Rival: close-run chase edge is a winner bounce-back after any flop, not the jumping error

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-48da1bb8 (Test-mechanism: one-off chase error after a close run, where the mista) survived falsification: 6 of 6 attacks left it standing; it is frozen for its one vault query. O-e01f5b46
   • H-8443fa31 (Rival: close-earlier-run chase edge is class relief, tested on GB runn) passed discovery as a niche: A/E 1.41 on 120 bets (p 0.079), same-race control 0.96. O-e836eb22
2. AskedWithin the one-off chase-error close-run pool, is the lift carried by horses whose close run was a win (best_last3_beaten = 0), not a placed loss? testing
3. MechanismA chaser that won in one of its last three runs and was then well beaten or failed to finish may simply be bouncing back. The flop came from a rise in the weights, a bad day or the ground, and the horse returns to its winning level. If that is the cause, clean-jumping recent winners who flopped last time should beat their price as much as the one-off-error group does.
4. Why the market misses itThe latest figure (P/U or a 10L+ beating) dominates the price. The earlier win is visible, but the market may treat the flop as a sign of a new, higher handicap ceiling.
5. Predictedpositive, A/E 1.15 to 1.30 on 300+ bets if bounce-back alone carries the edge
6. Falsified ifA/E of 1.05 or lower on this no-error pool, while the one-off-error winner subset stays above 1.25. That would mean the jumping error is what misleads the market.

Mechanism and rationale

A chaser that won in one of its last three runs and was then well beaten or failed to finish may simply be bouncing back. The flop came from a rise in the weights, a bad day or the ground, and the horse returns to its winning level. If that is the cause, clean-jumping recent winners who flopped last time should beat their price as much as the one-off-error group does.

The rule

race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 0 AND (last_position IS NULL OR last_beaten_lengths > 10) AND best_last3_beaten IS NOT NULL AND best_last3_beaten = 0

Reads: race_type, remarked_last3, jumping_errors_last3, last_position, last_beaten_lengths, best_last3_beaten

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

80a8d26921df78155d2c4187b5cc8229eb995aeea8090260516226b5d8b180d6

Frozen

not frozen

The definition can still change.

Confounders

• Non-finish without a recorded jumping error (pulled up) often signals a physical or wind problem; the finishers-only refinement separates this from a genuine flop
• Handicap mark rise after the win mixes class with rating; the flop may simply reflect the new mark, so bounce-back is not guaranteed
• The 8/1 to 14/1 band has carried much of every chase split; check the price-band distribution before crediting the mechanism
• remarked_last3, jumping_errors_last3 and best_last3_beaten start 2025-11 (GB and IRE), so the sample is one short window with seasonal (winter jumps) concentration
• jumping_errors_last3 = 0 makes this disjoint from the frozen one-off-error chase subsets, but the best_last3_beaten condition is shared; the result is a rival test, not independent replication
• Earlier finding: close-run lift in the error pool may be concentrated in recent winners; if both pools lift equally, the shared factor is the recent win, not the error
• Last-time-out class change (drop after the flop) could produce bounce-back via class relief rather than form recovery

Discovery result

fail broad rule · 17 Oct, 01:17

Bets

298

37 winners

A/E

0.97

95%: 0.68 to 1.34

SP return

-14.7%

level stakes

Betfair SP return

+4.1%

296 bets with a BSP

Strike rate 12.4%. p = 0.59. Same-race control: the other runners in these races had A/E 1.01, 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 · 35s · 17 Oct, 01:16

  Output (schema-checked)

  {
    "rule": "race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 0 AND (last_position IS NULL OR last_beaten_lengths > 10) AND best_last3_beaten IS NOT NULL AND best_last3_beaten = 0",
    "refinements": [
      {
        "label": "finishers-only flop (excludes pulled-up / non-finish, the physical-problem confounder)",
        "rule": "race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 0 AND last_position IS NOT NULL AND last_beaten_lengths > 10 AND best_last3_beaten IS NOT NULL AND best_last3_beaten = 0"
      },
      {
        "label": "handicaps only (the mechanism blames a weight rise and a misread handicap ceiling)",
        "rule": "race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 0 AND (last_position IS NULL OR last_beaten_lengths > 10) AND best_last3_beaten IS NOT NULL AND best_last3_beaten = 0 AND is_handicap = 1"
      },
      {
        "label": "14/1 or shorter (chase-split edges have died at big prices)",
        "rule": "race_type = 'Chase' AND remarked_last3 = 3 AND jumping_errors_last3 = 0 AND (last_position IS NULL OR last_beaten_lengths > 10) AND best_last3_beaten IS NOT NULL AND best_last3_beaten = 0 AND sp_decimal <= 15"
      }
    ],
    "control": {
      "description": "Other runners in the same races should score about 1.0 (0.95 to 1.02). If they are also lifted, the effect is a property of those races (small chase fields, overround), not of bounce-back winners. The decisive comparison is the rival pool: recent winners who flopped last time with a one-off jumping error. The falsification compares the clean-jumping pool against that subset. If the error subset stays above 1.25 while this clean pool is at 1.05 or below, the jumping error is what misleads the market.",
      "rule": "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 best_last3_beaten IS NOT NULL AND best_last3_beaten = 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": [
      "Non-finish without a recorded jumping error (pulled up) often signals a physical or wind problem; the finishers-only refinement separates this from a genuine flop",
      "Handicap mark rise after the win mixes class with rating; the flop may simply reflect the new mark, so bounce-back is not guaranteed",
      "The 8/1 to 14/1 band has carried much of every chase split; check the price-band distribution before crediting the mechanism",
      "remarked_last3, jumping_errors_last3 and best_last3_beaten start 2025-11 (GB and IRE), so the sample is one short window with seasonal (winter jumps) concentration",
      "jumping_errors_last3 = 0 makes this disjoint from the frozen one-off-error chase subsets, but the best_last3_beaten condition is shared; the result is a rival test, not independent replication",
      "Earlier finding: close-run lift in the error pool may be concentrated in recent winners; if both pools lift equally, the shared factor is the recent win, not the error",
      "Last-time-out class change (drop after the flop) could produce bounce-back via class relief rather than form recovery"
    ],
    "mechanism": "A chaser that won one of its previous three runs, then was beaten 10L+ or failed to finish last time without any recorded jumping error, returns to its winning level. The flop was caused by a weight rise, ground or a bad day, not a lasting decline. If the close-run chase edge is really a winner bounce-back, this clean-jumping pool should beat its price about as much as the one-off-error winner group does.",
    "market_miss": "The latest figure (P/U or a 10L+ beating) dominates the price. The earlier win is visible, but the market reads the flop as proof that the horse has hit a new, higher handicap ceiling and discounts a return to winning form.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.15 to 1.30 if bounce-back alone carries the edge (300+ bets hoped for; the short 2025-11 window may give fewer)",
      "falsified_if": "A/E of 1.05 or lower on this clean-jumping pool, while the one-off-error winner subset (control rule) stays above 1.25. That would mean the jumping error, not the bounce-back, is what the market misreads. A/E below 1.0 falsifies outright."
    },
    "notes": "Rule sketch kept unchanged: all columns are pre_off, there is no result leakage, and the slicing is minimal. best_last3_beaten = 0 together with a last run beaten 10L+ or not finished means the win was two or three runs back. All columns share the 2025-11-01 start. The refinements come straight from the reasoning's confounders: pulled-up non-finishers, the weight-rise mechanism, and big prices."
  }

Timeline

1. Job 2198 finished

   17 Oct, 01:17 · engine:compute-1

2. H-31ffebf7 Discovery → KILLED fails: A/E 0.971, needs 1.05; p 0.5939, needs 0.05 or lower; periods [1.03, 0.88], each needs 1.0; A/E 0.971 against same-race control 1.011, needs 0.03 above (not a niche pass: A/E 0.971 below 1.05; p 0.5938587852183083 above 0.5; does not beat the same-race control 1.011; a period points the other way [1.03, 0.88])

   17 Oct, 01:17 · engine:evaluator

3. discovery result for H-31ffebf7 (broad rule): fail

   17 Oct, 01:17 · engine:evaluator

4. H-31ffebf7 Designed → RESEARCHdiscovery measuring on the discovery zone

   17 Oct, 01:17 · engine:orchestrator

5. compute-1 took job 2198 (evaluate, attempt 1)

   17 Oct, 01:17 · compute-1

6. Job 2197 finished

   17 Oct, 01:17 · engine:ai-1

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

   17 Oct, 01:17 · engine:scheduler

8. H-31ffebf7 Proposed → RESEARCHdesigned pre-spec written and hashed

   17 Oct, 01:17 · agent:designer@v1

9. ai-1 took job 2197 (design, attempt 1)

   17 Oct, 01:16 · ai-1

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

    17 Oct, 01:16 · engine:scheduler

11. New idea H-31ffebf7: Rival: close-run chase edge is a winner bounce-back after any flop, not the jumping error (rival-explanation)

    17 Oct, 01:16 · 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.