Test-mechanism: GB one-off chase-error horses dropped in class are fairly priced

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-db7f42b8 (Rival: GB no-drop chase lift is a GB-region effect, not class choice) passed discovery as a niche: A/E 1.24 on 347 bets (p 0.073), same-race control 0.97. O-19c9c833
   • H-c2928b4e (Generalise: one-off chase error, well beaten, not dropping in class (G) survived falsification: 5 of 5 attacks left it standing; it is frozen for its one vault query. O-2f35030a
2. AskedWithin the GB one-off chase-error pool, do class droppers score at or below 1.0, so that the class choice, not the region, separates the pool? answered
3. MechanismA trainer who drops the horse in class after the error is saying the horse is in decline, not that it had a blip. The hidden ability the pool relies on is not there.
4. Why the market misses itThe market bets the class drop as relief and does not price the trainer's admission, so these runners should not beat their price. If they do, region or the error itself is doing the work, not the class choice.
5. Predictednegative, A/E 0.80 to 1.00 on about 90 bets
6. Falsified ifA/E of 1.15 or higher on droppers, which would mean class choice does not separate the pool
7. Rule sketch

Mechanism and rationale

A trainer who drops the horse in class after the error is saying the horse is in decline, not that it had a blip. The hidden ability the pool relies on is not there.

The rule

region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class > last_race_class

Reads: region, race_type, remarked_last3, jumping_errors_last3, last_jumping_error, last_position, last_beaten_lengths, race_class, last_race_class

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

acc428196da23b4f21adaa87dfc8352f58f0110ffeb3c05f0cf8480a14905b88

Frozen

not frozen

The definition can still change.

Confounders

• Small sample: about 90 droppers expected, so the 150-bet floor cannot be reached and the result is descriptive only
• Handicap mark drops come with class drops, so a lower mark could offset any decline signal
• Region confound: race_class is GB-only before July 2026, so the comparison is kept within GB
• Novice versus handicap mix differs between droppers and non-droppers
• Period effect: lifts in this thread have faded after April 2026
• Overlap: this is a subset of an already-tested pool, so it is not independent evidence

Control

{
  "description": "Main comparison is the complementary GB one-off chase-error pool: same conditions but running at the same class or higher (race_class <= last_race_class), which scored A/E 1.36 on 253. The hypothesis is that droppers sit at or below 1.0 while non-droppers stay well above it, so class choice separates the pool. The usual same-race control (other runners in the droppers' races) should show about 0.95 to 1.0. If the droppers score close to that and the complement stays near 1.3, the trainer-signal mechanism is supported.",
  "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class <= last_race_class"
}

Designer

Discovery result

fail broad rule · 17 Oct, 01:26

Bets

83

14 winners

A/E

1.01

95%: 0.56 to 1.70

SP return

-19.4%

level stakes

Betfair SP return

-8.3%

82 bets with a BSP

Strike rate 16.9%. p = 0.51. Same-race control: the other runners in these races had A/E 1.00, so the selection 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 · 31s · 17 Oct, 01:25

  Output (schema-checked)

  {
    "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class > last_race_class",
    "refinements": [
      {
        "label": "droppers 14/1 or shorter",
        "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class > last_race_class AND sp_decimal <= 15"
      },
      {
        "label": "handicap droppers only (class drop with a mark drop)",
        "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class > last_race_class AND is_handicap = 1"
      }
    ],
    "control": {
      "description": "Main comparison is the complementary GB one-off chase-error pool: same conditions but running at the same class or higher (race_class <= last_race_class), which scored A/E 1.36 on 253. The hypothesis is that droppers sit at or below 1.0 while non-droppers stay well above it, so class choice separates the pool. The usual same-race control (other runners in the droppers' races) should show about 0.95 to 1.0. If the droppers score close to that and the complement stays near 1.3, the trainer-signal mechanism is supported.",
      "rule": "region = 'GB' AND race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 1 AND last_jumping_error = 1 AND (last_position IS NULL OR last_beaten_lengths > 10) AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class <= last_race_class"
    },
    "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": [
      "Small sample: about 90 droppers expected, so the 150-bet floor cannot be reached and the result is descriptive only",
      "Handicap mark drops come with class drops, so a lower mark could offset any decline signal",
      "Region confound: race_class is GB-only before July 2026, so the comparison is kept within GB",
      "Novice versus handicap mix differs between droppers and non-droppers",
      "Period effect: lifts in this thread have faded after April 2026",
      "Overlap: this is a subset of an already-tested pool, so it is not independent evidence"
    ],
    "mechanism": "In the one-off chase-error pool, the edge depends on hidden ability that the error masked. When a trainer drops the horse in class after the error, that is a stable-side admission of decline, not a blip. So the hidden ability is missing, and these runners should lose the pool's lift.",
    "market_miss": "The market reads the class drop as easier opposition and prices the horse up for it. It does not discount the trainer's implicit admission of decline. So droppers should at best match their price and should not share the pool's lift. If they beat their price, the error or region drives the edge, not the class choice.",
    "prediction": {
      "direction": "negative",
      "size": "A/E 0.80 to 1.00 on about 90 bets, clearly below the 1.36 of the non-dropper complement",
      "falsified_if": "Droppers score A/E 1.15 or higher, or come within about 0.15 of the non-dropper complement. Either means class choice does not separate the pool."
    },
    "notes": "This is a test-mechanism split, not a candidate edge. Expected size (about 90) is below the 150-bet floor, so discovery cannot formally pass. Read it as a contrast against the control rule. An earlier subtraction estimate put GB droppers at about 0.9, and close-run droppers scored 0.997 on 40. race_class > last_race_class means a drop, because class 1 is the top. Refinements are only for diagnosis: the price cap checks whether any lift is longshot noise, and the handicap-only split isolates the mark-drop confound."
  }

Timeline

1. Job 2231 finished

   17 Oct, 01:26 · engine:compute-1

2. H-3f87586e Discovery → KILLED fails: 83 bets, needs 300; A/E 1.015, needs 0.95 or lower; p 0.5924 (lower tail), needs 0.05; periods [1.05], each needs 0.99 or lower; A/E 1.015 against same-race control 1.002, needs 0.03 below

   17 Oct, 01:26 · engine:evaluator

3. discovery result for H-3f87586e (broad rule): fail

   17 Oct, 01:26 · engine:evaluator

4. H-3f87586e Designed → RESEARCHdiscovery measuring on the discovery zone

   17 Oct, 01:26 · engine:orchestrator

5. compute-1 took job 2231 (evaluate, attempt 1)

   17 Oct, 01:26 · compute-1

6. Job 2229 finished

   17 Oct, 01:26 · engine:ai-1

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

   17 Oct, 01:26 · engine:scheduler

8. H-3f87586e Proposed → RESEARCHdesigned pre-spec written and hashed

   17 Oct, 01:26 · agent:designer@v1

9. ai-1 took job 2229 (design, attempt 1)

   17 Oct, 01:25 · ai-1

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

    17 Oct, 01:24 · engine:scheduler

11. New idea H-3f87586e: Test-mechanism: GB one-off chase-error horses dropped in class are fairly priced (test-mechanism)

    17 Oct, 01:24 · 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.