Sharpen: troubled narrow losers in handicaps, where the mark did not charge for the hidden run

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-b595a70a (Sharpen: troubled narrow loser back within 35 days) passed discovery as a niche: A/E 1.19 on 200 bets (p 0.134), same-race control 0.96. O-0fc94a65
   • H-fe5384b6 (Rival: closers that made late headway and were beaten narrowly without) was predicted positive but came out the other way: A/E 0.97 on 5540 bets (p 0.172 for the opposite direction). O-f4ebff20
   • H-4882f53a (Generalise: checked mid-run and beaten 2 to 5 lengths still beats its ) was predicted positive but came out the other way: A/E 0.8 on 263 bets (p 0.148 for the opposite direction). O-fd5d7674
2. AskedIs the H19 lift concentrated in handicaps, where a troubled narrow defeat earns little or no rise in the mark, so the horse runs off a rating below its true level? answered
3. MechanismA horse checked mid-run and beaten 2 lengths or less would probably have won. The handicapper rates the margin seen, not the run lost, so the horse runs next time off a mark a few pounds below its true level. In non-handicaps no rating is involved, and the edge has to come from the market alone.
4. Why the market misses itThe result shows a narrow defeat, which the market prices as an ordinary near-miss. Narrow defeat alone is priced correctly (0.979), so the market does not add the pounds the check cost. The handicap mark makes this worse by not charging for them either.
5. Predictedpositive, A/E 1.25 to 1.40 on 120+ bets, above H19 non-handicap runners
6. Falsified ifA/E below 1.10, or at or below the non-handicap H19 runners on comparable numbers

The question's answer so far: No. Handicap-only H19 scored A/E 1.145 on 190 (control 0.98), below the parent's 1.28 on 218, so the lift is not concentrated where the handicap mark could hide ability.

Mechanism and rationale

A horse checked mid-run and beaten 2 lengths or less would probably have won. The handicapper rates the margin seen, not the run lost, so the horse runs next time off a mark a few pounds below its true level. In non-handicaps no rating is involved, and the edge has to come from the market alone.

The rule

last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND is_handicap = 1

Reads: last_momentum_interrupted, last_position, last_beaten_lengths, is_handicap

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

3f0d386f14ec11aaf46b6966ecaa6658613fadbb2eb64f189c9114710df3c77d

Frozen

not frozen

The definition can still change.

Confounders

• Field size: handicaps have bigger fields, so long-priced runners are more common and A/E is noisier. Compare A/E within price bands.
• Price band: handicaps have few odds-on runners. Check the lift is not carried by one band, such as 16/1 or bigger, where H19's neighbour already showed wild swings.
• Nurseries are a separate juvenile population with unstable marks (handled by a refinement).
• Class level: lower-class handicaps may be priced less efficiently, whatever the trouble.
• Period drift: H19 went from 1.07 to 1.26 across periods. Check the handicap subset holds in each period.
• Sample split: H19 had only 242 bets in total, so the handicap subset may be thin and the non-handicap control thinner still. The comparison may lack power.
• Exchange vs SP: H19's edge was mostly at BSP (ROI 19% against 6% at SP). Judge the result by A/E, not SP ROI.

Control

{
  "description": "The same-race other runners should sit near par (A/E about 1.0), which shows the lift is not a property of the handicap races themselves. The main control is the H19 complement in non-handicaps (same trouble and margin conditions, is_handicap = 0). The mechanism predicts the handicap group scores clearly above it. If the non-handicap group scores the same or higher, the lift belongs to the market's failure to credit trouble, not to the handicap mark, and the sharpening fails.",
  "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND is_handicap = 0"
}

Discovery result

pass broad rule (niche) · 16 Oct, 14:52

Bets

190

37 winners

A/E

1.15

95%: 0.81 to 1.58

SP return

-2.4%

level stakes

Betfair SP return

+5.3%

188 bets with a BSP

Strike rate 19.5%. p = 0.23. Same-race control: the other runners in these races had A/E 0.98, 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. 6 of 6 survived.

• control: Control: same trouble and margin in non-handicapspass

  52 bets · A/E 1.28 (0.70 to 2.15) · control 0.91 · 16 Oct, 14:55

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-2715–

  after 2026-04-2737–

• partition: Core mechanism: troubled run was itself a handicap (mark set from the seen margin)pass

  180 bets · A/E 1.07 (0.74 to 1.51) · control 0.99 · 16 Oct, 14:55

  Breakdown

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.

Other experiments

• neighbour: > 1 -> 0 (-20%)pass

  275 bets · A/E 1.08 (0.81 to 1.42) · control 0.99 · 16 Oct, 14:55

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271091.09

  after 2026-04-271661.08

• neighbour: > 1 -> 2 (+20%)pass

  122 bets · A/E 1.21 (0.77 to 1.82) · control 0.96 · 16 Oct, 14:55

  Breakdown

Lineage

Agent runs

• designer v1ok · 27s · 16 Oct, 14:50

  Output (schema-checked)

  {
    "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND is_handicap = 1",
    "refinements": [
      {
        "label": "handicap-to-handicap: the troubled run itself was a handicap, so the mark was set straight from the margin seen",
        "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND is_handicap = 1 AND last_is_handicap = 1"
      },
      {
        "label": "exclude nurseries, a separate juvenile population with lightly-raced, fast-changing marks",
        "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND is_handicap = 1 AND is_nursery = 0"
      }
    ],
    "control": {
      "description": "The same-race other runners should sit near par (A/E about 1.0), which shows the lift is not a property of the handicap races themselves. The main control is the H19 complement in non-handicaps (same trouble and margin conditions, is_handicap = 0). The mechanism predicts the handicap group scores clearly above it. If the non-handicap group scores the same or higher, the lift belongs to the market's failure to credit trouble, not to the handicap mark, and the sharpening fails.",
      "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths <= 2 AND is_handicap = 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": [
      "Field size: handicaps have bigger fields, so long-priced runners are more common and A/E is noisier. Compare A/E within price bands.",
      "Price band: handicaps have few odds-on runners. Check the lift is not carried by one band, such as 16/1 or bigger, where H19's neighbour already showed wild swings.",
      "Nurseries are a separate juvenile population with unstable marks (handled by a refinement).",
      "Class level: lower-class handicaps may be priced less efficiently, whatever the trouble.",
      "Period drift: H19 went from 1.07 to 1.26 across periods. Check the handicap subset holds in each period.",
      "Sample split: H19 had only 242 bets in total, so the handicap subset may be thin and the non-handicap control thinner still. The comparison may lack power.",
      "Exchange vs SP: H19's edge was mostly at BSP (ROI 19% against 6% at SP). Judge the result by A/E, not SP ROI."
    ],
    "mechanism": "A horse checked in its run and beaten 2 lengths or less last time would probably have won without the trouble. The handicapper rates the margin seen, not the run lost, so in a handicap next time the horse races off a mark a few pounds below its true level. In non-handicaps no rating is involved, and any edge comes only from the market missing the trouble. The handicap group should therefore show the bigger lift.",
    "market_miss": "The form figures show an ordinary narrow defeat, and narrow defeat with no trouble is priced correctly (A/E 0.979 on 12,205). The market gives no visible credit for the trouble, whether or not the price shortened (findings: no concentration by price movement). In handicaps the mark compounds this, because the handicapper also fails to charge for the hidden pounds.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.25 to 1.40 on roughly 120 to 200 bets, clearly above the non-handicap H19 runners",
      "falsified_if": "The handicap subset scores A/E below 1.10, or scores at or below the non-handicap H19 complement on comparable numbers. That would mean the mark adds nothing beyond the market's general failure to credit trouble."
    },
    "notes": "This rule sharpens H19 (1.18 on 242). The draft rule is kept as written: every column is pre-off and none reads the result. Its value depends on the comparison with the is_handicap = 0 control, not only on passing the floor. The subset may fall short of 150 bets. If it points the right way, it goes to falsification and forward tracking rather than being killed. Refinements are only examined if the broad rule shows evidence."
  }

Timeline

1. Job 1867 finished

   16 Oct, 14:55 · engine:compute-1

2. hypothesis note H-c93d63ff

   16 Oct, 14:55 · engine:perturb

3. perturbation result for H-c93d63ff (neighbour: drop: last_position > 1): pass

   16 Oct, 14:55 · engine:perturb

4. perturbation result for H-c93d63ff (neighbour: drop: last_momentum_interrupted = 1): fail

   16 Oct, 14:55 · engine:perturb

5. perturbation result for H-c93d63ff (neighbour: <= 2 -> 3 (+20%)): pass

   16 Oct, 14:55 · engine:perturb

6. perturbation result for H-c93d63ff (neighbour: <= 2 -> 1 (-20%)): pass

   16 Oct, 14:55 · engine:perturb

7. perturbation result for H-c93d63ff (neighbour: > 1 -> 2 (+20%)): pass

   16 Oct, 14:55 · engine:perturb

8. perturbation result for H-c93d63ff (neighbour: > 1 -> 0 (-20%)): pass

   16 Oct, 14:55 · engine:perturb

9. compute-1 took job 1867 (mutate, attempt 1)

   16 Oct, 14:55 · compute-1

10. Job 1875 finished

    16 Oct, 14:55 · engine:compute-1

11. H-c93d63ff Falsification → KILLED falsified: period 2025-10-01 to 2026-04-27 A/E 0.96; IRE: too few bets to judge

    16 Oct, 14:55 · engine:falsification

12. falsification result for H-c93d63ff (partition: Exclude nurseries (unstable juvenile marks)): pass

    16 Oct, 14:55 · engine:evaluator

13. compute-1 took job 1875 (attack, attempt 1)

    16 Oct, 14:55 · compute-1

14. Job 1874 finished

    16 Oct, 14:55 · engine:compute-1

15. falsification result for H-c93d63ff (price_band: Price cut: 14/1 or shorter, drop the 16/1+ band (1 winner at A/E 1.93)): pass

    16 Oct, 14:55 · engine:evaluator

16. compute-1 took job 1874 (attack, attempt 1)

    16 Oct, 14:55 · compute-1