Generalise: checked mid-run and beaten 2 to 5 lengths still beats its price, by less

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 trouble edge shrink as the beaten margin grows (0-2 lengths strongest, 2-5 lengths weaker but still positive), as a dose-response would require? answered
3. MechanismBeing checked while making a move costs a horse roughly 1 to 3 lengths. A horse checked and beaten 2 to 5 lengths may have finished closer than the margin shows, but it has less hidden ability than one beaten 2 lengths or less.
4. Why the market misses itThe form book records the margin and the position. The checking appears only in the comment, which the price ignored even when the price shortened (H-48a2eca6 and the shortened half scored about the same).
5. Predictedpositive, A/E 1.05 to 1.15 on 300+ bets, below H19's 1.28
6. Falsified ifA/E at or below 1.0 on 300+ bets, which would mean the 2-length cut is a cliff rather than a dose-response
7. Rule sketchlast_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths > 2 AND last_beaten_lengths <= 5

Mechanism and rationale

Being checked while making a move costs a horse roughly 1 to 3 lengths. A horse checked and beaten 2 to 5 lengths may have finished closer than the margin shows, but it has less hidden ability than one beaten 2 lengths or less.

The rule

last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths > 2 AND last_beaten_lengths <= 5

Reads: last_momentum_interrupted, last_position, last_beaten_lengths

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

87b514b899cfd13ca8aae8dad18896efea6f9fa784f1457dcedceaa3b7546025

Frozen

not frozen

The definition can still change.

Confounders

• Horses beaten further may have been eased once the trouble ended their chance. That would change the margin's meaning (a refinement tests last_eased = 0).
• The mix of Flat and jumps runs may differ from H19's. Checking in Flat fields may cost a different number of lengths than in jumps races.
• After a beaten run, horses may drop in class or get handicap relief. The lift could come from that rather than from the hidden trouble.
• The edge may sit only at big prices or only in one price band, as happened in other families. The 14/1 refinement and the band split check this.
• The momentum-interrupted flag comes from comment wording, and the wording may change with the race reader or the period.

Control

{
  "description": "The other runners in the same races should score about A/E 1.0. The main control is the same margin band with no trouble: horses beaten more than 2 and up to 5 lengths whose momentum was not interrupted. We expect them to be priced correctly (about 0.97 to 1.01), as narrow defeats without trouble were (0.979 on 12,205). If they also beat their prices, the lift comes from the margin band and not from the checking.",
  "rule": "last_momentum_interrupted = 0 AND last_position > 1 AND last_beaten_lengths > 2 AND last_beaten_lengths <= 5"
}

Discovery result

fail broad rule · 16 Oct, 14:49

Bets

263

26 winners

A/E

0.80

95%: 0.52 to 1.18

SP return

-56.0%

level stakes

Betfair SP return

-48.6%

261 bets with a BSP

Strike rate 9.9%. p = 0.89. Same-race control: the other runners in these races had A/E 1.03, 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 · 25s · 16 Oct, 14:47

  Output (schema-checked)

  {
    "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths > 2 AND last_beaten_lengths <= 5",
    "refinements": [
      {
        "label": "not eased after the trouble",
        "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths > 2 AND last_beaten_lengths <= 5 AND last_eased = 0"
      },
      {
        "label": "14/1 or shorter today",
        "rule": "last_momentum_interrupted = 1 AND last_position > 1 AND last_beaten_lengths > 2 AND last_beaten_lengths <= 5 AND sp_decimal <= 15"
      }
    ],
    "control": {
      "description": "The other runners in the same races should score about A/E 1.0. The main control is the same margin band with no trouble: horses beaten more than 2 and up to 5 lengths whose momentum was not interrupted. We expect them to be priced correctly (about 0.97 to 1.01), as narrow defeats without trouble were (0.979 on 12,205). If they also beat their prices, the lift comes from the margin band and not from the checking.",
      "rule": "last_momentum_interrupted = 0 AND last_position > 1 AND last_beaten_lengths > 2 AND last_beaten_lengths <= 5"
    },
    "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": [
      "Horses beaten further may have been eased once the trouble ended their chance. That would change the margin's meaning (a refinement tests last_eased = 0).",
      "The mix of Flat and jumps runs may differ from H19's. Checking in Flat fields may cost a different number of lengths than in jumps races.",
      "After a beaten run, horses may drop in class or get handicap relief. The lift could come from that rather than from the hidden trouble.",
      "The edge may sit only at big prices or only in one price band, as happened in other families. The 14/1 refinement and the band split check this.",
      "The momentum-interrupted flag comes from comment wording, and the wording may change with the race reader or the period."
    ],
    "mechanism": "Being checked while making a move costs a horse about 1 to 3 lengths. A horse checked and beaten 2 to 5 lengths may have had the ability to finish within about 0 to 4 lengths, so the margin overstates how far it was beaten. It has less hidden ability than a horse checked and beaten 2 lengths or less (H19, A/E about 1.18 to 1.28). The edge should therefore be smaller but still positive, a dose-response rather than a cliff at 2 lengths.",
    "market_miss": "The form figures and the price record the margin and the finishing position. The checking appears only in the in-running comment. H19's split showed the market gives no credit for it: runners whose price did not shorten (H-48a2eca6, 1.24) did about as well as those whose price shortened. A beaten-2-to-5-lengths run looks like ordinary placed form, so the price takes the margin at face value.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.05 to 1.15 on 300+ bets, below H19's 1.18 to 1.28",
      "falsified_if": "A/E at or below 1.0 on 300+ bets, which would mean the 2-length cut is a cliff rather than a dose-response. The claim also fails if the no-trouble control in the same 2-to-5-length band scores as high as the rule."
    },
    "notes": "I kept the draft rule unchanged. It uses only pre-off columns and contains no post-race fields. last_position > 1 drops non-finishers, whose last_position is NULL, which is intended because they have no beaten margin. The predicted size sits close to the 1.05 discovery floor, so a thin result pointing the right way should go to falsification and forward tracking rather than be killed for its size. Read refinements only if the broad rule shows evidence."
  }

Timeline

1. Job 1842 finished

   16 Oct, 14:49 · engine:compute-1

2. H-4882f53a Discovery → KILLED fails: A/E 0.802, needs 1.05; p 0.8913, needs 0.05 or lower; periods [0.67, 0.86], each needs 1.0; A/E 0.802 against same-race control 1.03, needs 0.03 above (not a niche pass: A/E 0.802 below 1.05; p 0.8913184689610928 above 0.5; does not beat the same-race control 1.03; a period points the other way [0.67, 0.86])

   16 Oct, 14:49 · engine:evaluator

3. discovery result for H-4882f53a (broad rule): fail

   16 Oct, 14:49 · engine:evaluator

4. H-4882f53a Designed → RESEARCHdiscovery measuring on the discovery zone

   16 Oct, 14:49 · engine:orchestrator

5. compute-1 took job 1842 (evaluate, attempt 1)

   16 Oct, 14:49 · compute-1

6. Job 1839 finished

   16 Oct, 14:48 · engine:ai-1

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

   16 Oct, 14:48 · engine:scheduler

8. H-4882f53a Proposed → RESEARCHdesigned pre-spec written and hashed

   16 Oct, 14:48 · agent:designer@v1

9. ai-1 took job 1839 (design, attempt 1)

   16 Oct, 14:47 · ai-1

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

    16 Oct, 14:47 · engine:scheduler

11. New idea H-4882f53a: Generalise: checked mid-run and beaten 2 to 5 lengths still beats its price, by less (generalise)

    16 Oct, 14:47 · 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.