Sharpen: troubled narrow loser back within 35 days

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-48a2eca6 (Test-mechanism: the edge sits with horses the market did not shorten a) passed discovery as a niche: A/E 1.24 on 104 bets (p 0.25), same-race control 0.97. O-22d00ba5
2. AskedDoes the edge fade as the gap since the troubled run grows, as it should if the run shows current form that the market misses? answered
3. MechanismThe troubled run shows the horse's current ability. A quick return means that form still applies. A long break lets fitness and handicap mark drift, so the hidden ability says less.
4. Why the market misses itThe price follows the figures (beaten, unplaced) and does not separately credit the trouble. That gap is widest while the form is fresh.
5. Predictedpositive, A/E 1.25 to 1.40 on 120+ bets, above the runners off for longer
6. Falsified ifA/E at or below that of the H19 runners off for more than 35 days, or below 1.10
7. Rule sketchlast_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND days_since_last_run <= 35
8. question id

Mechanism and rationale

The troubled run shows the horse's current ability. A quick return means that form still applies. A long break lets fitness and handicap mark drift, so the hidden ability says less.

The rule

last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 35

Reads: last_momentum_interrupted, last_position, last_beaten_lengths, days_since_last_run

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

bfd00b53d12422187ed8afa5d455aa40c74d6a9614f8849f6d3d54234370f37f

Frozen

16 Oct, 14:52

Nothing in the definition can change after this.

Confounders

• Quick returners are often in-form handicappers from busy, well-run yards. Yard form and trainer placement could carry the lift without the trouble mattering, so check that narrow losers with no trouble back within 35 days are at par.
• AW winter programme: quick returns cluster on all-weather cards with frequent races, so the lift could be a surface or season effect. Check the split by race_type and surface.
• Handicap mark not yet raised: a horse back quickly runs off the same mark before the handicapper reacts. This sits close to the mechanism, so check the is_handicap split.
• Overlap with H19: this is a subset of a rule already measured (242 bets), so the result is not independent of it. Most of its evidence must come from the contrast with the runners off for longer, not from the overall A/E.
• Exchange vs SP: the H19 lift was about three times larger at BSP, so a result that only appears at SP should be doubted.

Control

{
  "description": "Main control: H19 runners (momentum interrupted, beaten 2 lengths or less) off for MORE than 35 days. The mechanism says the edge decays with the gap, so these should score clearly lower, near the H19 average or below (around 1.0 to 1.1). If they score the same as or better than the fresh group, recency is not the driver. Same-race other runners should sit at about par (A/E near 0.97 to 1.0 after the overround), which shows the lift is not a property of the races. Second check: narrow losers with no trouble back within 35 days should also be near par (the earlier finding showed 0.98 overall), so any lift belongs to the trouble flag and not to the quick return alone.",
  "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND days_since_last_run > 35"
}

Discovery result

pass broad rule (niche) · 16 Oct, 14:49

Bets

200

45 winners

A/E

1.19

95%: 0.87 to 1.60

SP return

+13.2%

level stakes

Betfair SP return

+25.7%

198 bets with a BSP

Strike rate 22.5%. p = 0.13. 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. 4 of 5 survived.

• perturbation: Tighter recency cut: back within 28 dayspass

  178 bets · A/E 1.25 (0.91 to 1.69) · control 0.94 · 16 Oct, 14:52

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-27691.22

  after 2026-04-271091.28

• perturbation: Looser recency cut: back within 42 dayspass

  214 bets · A/E 1.23 (0.91 to 1.62) · control 0.95 · 16 Oct, 14:52

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, 14:52

  Breakdown

  betsA/E0.71.01.3

  forward0–

Other experiments

• neighbour: > 1 -> 0 (-20%)pass

  279 bets · A/E 1.09 (0.82 to 1.41) · control 0.99 · 16 Oct, 14:49

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271021.23

  after 2026-04-271770.99

• neighbour: > 1 -> 2 (+20%)pass

  121 bets · A/E 1.31 (0.86 to 1.91) · control 0.93 · 16 Oct, 14:49

  Breakdown

Lineage

Agent runs

• designer v1ok · 29s · 16 Oct, 14:48

  Output (schema-checked)

  {
    "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 35",
    "refinements": [
      {
        "label": "very fresh: back within 21 days",
        "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 21"
      }
    ],
    "control": {
      "description": "Main control: H19 runners (momentum interrupted, beaten 2 lengths or less) off for MORE than 35 days. The mechanism says the edge decays with the gap, so these should score clearly lower, near the H19 average or below (around 1.0 to 1.1). If they score the same as or better than the fresh group, recency is not the driver. Same-race other runners should sit at about par (A/E near 0.97 to 1.0 after the overround), which shows the lift is not a property of the races. Second check: narrow losers with no trouble back within 35 days should also be near par (the earlier finding showed 0.98 overall), so any lift belongs to the trouble flag and not to the quick return alone.",
      "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND days_since_last_run > 35"
    },
    "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": [
      "Quick returners are often in-form handicappers from busy, well-run yards. Yard form and trainer placement could carry the lift without the trouble mattering, so check that narrow losers with no trouble back within 35 days are at par.",
      "AW winter programme: quick returns cluster on all-weather cards with frequent races, so the lift could be a surface or season effect. Check the split by race_type and surface.",
      "Handicap mark not yet raised: a horse back quickly runs off the same mark before the handicapper reacts. This sits close to the mechanism, so check the is_handicap split.",
      "Overlap with H19: this is a subset of a rule already measured (242 bets), so the result is not independent of it. Most of its evidence must come from the contrast with the runners off for longer, not from the overall A/E.",
      "Exchange vs SP: the H19 lift was about three times larger at BSP, so a result that only appears at SP should be doubted."
    ],
    "mechanism": "A run interrupted by trouble while the horse was beaten 2 lengths or less hides ability the bare result does not show. That hidden ability is a statement about the horse's condition at the time. A quick return (35 days or less) means fitness, condition and handicap mark have hardly moved, so the hidden ability still applies today. After a longer break, fitness, mark and condition drift, so the troubled run says less about today's chance and the edge should shrink.",
    "market_miss": "The SP follows the figures (beaten, unplaced, the beaten distance) and gives no separate credit for the trouble. Earlier work showed price movement after the excuse does not pick out the edge. The mis-pricing should be biggest while the form is fresh, because the market treats a recent defeat as the most informative and reads it at face value.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.25 to 1.40 on about 150 to 200 bets, clearly above the H19 runners off for more than 35 days",
      "falsified_if": "A/E below 1.10, or A/E at or below that of the control (H19 runners off for more than 35 days)"
    },
    "notes": "I kept the rule sketch and added only an explicit days_since_last_run IS NOT NULL guard. Every column is pre-off. H19 has 242 bets, so this subset may land near or under the 150-bet floor. If it does, the main evidence is the contrast between the fresh group and the runners off for longer, and a thin positive result goes to falsification and forward tracking rather than being killed. The one refinement (21 days or less) tests the recency gradient the mechanism implies. I added no price-band refinement because the reasoning gives no case for one. Also report the BSP A/E, since the H19 edge was mainly at exchange prices."
  }

Timeline

1. Job 1866 finished

   16 Oct, 14:52 · engine:compute-1

2. forward result for H-b595a70a (forward vs bar {'min_bets': 100, 'min_ae': 1.0, 'max_p': 0.2}): info

   16 Oct, 14:52 · engine:evaluator

3. compute-1 took job 1866 (forward_check, attempt 1)

   16 Oct, 14:52 · compute-1

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

   16 Oct, 14:52 · engine:scheduler

5. Job 1863 finished

   16 Oct, 14:52 · engine:compute-1

6. H-b595a70a Frozen → SHADOW vault skipped: last_momentum_interrupted, last_beaten_lengths do not exist before the discovery zone; track it forward instead

   16 Oct, 14:52 · engine:orchestrator

7. compute-1 took job 1863 (vault, attempt 1)

   16 Oct, 14:52 · compute-1

8. Job 1862 finished

   16 Oct, 14:52 · engine:compute-1

9. Queued vault on the compute lane: frozen: one sealed vault query

   16 Oct, 14:52 · engine:scheduler

10. H-b595a70a Falsification → FROZEN survived falsification

    16 Oct, 14:52 · engine:falsification

11. falsification result for H-b595a70a (partition: Handicaps only: mark not yet raised): pass

    16 Oct, 14:52 · engine:evaluator

12. compute-1 took job 1862 (attack, attempt 1)

    16 Oct, 14:52 · compute-1

13. Job 1861 finished

    16 Oct, 14:52 · engine:compute-1

14. falsification result for H-b595a70a (price_band: Remove the up-to-7/1 band (outside 3/1 to 7/1)): pass

    16 Oct, 14:52 · engine:evaluator

15. compute-1 took job 1861 (attack, attempt 1)

    16 Oct, 14:52 · compute-1

16. Job 1860 finished

    16 Oct, 14:52 · engine:compute-1

17. falsification result for H-b595a70a (control: Control: untroubled narrow losers back within 35 days): fail

    16 Oct, 14:52 · engine:evaluator

18. compute-1 took job 1860 (attack, attempt 1)

    16 Oct, 14:52 · compute-1

19. Job 1859 finished

    16 Oct, 14:52 · engine:compute-1

20. falsification result for H-b595a70a (perturbation: Looser recency cut: back within 42 days): pass

    16 Oct, 14:52 · engine:evaluator