Fresh from a yard good with returners, when the last run before the break was ugly

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-6243bb06 (Test-mechanism: undropped recent winners after a heavy beating, given ) survived falsification: 6 of 6 attacks left it standing; it is frozen for its one vault query. O-3761601e
   • H-2af132dc (Generalise: an unchanged mark after a heavy beating lifts close-form j) passed discovery: A/E 1.31 on 506 bets (p 0.026), same-race control 0.98. O-388a8a14
   • H-8cff5453 (Replicate: recent winners beaten 20+ lengths in non-handicap jumps rac) was predicted positive but came out the other way: A/E 0.88 on 211 bets (p 0.268 for the opposite direction). O-6896f0b0
2. AskedDoes a yard's good record with fresh horses pay most when the break followed a bad run (pulled up, or beaten 20+ lengths), i.e. when the break was a repair?
3. MechanismA heavy beating or pull-up followed by a 90+ day break usually means something physical: a breathing problem, sore shins or a muscle issue that was then treated. Yards with a strong fresh record are good at bringing a horse back ready after time off. The old problem is gone, but the bad run still sits on top of the form.
4. Why the market misses itPunters and the odds compilers weight the last visible run. A PU or a 20-length beating followed by a long absence reads as a horse with a problem, so its price drifts. The lab's rule is to forgive what was ugly and physical. The yard's fresh record, which shows it fixes horses, is not on the card.
5. Predictedpositive, A/E 1.12 to 1.30 on 250+ bets
6. Falsified if

Mechanism and rationale

A heavy beating or pull-up followed by a 90+ day break usually means something physical: a breathing problem, sore shins or a muscle issue that was then treated. Yards with a strong fresh record are good at bringing a horse back ready after time off. The old problem is gone, but the bad run still sits on top of the form.

The rule

days_since_last_run >= 90 AND tx_fresh_runs >= 25 AND tx_fresh_wins >= 1.3 * tx_fresh_exp AND ((last_position IS NULL AND COALESCE(last_non_finish_code, 'PU') = 'PU') OR (last_field_size >= 6 AND last_position >= 0.75 * last_field_size))

Reads: days_since_last_run, tx_fresh_runs, tx_fresh_wins, tx_fresh_exp, last_position, last_non_finish_code, last_field_size

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

5d9bc895464ed5efc668300c36b15b300f9f8ebb2d38c39b0c7627e7ae907f90

Frozen

not frozen

The definition can still change.

Confounders

• Seasonal: October-November returners (jumps season openers) may carry an existing seasonal lift; check lift outside Oct-Nov.
• Code: jumps returners will dominate the bets, and PU is overwhelmingly a jumps outcome; report Flat and jumps separately.
• Price band: lift may be carried by a few big-priced winners; the 14/1-or-shorter refinement checks this.
• Yard concentration: a handful of big jumps yards could supply most of the bets and winners; report top-yard share and leave-one-yard-out.
• Non-finish coding: before 2025 every non-finisher is coded PU, so falls/unseats leak into the PU arm pre-2025; from 2025 they are excluded via last_non_finish_code.
• Proxy: last_beaten_lengths only exists from 2025-11, so 'ugly' is measured as PU or finishing in the back quarter of a 6+ runner field, which is not the same as a 20-length beating.
• Overlap with the yard-fresh finding (1.07) and 30-day yard form; the rule must beat the good-last-run control, not just 1.00.

Control

{
  "description": "Same strong-fresh yards, same 90+ day break, but the last run before the break was good (placed 1st-3rd). This is the falsification arm the hypothesis names: if these score as high or higher than the ugly-last-run group, the bad run adds nothing and any lift is just the yard's fresh record (already a known finding at ~1.07). Expect the control near 1.05-1.10 and the main rule at least ~0.07 A/E above it. The other runners in the same races should sit near 1.00.",
  "rule": "days_since_last_run >= 90 AND tx_fresh_runs >= 25 AND tx_fresh_wins >= 1.3 * tx_fresh_exp AND last_position IS NOT NULL AND last_position <= 3"
}

Discovery result

fail broad rule · 17 Oct, 05:01

Bets

500

24 winners

A/E

0.89

95%: 0.57 to 1.33

SP return

-46.0%

level stakes

Betfair SP return

-35.3%

485 bets with a BSP

Strike rate 4.8%. p = 0.74. Same-race control: the other runners in these races had A/E 1.00, 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, 04:57

  Output (schema-checked)

  {
    "rule": "days_since_last_run >= 90 AND tx_fresh_runs >= 25 AND tx_fresh_wins >= 1.3 * tx_fresh_exp AND ((last_position IS NULL AND COALESCE(last_non_finish_code, 'PU') = 'PU') OR (last_field_size >= 6 AND last_position >= 0.75 * last_field_size))",
    "refinements": [
      {
        "label": "14/1 or shorter",
        "rule": "days_since_last_run >= 90 AND tx_fresh_runs >= 25 AND tx_fresh_wins >= 1.3 * tx_fresh_exp AND ((last_position IS NULL AND COALESCE(last_non_finish_code, 'PU') = 'PU') OR (last_field_size >= 6 AND last_position >= 0.75 * last_field_size)) AND sp_decimal <= 15.0"
      },
      {
        "label": "pulled up last time only (the clearest physical-problem signal)",
        "rule": "days_since_last_run >= 90 AND tx_fresh_runs >= 25 AND tx_fresh_wins >= 1.3 * tx_fresh_exp AND last_position IS NULL AND COALESCE(last_non_finish_code, 'PU') = 'PU'"
      }
    ],
    "control": {
      "description": "Same strong-fresh yards, same 90+ day break, but the last run before the break was good (placed 1st-3rd). This is the falsification arm the hypothesis names: if these score as high or higher than the ugly-last-run group, the bad run adds nothing and any lift is just the yard's fresh record (already a known finding at ~1.07). Expect the control near 1.05-1.10 and the main rule at least ~0.07 A/E above it. The other runners in the same races should sit near 1.00.",
      "rule": "days_since_last_run >= 90 AND tx_fresh_runs >= 25 AND tx_fresh_wins >= 1.3 * tx_fresh_exp AND last_position IS NOT NULL AND last_position <= 3"
    },
    "pass_criteria": {
      "min_bets": 200,
      "min_ae": 1.08,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1.03,
      "max_p": 0.2
    },
    "confounders": [
      "Seasonal: October-November returners (jumps season openers) may carry an existing seasonal lift; check lift outside Oct-Nov.",
      "Code: jumps returners will dominate the bets, and PU is overwhelmingly a jumps outcome; report Flat and jumps separately.",
      "Price band: lift may be carried by a few big-priced winners; the 14/1-or-shorter refinement checks this.",
      "Yard concentration: a handful of big jumps yards could supply most of the bets and winners; report top-yard share and leave-one-yard-out.",
      "Non-finish coding: before 2025 every non-finisher is coded PU, so falls/unseats leak into the PU arm pre-2025; from 2025 they are excluded via last_non_finish_code.",
      "Proxy: last_beaten_lengths only exists from 2025-11, so 'ugly' is measured as PU or finishing in the back quarter of a 6+ runner field, which is not the same as a 20-length beating.",
      "Overlap with the yard-fresh finding (1.07) and 30-day yard form; the rule must beat the good-last-run control, not just 1.00."
    ],
    "mechanism": "A pull-up or a run near the back of the field, followed by 90+ days off, usually marks a physical problem (breathing, sore shins, muscle) that the break was used to fix. Yards with a proven record of winning with fresh horses (A/E 1.3+ on 25+ such runners) are the ones that reliably bring a horse back fit, so with them the break is a repair, not a decline. The horse returns with the problem gone but the ugly run still heading its form.",
    "market_miss": "Punters and compilers weight the last visible run. A PU or a tailed-off run followed by a long absence reads as a horse with a lingering problem, so the price drifts. The yard's fresh-horse record, which is the evidence that the problem was fixed, is not on the racecard and is rarely added up by bettors.",
    "prediction": {
      "direction": "positive",
      "size": "A/E about 1.12 to 1.30 on 250+ bets, and at least ~0.07 above the good-last-run control from the same yards",
      "falsified_if": "A/E below 1.03 on 200+ bets, or the control (same yards, 90+ days off, placed 1st-3rd last time) scores as high as or higher than the main rule, which would mean the ugly last run adds nothing beyond the yard's fresh record."
    },
    "notes": "I changed the draft rule. last_beaten_lengths (and best_last3_beaten, last_pulled_up and last_code) only exist from 2025-11-01, so 'beaten 20+ lengths' cannot be tested in a discovery zone that starts around 2021. tx_fresh is available from 2021-03. The rule therefore uses a proxy that exists across the full window: pulled up (last_position NULL), or finishing in the back quarter of a field of 6 or more. From 2025, falls, unseats and brought-downs are excluded with COALESCE(last_non_finish_code,'PU')='PU', because those are not physical-repair stories. sp_decimal is the price at the off and is used only for the 14/1 cut. Pass bar raised to 200 bets and A/E 1.08 so the rule has to beat the known yard-fresh baseline of 1.07 rather than just 1.00. Forward A/E bar raised to 1.03 for the same reason."
  }

Timeline

1. Job 3084 finished

   17 Oct, 05:01 · engine:compute-1

2. H-348bbae0 Discovery → KILLED fails: A/E 0.893, needs 1.08; p 0.7371, needs 0.05 or lower; periods [0.95, 0.82], each needs 1.0; A/E 0.893 against same-race control 1.004, needs 0.03 above (not a niche pass: A/E 0.893 below 1.08; p 0.7371136117870686 above 0.5; does not beat the same-race control 1.004; a period points the other way [0.95, 0.82])

   17 Oct, 05:01 · engine:evaluator

3. discovery result for H-348bbae0 (broad rule): fail

   17 Oct, 05:01 · engine:evaluator

4. H-348bbae0 Designed → RESEARCHdiscovery measuring on the discovery zone

   17 Oct, 05:01 · engine:orchestrator

5. compute-1 took job 3084 (evaluate, attempt 1)

   17 Oct, 05:01 · compute-1

6. Job 3080 finished

   17 Oct, 04:57 · engine:ai-1

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

   17 Oct, 04:57 · engine:scheduler

8. H-348bbae0 Proposed → RESEARCHdesigned pre-spec written and hashed

   17 Oct, 04:57 · agent:designer@v1

9. ai-1 took job 3080 (design, attempt 1)

   17 Oct, 04:57 · ai-1

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

    17 Oct, 04:56 · engine:scheduler

11. New idea H-348bbae0: Fresh from a yard good with returners, when the last run before the break was ugly (scout)

    17 Oct, 04:56 · agent:scout@v1

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