Rival: big-rise quick-returning handicap winners stepping up in class are a real obstacle, not an over-fade

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-21f9e6e7 (Long-priced raised handicap winners turned out within 14 days are over) pointed the predicted way but failed the bar: A/E 0.57 on 255 bets (p 0.967), control 1.03. O-7c9eecf8
   • H-d9dabea2 (Rival: big-rise quick-returning handicap winners beat their price with) survived falsification: 5 of 6 attacks left it standing; it is frozen for its one vault query. O-8f406178
2. AskedIs the big-rise lift strongest when the rise also forces a step up in class, which doubles the visible deterrent? answered
3. MechanismA big rise often lifts a horse out of its class band. Facing better-handicapped, more progressive rivals is a real obstacle, unlike the rise alone, and would explain why the long-priced end of the raised group fails.
4. Why the market misses itIf the fade-the-visible reading is right, the market fades a double deterrent too hard and these runners beat their price. The rival is that the class rise is priced correctly or under-priced, because punters still back a recent winner.
5. Predictednegative, A/E 0.85 to 1.00 on 100+ GB bets, well below the 1.2+ of the frozen parents
6. Falsified ifA/E at or above 1.15 on 100+ bets, which would favour the over-fade reading
7. Rule sketch

Mechanism and rationale

A big rise often lifts a horse out of its class band. Facing better-handicapped, more progressive rivals is a real obstacle, unlike the rise alone, and would explain why the long-priced end of the raised group fails.

The rule

last_position = 1 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change >= 6 AND region = 'GB' AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class < last_race_class

Reads: last_position, days_since_last_run, is_handicap, last_is_handicap, rating_change, region, race_class, last_race_class

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

6149a671ddb696c9a02ef83b39bd3722847f8621d2bc04a23c91dfff12a205e1

Frozen

not frozen

The definition can still change.

Confounders

• Price band: class step-ups drift to longer odds, and the raised group already fails at 8/1 or bigger (0 winners from 32 at 16/1 or bigger), so a low A/E may be the long-price effect again. The 7/1-or-shorter refinement checks this.
• race_type mix: Flat and jumps class structures differ, and Flat nurseries and handicaps step up more often.
• Field size is larger in better classes, which lowers the base win rate. A/E adjusts for price but not fully for competitiveness.
• Thin sample: rating_change exists only from 2025-11-01, so with the GB and step-up filters the group may fall short of 150 bets.
• Overlap with the frozen parents (H-dd3724ed, H-d9dabea2, H-528222df): this is a subset of them, not independent evidence.
• Easy-win mix: the step-up group may hold a different share of easy winners than the parent, and easy winners carry their own lift.

Control

{
  "description": "Same-race other runners (lab default): expected near 1.0 (roughly 0.9 to 1.05), so a low A/E for the selection is not a property of these better-class races. The decisive comparison is the same big-rise quick-returning GB handicap winners who stay in the same class or drop in class. They should hold the parents' lift (A/E about 1.15 or higher). If the step-up group scores at or below 1.0 and the non-step-up group scores 1.15 or higher, the class step-up is the obstacle. If both score alike, class has nothing to do with it.",
  "rule": "last_position = 1 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change >= 6 AND region = 'GB' AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class >= last_race_class"
}

Discovery result

fail broad rule · 16 Oct, 16:36

Bets

138

31 winners

A/E

1.16

95%: 0.79 to 1.65

SP return

-12.6%

level stakes

Betfair SP return

+0.5%

138 bets with a BSP

Strike rate 22.5%. p = 0.22. Same-race control: the other runners in these races had A/E 0.97, 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 · 45s · 16 Oct, 16:35

  Output (schema-checked)

  {
    "rule": "last_position = 1 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change >= 6 AND region = 'GB' AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class < last_race_class",
    "refinements": [
      {
        "label": "7/1 or shorter",
        "rule": "last_position = 1 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change >= 6 AND region = 'GB' AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class < last_race_class AND sp_decimal IS NOT NULL AND sp_decimal <= 8.0"
      },
      {
        "label": "one-class step only",
        "rule": "last_position = 1 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change >= 6 AND region = 'GB' AND race_class IS NOT NULL AND last_race_class IS NOT NULL AND race_class = last_race_class - 1"
      }
    ],
    "control": {
      "description": "Same-race other runners (lab default): expected near 1.0 (roughly 0.9 to 1.05), so a low A/E for the selection is not a property of these better-class races. The decisive comparison is the same big-rise quick-returning GB handicap winners who stay in the same class or drop in class. They should hold the parents' lift (A/E about 1.15 or higher). If the step-up group scores at or below 1.0 and the non-step-up group scores 1.15 or higher, the class step-up is the obstacle. If both score alike, class has nothing to do with it.",
      "rule": "last_position = 1 AND days_since_last_run IS NOT NULL AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change >= 6 AND region = 'GB' 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": [
      "Price band: class step-ups drift to longer odds, and the raised group already fails at 8/1 or bigger (0 winners from 32 at 16/1 or bigger), so a low A/E may be the long-price effect again. The 7/1-or-shorter refinement checks this.",
      "race_type mix: Flat and jumps class structures differ, and Flat nurseries and handicaps step up more often.",
      "Field size is larger in better classes, which lowers the base win rate. A/E adjusts for price but not fully for competitiveness.",
      "Thin sample: rating_change exists only from 2025-11-01, so with the GB and step-up filters the group may fall short of 150 bets.",
      "Overlap with the frozen parents (H-dd3724ed, H-d9dabea2, H-528222df): this is a subset of them, not independent evidence.",
      "Easy-win mix: the step-up group may hold a different share of easy winners than the parent, and easy winners carry their own lift."
    ],
    "mechanism": "A rise of 6 lb or more often lifts a handicap winner out of its class band. In the higher class it meets better-handicapped, more progressive rivals, many of them also improving. That is a real obstacle, unlike the rise alone, where the horse stays among the same rivals and its current form outruns the handicapper. If so, the step-up subset loses the lift the frozen big-rise parents show, and the step-up may explain the poor long-priced end of the raised group.",
    "market_miss": "The rival to the over-fade reading. If the market fades a double deterrent (a big rise plus a class rise) too hard, these runners beat their price. The rival view is that the class step-up is priced correctly or even under-priced, because punters keep backing a recent, quick-returning winner and do not mark it down enough for the stronger opposition. The selection would then score at or below its price.",
    "prediction": {
      "direction": "negative",
      "size": "A/E 0.85 to 1.00 on 100+ GB bets, well below the 1.2+ of the frozen big-rise parents, with the same-class/class-drop control at about 1.15 or higher",
      "falsified_if": "A/E at or above 1.15 on 100+ bets, which would favour the over-fade reading; or step-up and non-step-up groups scoring within 0.1 of each other, which would mean class does not explain the gap"
    },
    "notes": "Changes from the sketch: added IS NOT NULL guards on days_since_last_run, rating_change, race_class and last_race_class, matching the frozen parents. GB only is kept for two reasons: race_class is available for IRE only from 2026-07-01, and the hypothesis concerns GB. rating_change starts 2025-11-01, so the sample covers about a year and may fall short of 150. The prediction is negative, so the vault judges over-betting (A/E below 1). The pass_criteria are the lab floors, read in the pre-registered negative direction. Refinements: (1) 7/1 or shorter separates the class effect from the known long-price failure. (2) A one-class step tests whether the obstacle needs a big jump in class."
  }

Timeline

1. Job 2041 finished

   16 Oct, 16:36 · engine:compute-1

2. H-7dbfdd08 Discovery → KILLED fails: 138 bets, needs 300; A/E 1.162, needs 0.95 or lower; p 0.8263 (lower tail), needs 0.05; periods [1.09, 1.24], each needs 0.99 or lower; A/E 1.162 against same-race control 0.971, needs 0.03 below

   16 Oct, 16:36 · engine:evaluator

3. discovery result for H-7dbfdd08 (broad rule): fail

   16 Oct, 16:36 · engine:evaluator

4. H-7dbfdd08 Designed → RESEARCHdiscovery measuring on the discovery zone

   16 Oct, 16:36 · engine:orchestrator

5. compute-1 took job 2041 (evaluate, attempt 1)

   16 Oct, 16:36 · compute-1

6. Job 2039 finished

   16 Oct, 16:36 · engine:ai-1

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

   16 Oct, 16:36 · engine:scheduler

8. H-7dbfdd08 Proposed → RESEARCHdesigned pre-spec written and hashed

   16 Oct, 16:36 · agent:designer@v1

9. ai-1 took job 2039 (design, attempt 1)

   16 Oct, 16:35 · ai-1

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

    16 Oct, 16:35 · engine:scheduler

11. New idea H-7dbfdd08: Rival: big-rise quick-returning handicap winners stepping up in class are a real obstacle, not an over-fade (rival-explanation)

    16 Oct, 16:35 · 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.