('led', 'prominent') AND last_race_tilt >= 0.2 AND last_position > 1 AND last_field_size >= 8 AND last_beaten_lengths >= 3"
      },
      {
        "label": "14/1 or shorter, because edges die at big prices",
        "rule": "last_run_style IN ('led', 'prominent') AND last_race_tilt >= 0.2 AND last_position > 1 AND last_field_size >= 8 AND sp_decimal <= 15"
      }
    ],
    "control": {
      "description": "Same-race other runners should sit at the usual RaceiQ control level (about 1.03 to 1.05); the rule must clear them clearly. A second control separates the race-shape effect from the front-runner trait: beaten leaders/prominent runners whose last race was neutrally run (tilt near zero). If the beaten pace horses from neutral races show the same A/E as those from collapsed races, the edge (if any) is a front-runner trait, not a pace-victim effect, and the hypothesis fails.",
      "rule": "last_run_style IN ('led', 'prominent') AND last_race_tilt > -0.1 AND last_race_tilt < 0.1 AND last_position > 1 AND last_field_size >= 8"
    },
    "pass_criteria": {
      "min_bets": 300,
      "min_ae": 1.05,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 150,
      "min_ae": 1.03,
      "max_p": 0.2
    },
    "confounders": [
      "Self-inclusion in the tilt: a beaten leader pushes its own race's tilt towards 'closers favoured', so a high tilt can partly be the horse's own failure. last_field_size >= 8 dilutes this; the neutral-tilt control checks it.",
      "Genuine non-stayers: 'led, weakened' can be a trip problem; refinement 1 removes horses stepping up in trip, and the lab should split by trip change.",
      "Soft-ground front-runner that weakened, already a negative inside H120: split by going_group and last_going_group.",
      "Front-runners as a trait group (lone-front-runner pace maps were 0.97): the neutral-tilt control separates trait from race shape.",
      "Track pace bias: some courses favour closers by layout; split by last_track_pace_bias and walk any course pattern forward.",
      "Today's pace: a leader meeting another collapse is not helped; split by n_front_last / n_led_last in today's race.",
      "Short window: last_race_tilt, last_run_style and last_beaten_lengths exist only from 2025-11-01, so there is about a year of data; the period test is two halves, not three periods."
    ],
    "mechanism": "When the early pace of a race is too strong and it collapses (last_race_tilt >= 0.2, closers favoured), the horses that led or raced prominently are tired by the early speed and drop back whatever their ability. Their finishing position and beaten distance understate them. This is the mirror image of the flattered closer, already a strong negative (0.76 at best): if the hold-up horses were flattered by the shape, the pace horses were understated. In a normal-shaped race next time they should run to their real level.",
    "market_miss": "The form book and punters read 'led, headed, weakened' as a horse that doesn't stay or isn't good enough, and price it on its position and beaten distance. Only a race-shape view shows that every horse up with the pace suffered. The flattered-closer negative shows the market does not correct for race shape, so the same blind spot should leave the pace victims too long at SP.",
    "notes": "Changes from the draft: added last_field_size >= 8 so the horse's own result has less weight in its race's tilt (the tilt includes every finisher, this horse too). last_position > 1 already leaves out non-finishers (NULL), which is right: pulled up or fell is a different story. The style values 'led' and 'prominent' come from the draft and could not be checked in the catalogue (it gives no values for last_run_style). If the lab finds the labels are different (e.g. 'front'), the rule must be read as led-or-prominent with the same meaning, not widened. The tilt scale is not documented either; 0.2 is kept as the pre-registered cutoff, with no tuning. Data only runs from 2025-11-01, so judge it on two halves."
  }

Timeline

1. Job 2116 finished

   16 Oct, 23:51 · engine:compute-1

2. forward result for H-22ef2116 (forward vs bar {'min_bets': 150, 'min_ae': 1.03, 'max_p': 0.2}): info

   16 Oct, 23:51 · engine:evaluator

3. compute-1 took job 2116 (forward_check, attempt 1)

   16 Oct, 23:51 · compute-1

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

   16 Oct, 23:51 · engine:scheduler

5. Job 1219 finished

   16 Oct, 11:02 · engine:compute-1

6. hypothesis note H-22ef2116

   16 Oct, 11:02 · engine:perturb

7. perturbation result for H-22ef2116 (neighbour: >= 8 -> 10 (+20%)): pass

   16 Oct, 11:02 · engine:perturb

8. perturbation result for H-22ef2116 (neighbour: >= 8 -> 6 (-20%)): pass

   16 Oct, 11:02 · engine:perturb

9. perturbation result for H-22ef2116 (neighbour: > 1 -> 2 (+20%)): pass

   16 Oct, 11:02 · engine:perturb

10. perturbation result for H-22ef2116 (neighbour: > 1 -> 0 (-20%)): pass

    16 Oct, 11:02 · engine:perturb

11. perturbation result for H-22ef2116 (neighbour: >= 0.2 -> 0.24 (+20%)): pass

    16 Oct, 11:02 · engine:perturb

12. perturbation result for H-22ef2116 (neighbour: >= 0.2 -> 0.16 (-20%)): pass

    16 Oct, 11:02 · engine:perturb

13. compute-1 took job 1219 (mutate, attempt 1)

    16 Oct, 11:02 · compute-1

14. Queued mutate on the compute lane: parameter stability: do neighbouring thresholds hold?

    16 Oct, 11:02 · engine:scheduler

15. Job 39 finished

    15 Oct, 23:24 · engine:compute-1

16. forward result for H-22ef2116 (forward vs bar {'min_bets': 150, 'min_ae': 1.03, 'max_p': 0.2}): info

    15 Oct, 23:24 · engine:evaluator