Rival: pulled up after leading or racing prominently shows the same lift

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
   • H36-pulled-up-after-being-held-up (H36: pulled up after being held up) was GREEN in the earlier research: A/E 1.19 on 1451 bets (periods 1.07 / 1.41); now PROMISING. O-309e6a2d
2. AskedDo pulled-up horses that led or raced prominently (and so likely stopped) fail to show the lift, confirming that run style is doing the work? answered
3. MechanismIf the market simply over-punishes any P, front-runners who pulled up should beat their price as well. If H36's mechanism is right, they stopped while trying, and their price is fair or too short.
4. Why the market misses itUnder the rival, punters treat every P alike. Under H36, the front-runner's P really does reflect the horse failing, so it is priced correctly.
5. Predictednegative, A/E 0.85 to 1.00 on 600+ bets (the H36 mechanism predicts no lift)
6. Falsified ifA/E at or above 1.15, which would mean the held-up condition is not the cause
7. Rule sketchlast_pulled_up = 1 AND last_run_style IN ('led', 'prominent')
8. question idQ-a48c137a
9. operatorrival-explanation

The question's answer so far: Mostly yes. Led/prominent pulled-up horses showed no lift (A/E 0.916 on 727, control 1.01) against 1.198 for held-up ones, so run style is doing the work, though the negative side is not significant (CI 0.67 to 1.22).

Mechanism and rationale

If the market simply over-punishes any P, front-runners who pulled up should beat their price as well. If H36's mechanism is right, they stopped while trying, and their price is fair or too short.

The rule

last_pulled_up = 1 AND last_run_style IN ('led', 'prominent')

Reads: last_pulled_up, last_run_style

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

63cbcb0ecd95f28cd9b24692ebc4ca72af8fb66566880e0dbe84ab5056d1da41

Frozen

not frozen

The definition can still change.

Confounders

• Front-runners that pull up are more likely to have had a physical problem (bleeding, wind, lameness). That would push their next-run A/E down for reasons other than 'stopped while trying'.
• Price bands differ between styles. Front-runner PUs may be shorter in the betting, and the favourite-longshot bias then differs. Compare at 14/1 or shorter as well as broad.
• Run style labels may be noisy or missing for PU runs, especially when the horse was pulled up early. Mislabelled held-up horses would leak H36's lift into this pool.
• Data window: last_run_style and last_pulled_up only start from 2025-11-01, so the sample is about one season and overlaps H36's period exactly.
• Race type mix: front-running PUs may be more common in chases, where jumping errors (a separate story edge) are involved.

Control

{
  "description": "Two comparisons. (1) Same-race control: the other runners in the same races should come out near 1.00, so any departure in the rule group is about the horse, not the race. (2) The real contrast is with H36, the held-up/midfield PU pool, measured over the same window (broad A/E 1.19 on 1451; 1.253 on 510 at 14/1 or shorter). The control rule restates H36 so both groups are scored on the same dates and data. If front-runner PUs show a lift close to H36's, the 'any P is over-punished' rival wins. If they sit at or below 1.00 while H36 holds up, the lift is specific to held-up horses.",
  "rule": "last_pulled_up = 1 AND last_run_style IN ('held_up', 'midfield')"
}

Discovery result

fail broad rule · 17 Oct, 01:36

Bets

727

47 winners

A/E

0.92

95%: 0.67 to 1.22

SP return

-28.9%

level stakes

Betfair SP return

-6.1%

726 bets with a BSP

Strike rate 6.5%. p = 0.74. Same-race control: the other runners in these races had A/E 1.01, 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 · 50s · 17 Oct, 01:34

  Output (schema-checked)

  {
    "rule": "last_pulled_up = 1 AND last_run_style IN ('led', 'prominent')",
    "refinements": [
      {
        "label": "14/1 or shorter (matches the live H36 system band)",
        "rule": "last_pulled_up = 1 AND last_run_style IN ('led', 'prominent') AND sp_decimal <= 15"
      }
    ],
    "control": {
      "description": "Two comparisons. (1) Same-race control: the other runners in the same races should come out near 1.00, so any departure in the rule group is about the horse, not the race. (2) The real contrast is with H36, the held-up/midfield PU pool, measured over the same window (broad A/E 1.19 on 1451; 1.253 on 510 at 14/1 or shorter). The control rule restates H36 so both groups are scored on the same dates and data. If front-runner PUs show a lift close to H36's, the 'any P is over-punished' rival wins. If they sit at or below 1.00 while H36 holds up, the lift is specific to held-up horses.",
      "rule": "last_pulled_up = 1 AND last_run_style IN ('held_up', 'midfield')"
    },
    "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": [
      "Front-runners that pull up are more likely to have had a physical problem (bleeding, wind, lameness). That would push their next-run A/E down for reasons other than 'stopped while trying'.",
      "Price bands differ between styles. Front-runner PUs may be shorter in the betting, and the favourite-longshot bias then differs. Compare at 14/1 or shorter as well as broad.",
      "Run style labels may be noisy or missing for PU runs, especially when the horse was pulled up early. Mislabelled held-up horses would leak H36's lift into this pool.",
      "Data window: last_run_style and last_pulled_up only start from 2025-11-01, so the sample is about one season and overlaps H36's period exactly.",
      "Race type mix: front-running PUs may be more common in chases, where jumping errors (a separate story edge) are involved."
    ],
    "mechanism": "H36 says a held-up or midfield horse that pulled up was often never asked for an effort, so the P overstates how badly it ran and the market over-punishes it. A horse that led or raced prominently and then pulled up was in the race and stopped: it tired, failed, or had a problem. Its P is real information. The rival says the market discounts every P alike. If so, front-runner PUs should show the same lift as held-up PUs. If H36 is right, they should be priced fairly or too short.",
    "market_miss": "Under the rival, punters and odds compilers read the P in the form figures without looking at how the race was run, so every PU is marked down alike. Under H36, the market's discount is right for front-runners, whose P reflects a real failure. The mispricing exists only where the P hides a non-effort, as with held-up horses.",
    "prediction": {
      "direction": "negative",
      "size": "A/E 0.85 to 1.00 on 600+ bets, i.e. no lift. The H36 mechanism predicts fair or slightly over-bet prices.",
      "falsified_if": "A broad A/E at or above 1.15, or at or above 1.15 at 14/1 or shorter, on 150 or more bets. That would mean the held-up condition is not the cause of H36's lift and the rival 'any P is over-punished' explanation holds."
    },
    "notes": "This is a rival-explanation test, not an edge hunt. The result that supports H36 is a null (A/E near or below 1.00), which will not, and need not, pass the discovery bar as a negative edge. Read the outcome as follows: A/E of 1.15 or more kills the held-up specificity of H36. A/E in 0.85–1.05 while the H36 control holds supports H36. A/E of 0.95 or below with p ≤ 0.05 on 150+ bets would be a reportable over-bet effect. The only refinement is the sp_decimal ≤ 15 band, which mirrors the live H36 system so the contrast is like for like. Run style values follow the H36 labels ('held_up', 'midfield'); 'led' and 'prominent' are assumed to be the matching labels. If either returns zero rows, the rule is mis-specified and must be flagged, not reinterpreted."
  }

Timeline

1. Job 2281 finished

   17 Oct, 01:36 · engine:compute-1

2. H-a2870576 Discovery → KILLED fails: p 0.3043 (lower tail), needs 0.05; periods [0.79, 1.11], each needs 0.99 or lower

   17 Oct, 01:36 · engine:evaluator

3. discovery result for H-a2870576 (broad rule): fail

   17 Oct, 01:36 · engine:evaluator

4. H-a2870576 Designed → RESEARCHdiscovery measuring on the discovery zone

   17 Oct, 01:36 · engine:orchestrator

5. compute-1 took job 2281 (evaluate, attempt 1)

   17 Oct, 01:36 · compute-1

6. Job 2278 finished

   17 Oct, 01:35 · engine:ai-1

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

   17 Oct, 01:35 · engine:scheduler

8. H-a2870576 Proposed → RESEARCHdesigned pre-spec written and hashed

   17 Oct, 01:35 · agent:designer@v1

9. ai-1 took job 2278 (design, attempt 1)

   17 Oct, 01:34 · ai-1

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

    17 Oct, 01:33 · engine:scheduler

11. New idea H-a2870576: Rival: pulled up after leading or racing prominently shows the same lift (rival-explanation)

    17 Oct, 01:33 · 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.