Sharpen: easy handicap winners back within 14 days with no rise in the official rating yet

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-5c714267 (Test-mechanism: easy-winner quick-return lift lives in handicap-to-han) passed discovery: A/E 1.19 on 478 bets (p 0.019), same-race control 0.92. O-dd247331
2. AskedWithin the handicap-to-handicap fortnight rule, is the lift larger when the official rating has not yet gone up (rating_change <= 0), meaning the horse runs under a fixed penalty rather than a reassessed mark? answered
3. MechanismA fixed penalty is the same whatever the manner of victory, so for an easy winner it is the bluntest possible adjustment. Before reassessment the horse is most under-burdened.
4. Why the market misses itThe racecard shows the same rating plus a standard penalty, which looks like a fair rise. Nothing on it shows how much the horse had left.
5. Predictedpositive, A/E 1.20 to 1.35 on 150+ bets
6. Falsified ifA/E at or below the parent's 1.188, or below 1 in either period.
7. Rule sketchlast_position = 1 AND last_won_easily = 1 AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change <= 0
8. question idQ-5b6717c9
9. operatorsharpen

The question's answer so far: No, and the opposite looks true. Runners with no rating rise scored 1.096 on 247, below the parent's 1.188 on 478. That implies about 1.28 for raised or missing-rating runners, so the fixed-penalty mechanism is not where the lift lives.

Mechanism and rationale

A fixed penalty is the same whatever the manner of victory, so for an easy winner it is the bluntest possible adjustment. Before reassessment the horse is most under-burdened.

The rule

last_position = 1 AND last_won_easily = 1 AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change <= 0

Reads: last_position, last_won_easily, days_since_last_run, is_handicap, last_is_handicap, rating_change

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

14c61aa5f1b86f47f4439be744b9ddae485bead35e7baab3aff384fff2eddf03

Frozen

not frozen

The definition can still change.

Confounders

• rating_change, last_won_easily and last_official_rating all start 2025-11, so the whole parent (478) lives in this short window; the subset is a slice of an already thin sample and the two periods are short
• Reassessment timing differs: GB ratings are revised weekly and take effect days later; IRE practice differs, so rating_change <= 0 means different things by region
• Flat vs jumps penalty sizes and reassessment lags differ
• Quickest returners (under about a week) are almost all unreassessed, so the split may be days_since_last_run in disguise
• rating_change < 0 (mark cut despite winning) is rare and may be data quirks; NULL rating_change is excluded explicitly
• Long-priced runners added nothing to the parent (0 from 24 at 16/1+), so price mix can drive the result

Control

{
  "description": "The same-race other runners should sit near A/E 1.0 (the parent's control was 0.92), showing the lift is not a property of these races. The key comparison is the reassessed sibling inside the parent: easy handicap winners back within 14 days whose official rating has already gone up. If the fixed-penalty mechanism is right, they should sit clearly below the rule (near or under 1.0); if they match or beat it, the rating_change split adds nothing.",
  "rule": "last_position = 1 AND last_won_easily = 1 AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change > 0"
}

Discovery result

fail broad rule · 16 Oct, 16:19

Bets

247

89 winners

A/E

1.10

95%: 0.88 to 1.35

SP return

+0.7%

level stakes

Betfair SP return

+14.5%

247 bets with a BSP

Strike rate 36.0%. p = 0.21. Same-race control: the other runners in these races had A/E 0.95, 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 · 41s · 16 Oct, 16:18

  Output (schema-checked)

  {
    "rule": "last_position = 1 AND last_won_easily = 1 AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change <= 0",
    "refinements": [
      {
        "label": "Shorter than 16/1",
        "rule": "last_position = 1 AND last_won_easily = 1 AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change <= 0 AND sp_decimal < 17"
      },
      {
        "label": "Flat only (penalty structure most uniform)",
        "rule": "last_position = 1 AND last_won_easily = 1 AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change <= 0 AND race_type = 'Flat'"
      }
    ],
    "control": {
      "description": "The same-race other runners should sit near A/E 1.0 (the parent's control was 0.92), showing the lift is not a property of these races. The key comparison is the reassessed sibling inside the parent: easy handicap winners back within 14 days whose official rating has already gone up. If the fixed-penalty mechanism is right, they should sit clearly below the rule (near or under 1.0); if they match or beat it, the rating_change split adds nothing.",
      "rule": "last_position = 1 AND last_won_easily = 1 AND days_since_last_run <= 14 AND is_handicap = 1 AND last_is_handicap = 1 AND rating_change IS NOT NULL AND rating_change > 0"
    },
    "pass_criteria": {
      "min_bets": 150,
      "min_ae": 1.1,
      "max_p": 0.05
    },
    "forward_bar": {
      "min_bets": 100,
      "min_ae": 1,
      "max_p": 0.2
    },
    "confounders": [
      "rating_change, last_won_easily and last_official_rating all start 2025-11, so the whole parent (478) lives in this short window; the subset is a slice of an already thin sample and the two periods are short",
      "Reassessment timing differs: GB ratings are revised weekly and take effect days later; IRE practice differs, so rating_change <= 0 means different things by region",
      "Flat vs jumps penalty sizes and reassessment lags differ",
      "Quickest returners (under about a week) are almost all unreassessed, so the split may be days_since_last_run in disguise",
      "rating_change < 0 (mark cut despite winning) is rare and may be data quirks; NULL rating_change is excluded explicitly",
      "Long-priced runners added nothing to the parent (0 from 24 at 16/1+), so price mix can drive the result"
    ],
    "mechanism": "A penalty for a recent win is a fixed weight, whatever the manner of victory. For a horse that won easily with more in hand, a fixed penalty is the bluntest adjustment. Until the handicapper reassesses and raises the official rating, the horse is at its most under-burdened, so the parent's quick-return lift (A/E 1.188 on 478) should concentrate in runners whose rating has not yet gone up.",
    "market_miss": "The racecard shows the old rating plus a standard penalty, which looks like a fair rise. Nothing on it shows how much the horse had in hand, so punters and layers price the penalty as if it captured the improvement.",
    "prediction": {
      "direction": "positive",
      "size": "A/E 1.20 to 1.35 on 150+ bets, above the reassessed sibling (rating_change > 0)",
      "falsified_if": "A/E at or below the parent's 1.188, or below 1.0 in either period, or the reassessed control (rating_change > 0) matching or beating the rule."
    },
    "notes": "Draft rule kept, with rating_change IS NOT NULL made explicit. All three easy-win and rating columns are pre_off from 2025-11-01 in GB and IRE. Refinements follow the reasoning and earlier findings: the long end added nothing to the parent, and Flat penalties are the most uniform. If the subset is under 150 it goes to falsification and forward tracking as thin, not killed. min_ae is set above the floor at 1.10 because the hypothesis has to beat a parent of 1.188."
  }

Timeline

1. Job 1968 finished

   16 Oct, 16:19 · engine:compute-1

2. H-97b1373c Discovery → KILLED fails: A/E 1.096, needs 1.1; p 0.2082, needs 0.05 or lower (not a niche pass: A/E 1.096 below 1.1)

   16 Oct, 16:19 · engine:evaluator

3. discovery result for H-97b1373c (broad rule): fail

   16 Oct, 16:19 · engine:evaluator

4. H-97b1373c Designed → RESEARCHdiscovery measuring on the discovery zone

   16 Oct, 16:19 · engine:orchestrator

5. compute-1 took job 1968 (evaluate, attempt 1)

   16 Oct, 16:19 · compute-1

6. Job 1966 finished

   16 Oct, 16:18 · engine:ai-1

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

   16 Oct, 16:18 · engine:scheduler

8. H-97b1373c Proposed → RESEARCHdesigned pre-spec written and hashed

   16 Oct, 16:18 · agent:designer@v1

9. ai-1 took job 1966 (design, attempt 1)

   16 Oct, 16:18 · ai-1

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

    16 Oct, 16:17 · engine:scheduler

11. New idea H-97b1373c: Sharpen: easy handicap winners back within 14 days with no rise in the official rating yet (sharpen)

    16 Oct, 16:17 · 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.