Contrast: fell or unseated while NOT going well, at 14/1 or shorter, only matches its price

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-9f384f7c (Fell or unseated while going well: hidden form at normal prices) passed discovery as a niche: A/E 1.1 on 245 bets (p 0.291), same-race control 0.98. O-cae9b321
2. AskedDoes the 'going well' flag add information beyond a fall or unseat, i.e. do jumps runners that fell or unseated while not going well score clearly below the going-well group at 14/1 or shorter? answered
3. MechanismThe hidden-form idea says only a horse travelling well when it departed has form the result hides. A horse that fell or unseated while struggling showed nothing the price misses, and it may carry a jumping problem.
4. Why the market misses itThe form figures show F or U for both groups, so the market prices them alike. If the in-running detail matters, it separates them and the price does not.
5. Predictednegative, A/E 0.90 to 1.00 on 300+ bets, at least 0.07 below H-9f384f7c's 1.099
6. Falsified ifA/E 1.06 or more, or within 0.03 of the going-well group, which would mean the flag adds nothing
7. Rule sketchlast_fell_or_unseated = 1 AND last_fell_going_well = 0 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15

Mechanism and rationale

The hidden-form idea says only a horse travelling well when it departed has form the result hides. A horse that fell or unseated while struggling showed nothing the price misses, and it may carry a jumping problem.

The rule

last_fell_or_unseated = 1 AND last_fell_going_well = 0 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15

Reads: last_fell_or_unseated, last_fell_going_well, race_type, sp_decimal

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

9963199c715016500497e25f043beba2bae7d206e2e9212f40e521803b31af99

Frozen

not frozen

The definition can still change.

Confounders

• Jumping problem: some of the not-going-well group may have a real jumping fault, which lowers their score for a reason other than missing hidden form, so a low A/E does not by itself prove the going-well lift
• How well the going-well flag is coded: it comes from comment wording, so a vague or missing 'going well' phrase puts real going-well departures into this group and narrows the contrast
• GB and IRE mix: the going-well lift sat in GB (1.27) and not Ireland (0.80), so the contrast should be read by region in case flag coverage differs
• Mix of price bands: the going-well excess came from mid-short prices, so any contrast should compare like bands
• Short window: both flags start on 2025-11-01, so the sample covers about one season and the two groups share the same period noise
• F and UR mix: unseats are often down to the jockey or bad luck, falls more to the horse, and the two may be split differently across the groups

Control

{
  "description": "Two comparisons. (1) The lab's standard control, the other runners in the same races, should score about 0.98 to 1.00, as it did for H-9f384f7c (0.983). (2) The contrast this hypothesis exists for is the going-well group under the same filters (H-9f384f7c, 1.099 on 245). The flag counts as adding something only if this rule scores at least 0.07 below that group. If this rule lands within 0.03 of the going-well group, the in-running detail adds nothing.",
  "rule": "last_fell_going_well = 1 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15"
}

Discovery result

fail broad rule · 16 Oct, 13:34

Bets

367

58 winners

A/E

1.13

95%: 0.85 to 1.46

SP return

-16.6%

level stakes

Betfair SP return

-1.1%

367 bets with a BSP

Strike rate 15.8%. p = 0.20. 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

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 · 36s · 16 Oct, 13:32

  Output (schema-checked)

  {
    "rule": "last_fell_or_unseated = 1 AND last_fell_going_well = 0 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15",
    "refinements": [
      {
        "label": "fell (not unseated) while not going well: the jumping-problem end",
        "rule": "last_fell_or_unseated = 1 AND last_fell_going_well = 0 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15 AND last_non_finish_code = 'F'"
      },
      {
        "label": "chases only, where a jumping problem costs most",
        "rule": "last_fell_or_unseated = 1 AND last_fell_going_well = 0 AND race_type = 'Chase' AND sp_decimal <= 15"
      }
    ],
    "control": {
      "description": "Two comparisons. (1) The lab's standard control, the other runners in the same races, should score about 0.98 to 1.00, as it did for H-9f384f7c (0.983). (2) The contrast this hypothesis exists for is the going-well group under the same filters (H-9f384f7c, 1.099 on 245). The flag counts as adding something only if this rule scores at least 0.07 below that group. If this rule lands within 0.03 of the going-well group, the in-running detail adds nothing.",
      "rule": "last_fell_going_well = 1 AND race_type IN ('Hurdle', 'Chase') AND sp_decimal <= 15"
    },
    "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": [
      "Jumping problem: some of the not-going-well group may have a real jumping fault, which lowers their score for a reason other than missing hidden form, so a low A/E does not by itself prove the going-well lift",
      "How well the going-well flag is coded: it comes from comment wording, so a vague or missing 'going well' phrase puts real going-well departures into this group and narrows the contrast",
      "GB and IRE mix: the going-well lift sat in GB (1.27) and not Ireland (0.80), so the contrast should be read by region in case flag coverage differs",
      "Mix of price bands: the going-well excess came from mid-short prices, so any contrast should compare like bands",
      "Short window: both flags start on 2025-11-01, so the sample covers about one season and the two groups share the same period noise",
      "F and UR mix: unseats are often down to the jockey or bad luck, falls more to the horse, and the two may be split differently across the groups"
    ],
    "mechanism": "Hidden-form idea: a horse that fell or unseated while travelling well ran better than its F/U figure shows, so its next price is too long. A horse that fell or unseated while struggling or already beaten has no such hidden form. Its F/U hides nothing, and the departure may point to a real jumping weakness. So the lift should belong only to the going-well group, and this group should run at or below its price.",
    "market_miss": "The form figures show F or U for both groups, so a form-figure reader prices them alike. Only the in-running comment separates them. If the market does not use that comment, the going-well group is underpriced and this group is priced fairly or overbet. If this group scores as well as the going-well group, the market is not missing the comment detail. Any lift would then be a general F/U effect.",
    "prediction": {
      "direction": "negative",
      "size": "A/E 0.90 to 1.00 on 300+ bets, at least 0.07 below the going-well group's 1.099 under the same filters",
      "falsified_if": "A/E 1.06 or more, or within 0.03 of the going-well group (H-9f384f7c) under the same filters, which would mean the going-well flag adds nothing"
    },
    "notes": "This is a contrast test of H-9f384f7c, pre-registered as negative: the rule should not beat its price. The schema's pass_criteria and forward_bar are set at the lab floors and are written for positive edges. For this negative prediction, read them mirrored: the rule supports the mechanism only if A/E is at or below about 1.00 and at least 0.07 under the going-well control. The schema floors are not the success bar here. Both flag columns are pre-off and available from 2025-11-01 in GB and IRE. last_fell_going_well = 0 drops rows where the flag is NULL, which is intended. No result columns are used. The refinements come from the jumping-problem part of the reasoning (fallers, and chases) and are looked at only if the broad rule shows a clear effect."
  }

Timeline

1. Job 1623 finished

   16 Oct, 13:34 · engine:compute-1

2. H-9519a139 Discovery → KILLED fails: A/E 1.126, needs 0.95 or lower; p 0.8359 (lower tail), needs 0.05; periods [1.12, 1.14], each needs 0.99 or lower; A/E 1.126 against same-race control 0.98, needs 0.03 below

   16 Oct, 13:34 · engine:evaluator

3. discovery result for H-9519a139 (broad rule): fail

   16 Oct, 13:34 · engine:evaluator

4. H-9519a139 Designed → RESEARCHdiscovery measuring on the discovery zone

   16 Oct, 13:34 · engine:orchestrator

5. compute-1 took job 1623 (evaluate, attempt 1)

   16 Oct, 13:34 · compute-1

6. Job 1621 finished

   16 Oct, 13:33 · engine:ai-1

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

   16 Oct, 13:33 · engine:scheduler

8. H-9519a139 Proposed → RESEARCHdesigned pre-spec written and hashed

   16 Oct, 13:33 · agent:designer@v1

9. ai-1 took job 1621 (design, attempt 1)

   16 Oct, 13:32 · ai-1

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

    16 Oct, 13:32 · engine:scheduler

11. New idea H-9519a139: Contrast: fell or unseated while NOT going well, at 14/1 or shorter, only matches its price (test-mechanism)

    16 Oct, 13:32 · 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.