Test-mechanism: chase fallers that were not going well are at or below par

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-f8d6b9d7 (Generalise: all last-time chase unseaters beat their price, going well) passed discovery: A/E 1.4 on 213 bets (p 0.048), same-race control 0.96. O-4c8868a6
   • H-2ac5dd09 (Test-mechanism: unseated (not fell) chasers not going well carry the g) survived falsification: 5 of 5 attacks left it standing; it is frozen for its one vault query. O-8ba5d41c
   • H-dcbc54fb (Rival: the going-well chase edge is carried by actual fallers, not uns) passed discovery as a niche: A/E 1.26 on 85 bets (p 0.223), same-race control 0.96. O-da3d9ee2
2. AskedDo chase fallers proper (F) that were not going well sit at or below par, completing the picture that a plain fall carries no bounce-back? answered
3. MechanismA fall while the horse is under pressure or struggling usually reflects tiredness or poor jumping, which are real weaknesses that repeat. A fall while the horse is travelling well hides a good run. An unseat says little either way.
4. Why the market misses itThe market treats every F alike. If anything it should under-punish tired fallers relative to the hidden good runs, so this group is fairly priced or over-bet.
5. Predictednegative, A/E 0.80 to 0.97 on 70+ bets
6. Falsified ifA/E at or above 1.10, which would mean fallers bounce back whether or not they were going well.

Mechanism and rationale

A fall while the horse is under pressure or struggling usually reflects tiredness or poor jumping, which are real weaknesses that repeat. A fall while the horse is travelling well hides a good run. An unseat says little either way.

The rule

race_type = 'Chase' AND last_non_finish_code = 'F' AND COALESCE(last_fell_going_well, 0) = 0 AND race_date >= '2025-11-01'

Reads: race_type, last_non_finish_code, last_fell_going_well, race_date

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

a20b3b4f8af904bf65d8d3d417e55597bab27bfbaa5ee701b79b66fed434da39

Frozen

not frozen

The definition can still change.

Confounders

• Small sample: about 80 bets are implied (270 non-going-well chase F/U minus 190 unseaters), so the result will be thin and a few long-price winners will swing it
• last_fell_going_well is inconsistently flagged for falls at the last fence while still in contention: hidden good runs leak into the test group and bias it upward
• Coverage: last_fell_going_well only exists from 2025-11-01, and COALESCE would put every earlier faller into the 'not going well' group, hence the date bound
• Confidence or physical damage after a heavy fall, which shows up in today's run and not in jumping ability
• Long layoffs after injury: the fallers that return quickly are a selected, sounder subset
• Today's price: tired fallers may go off at long prices where the F/U long-shot profit sits (0 from 98 at 16/1+ in the F/U not-going-well cut)

Control

{
  "description": "Same-race other runners should sit at about A/E 0.96 to 1.00, as in the sibling chase F/U cuts. Those cuts got 0.96 to 0.99, so any shortfall below that is the faller effect and not the race. The comparison this question needs is the going-well chase fallers proper over the same window. They scored 1.26 on 85 before. If the going-well flag matters for fallers, the control should sit clearly above the test group. If both groups come out at about the same level, the flag adds nothing for fallers.",
  "rule": "race_type = 'Chase' AND last_non_finish_code = 'F' AND last_fell_going_well = 1 AND race_date >= '2025-11-01'"
}

Discovery result

fail broad rule · 17 Oct, 03:02

Bets

167

20 winners

A/E

1.04

95%: 0.63 to 1.60

SP return

-38.4%

level stakes

Betfair SP return

-24.5%

167 bets with a BSP

Strike rate 12.0%. p = 0.47. Same-race control: the other runners in these races had A/E 1.01, 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 · 42s · 17 Oct, 03:01

  Output (schema-checked)

  {
    "rule": "race_type = 'Chase' AND last_non_finish_code = 'F' AND COALESCE(last_fell_going_well, 0) = 0 AND race_date >= '2025-11-01'",
    "refinements": [
      {
        "label": "Fell under pressure: the tired-faller reading itself",
        "rule": "race_type = 'Chase' AND last_non_finish_code = 'F' AND COALESCE(last_fell_going_well, 0) = 0 AND race_date >= '2025-11-01' AND last_under_pressure = 1"
      },
      {
        "label": "Under 16/1: strip the long-price tail where the F/U profit sits",
        "rule": "race_type = 'Chase' AND last_non_finish_code = 'F' AND COALESCE(last_fell_going_well, 0) = 0 AND race_date >= '2025-11-01' AND sp_decimal < 17"
      },
      {
        "label": "Fall came in a chase: like-for-like jumping test",
        "rule": "race_type = 'Chase' AND last_non_finish_code = 'F' AND COALESCE(last_fell_going_well, 0) = 0 AND race_date >= '2025-11-01' AND last_race_type = 'Chase'"
      }
    ],
    "control": {
      "description": "Same-race other runners should sit at about A/E 0.96 to 1.00, as in the sibling chase F/U cuts. Those cuts got 0.96 to 0.99, so any shortfall below that is the faller effect and not the race. The comparison this question needs is the going-well chase fallers proper over the same window. They scored 1.26 on 85 before. If the going-well flag matters for fallers, the control should sit clearly above the test group. If both groups come out at about the same level, the flag adds nothing for fallers.",
      "rule": "race_type = 'Chase' AND last_non_finish_code = 'F' AND last_fell_going_well = 1 AND race_date >= '2025-11-01'"
    },
    "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 80 bets are implied (270 non-going-well chase F/U minus 190 unseaters), so the result will be thin and a few long-price winners will swing it",
      "last_fell_going_well is inconsistently flagged for falls at the last fence while still in contention: hidden good runs leak into the test group and bias it upward",
      "Coverage: last_fell_going_well only exists from 2025-11-01, and COALESCE would put every earlier faller into the 'not going well' group, hence the date bound",
      "Confidence or physical damage after a heavy fall, which shows up in today's run and not in jumping ability",
      "Long layoffs after injury: the fallers that return quickly are a selected, sounder subset",
      "Today's price: tired fallers may go off at long prices where the F/U long-shot profit sits (0 from 98 at 16/1+ in the F/U not-going-well cut)"
    ],
    "mechanism": "A chase fall while under pressure or struggling usually means tiredness, limited ability or poor jumping under fatigue. These weaknesses repeat next time. A fall while travelling well hides a good run that the form figures do not show, which explains the going-well faller lift (1.26 on 85). An unseat says little about the horse either way, so the unseater lift (1.26 to 1.40) comes from somewhere else. Prediction: once the going-well cases are removed, plain fallers carry no bounce-back. That would make the going-well flag what separates fallers, and the F/U lift an unseat effect plus a going-well effect.",
    "market_miss": "The market reads every F in the form figures alike and discounts it by roughly the same amount. If anything this flat discount under-punishes the tired fallers, whose fall followed a genuinely poor run, compared with the going-well fallers who hid a good one. So this group should be fairly priced or over-bet, not under-bet.",
    "prediction": {
      "direction": "negative",
      "size": "A/E 0.80 to 0.97 on roughly 70 to 100 bets, below the going-well faller control and near or below same-race other runners",
      "falsified_if": "A/E at or above 1.10, which would mean chase fallers bounce back whether or not they were going well. Also falsified if the test group matches or beats the going-well faller control, which would mean the flag does not separate fallers."
    },
    "notes": "Change from the sketch: I added race_date >= '2025-11-01'. last_fell_going_well only exists from 2025-11-01, while last_non_finish_code runs from 2025-01-01. Without the bound, COALESCE(...,0)=0 would quietly class every faller from Jan to Oct 2025 as 'not going well' and mix in the hidden good runs. The predicate is the same shape as the earlier sibling cuts, so the results stay comparable. The direction is negative, so the pass floors (150 / 1.05 / 0.05) are to be applied in the pre-registered direction, i.e. A/E at or below about 0.95. The expected sample is under 150, so this will probably be a thin result sent on to falsification and forward tracking. Refinement 2 uses sp_decimal < 17, i.e. shorter than 16/1. The question is answered mainly by the contrast with the going-well control, not by profit at the prices."
  }

Timeline

1. Job 2555 finished

   17 Oct, 03:02 · engine:compute-1

2. H-a50e3885 Discovery → KILLED fails: 167 bets, needs 300; A/E 1.036, needs 0.95 or lower; p 0.6207 (lower tail), needs 0.05; periods [0.81, 1.58], each needs 0.99 or lower; A/E 1.036 against same-race control 1.008, needs 0.03 below

   17 Oct, 03:02 · engine:evaluator

3. discovery result for H-a50e3885 (broad rule): fail

   17 Oct, 03:02 · engine:evaluator

4. H-a50e3885 Designed → RESEARCHdiscovery measuring on the discovery zone

   17 Oct, 03:02 · engine:orchestrator

5. compute-1 took job 2555 (evaluate, attempt 1)

   17 Oct, 03:02 · compute-1

6. Job 2553 finished

   17 Oct, 03:02 · engine:ai-1

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

   17 Oct, 03:02 · engine:scheduler

8. H-a50e3885 Proposed → RESEARCHdesigned pre-spec written and hashed

   17 Oct, 03:02 · agent:designer@v1

9. ai-1 took job 2553 (design, attempt 1)

   17 Oct, 03:01 · ai-1

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

    17 Oct, 03:00 · engine:scheduler

11. New idea H-a50e3885: Test-mechanism: chase fallers that were not going well are at or below par (test-mechanism)

    17 Oct, 03:00 · 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.