Rival: chase fallers/unseaters NOT going well show the same lift

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
   • H85-fell-when-going-well (H85: fell when going well) was GREEN in the earlier research: A/E 1.13 on 405 bets (periods 1.15 / 1.11); now PROMISING. O-add18290
2. AskedDoes the chase lift need the horse to have been going well when it fell or unseated, or do all last-time chase fallers and unseaters beat their price? answered
3. MechanismIf punters simply over-penalise an F or U, the bounce-back would show in every chaser that failed to complete through a fall or unseat, whatever its position at the time.
4. Why the market misses itForm figures show F or U the same way whether the horse was going well or already beaten. So if the market overreacts to the letter itself, it overreacts to both groups.
5. Predictedpositive, A/E 1.00 to 1.05 on 400+ bets if the mechanism is specific (the rival fails). A/E 1.15+ would mean the rival holds.
6. Falsified ifThe rival is ruled out if A/E is at or below 1.05 with the same-race control near 1.0, which supports 'going well' as the active ingredient.
7. Rule sketchrace_type = 'Chase' AND last_fell_or_unseated = 1 AND last_fell_going_well = 0

Mechanism and rationale

If punters simply over-penalise an F or U, the bounce-back would show in every chaser that failed to complete through a fall or unseat, whatever its position at the time.

The rule

race_type = 'Chase' AND last_fell_or_unseated = 1 AND COALESCE(last_fell_going_well, 0) = 0

Reads: race_type, last_fell_or_unseated, last_fell_going_well

Pre-registration

Spec hash (sha256 of the canonical pre-spec)

f83ce1cef4e7771721145bd8453454b50460278a8bec7aab4619e8999dbc2d84

Frozen

17 Oct, 02:11

Nothing in the definition can change after this.

Confounders

• Horses that fell or unseated early never showed their ability, so their form is uninformative rather than over-penalised. That dilutes any lift and may mix in horses that jump poorly by habit.
• Price band: horses not going well when they fell are likely longer-priced next time, and edges die at 16/1+ (prior finding), so a pooled A/E can hide an effect that exists only at shorter prices.
• Field size: big handicap chases produce more F/U and more bystander unseats (brought down, hampered), and have different A/E baselines.
• Region: the chase-error edge so far is GB-led (IRE non-finishers 0.433 on 23), so a mix of regions could mask or manufacture a difference.
• Calendar: the data window starts on 1 Nov 2025 (last_fell_* availability), so this covers one season. Late-period skews seen elsewhere in this thread (after 27 April 2026) could drive the result.
• General bounce-back by well-beaten chasers (fancied, well-beaten no-error chasers 1.112 on 397) may lift any last-time disappointment, whether or not it was an F/U.
• last_fell_going_well may be NULL when not coded. COALESCE treats NULL as 'not going well', which could put unknown-status fallers into this group.

Discovery result

pass broad rule (niche) · 17 Oct, 02:09

Bets

270

29 winners

A/E

1.08

95%: 0.73 to 1.56

SP return

-41.6%

level stakes

Betfair SP return

-32.5%

270 bets with a BSP

Strike rate 10.7%. p = 0.36. Same-race control: the other runners in these races had A/E 0.99, so the selection beat its own races. Zone: discovery. Evaluator eval-1.0, data 6a4e98be18e83933

Falsification

Each attack is a deliberate attempt to kill the idea. 5 of 5 survived.

• partition: Explicit not-going-well only (drop NULL-coded fallers)pass

  270 bets · A/E 1.08 (0.73 to 1.56) · control 0.99 · 17 Oct, 02:11

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-271881.03

  after 2026-04-27821.24

• partition: F/U came over fences last timepass

  244 bets · A/E 1.19 (0.80 to 1.71) · control 0.98 · 17 Oct, 02:11

Sealed holdout (vault)

Not queried. Only a frozen idea can spend its one query against the 2020-25 holdout.

Forward record

• forward vs bar {'min_bets': 100, 'min_ae': 1.0, 'max_p': 0.2}info

  0 bets · · 17 Oct, 02:11

  Breakdown

  betsA/E0.71.01.3

  forward0–

Other experiments

• neighbour: drop: race_type = 'Chase'pass

  958 bets · A/E 1.09 (0.87 to 1.37) · control 0.99 · 17 Oct, 02:11

  Breakdown

  betsA/E0.71.01.3

  2025-10-01 to 2026-04-275291.06

  after 2026-04-274291.15

• neighbour: drop: last_fell_or_unseated = 1fail

  13,078 bets · A/E 1.00 (0.95 to 1.05) · control 1.30 · 17 Oct, 02:11

  Breakdown

Lineage

Agent runs

• designer v1ok · 59s · 17 Oct, 02:07

  Output (schema-checked)

  {
    "rule": "race_type = 'Chase' AND last_fell_or_unseated = 1 AND COALESCE(last_fell_going_well, 0) = 0",
    "refinements": [
      {
        "label": "GB only",
        "rule": "race_type = 'Chase' AND last_fell_or_unseated = 1 AND COALESCE(last_fell_going_well, 0) = 0 AND region = 'GB'"
      },
      {
        "label": "14/1 or shorter",
        "rule": "race_type = 'Chase' AND last_fell_or_unseated = 1 AND COALESCE(last_fell_going_well, 0) = 0 AND sp_decimal <= 15.0"
      },
      {
        "label": "F/U was over fences",
        "rule": "race_type = 'Chase' AND last_fell_or_unseated = 1 AND COALESCE(last_fell_going_well, 0) = 0 AND last_race_type = 'Chase'"
      }
    ],
    "control": {
      "description": "The lab's standard comparison is the other runners in the same chases. They should come out near A/E 1.0; if they are well above 1.0, the lift belongs to the races and not to the F/U. The comparison that settles the question is the fell-going-well chase group (the parent edge, A/E 1.297 on 165), measured over the same window. The rival holds only if this not-going-well group gets close to that figure. If this group sits near par while the going-well group is above it, 'going well' is what drives the edge.",
      "rule": "race_type = 'Chase' AND last_fell_or_unseated = 1 AND last_fell_going_well = 1"
    },
    "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": [
      "Horses that fell or unseated early never showed their ability, so their form is uninformative rather than over-penalised. That dilutes any lift and may mix in horses that jump poorly by habit.",
      "Price band: horses not going well when they fell are likely longer-priced next time, and edges die at 16/1+ (prior finding), so a pooled A/E can hide an effect that exists only at shorter prices.",
      "Field size: big handicap chases produce more F/U and more bystander unseats (brought down, hampered), and have different A/E baselines.",
      "Region: the chase-error edge so far is GB-led (IRE non-finishers 0.433 on 23), so a mix of regions could mask or manufacture a difference.",
      "Calendar: the data window starts on 1 Nov 2025 (last_fell_* availability), so this covers one season. Late-period skews seen elsewhere in this thread (after 27 April 2026) could drive the result.",
      "General bounce-back by well-beaten chasers (fancied, well-beaten no-error chasers 1.112 on 397) may lift any last-time disappointment, whether or not it was an F/U.",
      "last_fell_going_well may be NULL when not coded. COALESCE treats NULL as 'not going well', which could put unknown-status fallers into this group."
    ],
    "mechanism": "If punters simply over-penalise the F or U letter in the form figures, the bounce-back should show in every chaser whose last run ended in a fall or unseat, whatever its position at the time. If the lift exists only for horses that were going well, the edge comes from hidden ability: the horse was running well enough to win and the market cannot see that behind the letter. This test separates the two by measuring the F/U chasers that were NOT going well.",
    "market_miss": "Form figures show F or U the same way whether the horse was travelling strongly or already beaten. If the market reacts to the letter itself, it will under-price both groups equally. If it reads the race comments, only the going-well group, whose ability is hidden, will be mispriced.",
    "prediction": {
      "direction": "positive",
      "size": "Expected A/E 1.00–1.05 on roughly 300–500 bets if the edge is specific to going well (the rival fails). A/E 1.15 or more would mean the rival holds.",
      "falsified_if": "The rival is ruled out if A/E is at or below 1.05 while the same-race control is near 1.0 and the going-well control stays clearly above it. That would make 'going well' the active ingredient. The rival is supported only if A/E is 1.15 or more with p ≤ 0.05."
    },
    "notes": "This is a rival-explanation test. The expected outcome is a FAIL, and a fail is informative because it backs the specificity of SYS-fell-when-going-well-chase. The NULL handling for last_fell_going_well is the only change from the sketch. The window starts on 2025-11-01 (one season), so any pass needs forward confirmation across a second winter. Read the result at BSP as well as SP, because the parent edge paid only at BSP."
  }

Timeline

1. Job 2442 finished

   17 Oct, 02:11 · engine:compute-1

2. hypothesis note H-f2244c8b

   17 Oct, 02:11 · engine:perturb

3. perturbation result for H-f2244c8b (neighbour: drop: COALESCE(last_fell_going_well, 0) = 0): pass

   17 Oct, 02:11 · engine:perturb

4. perturbation result for H-f2244c8b (neighbour: drop: last_fell_or_unseated = 1): fail

   17 Oct, 02:11 · engine:perturb

5. perturbation result for H-f2244c8b (neighbour: drop: race_type = 'Chase'): pass

   17 Oct, 02:11 · engine:perturb

6. compute-1 took job 2442 (mutate, attempt 1)

   17 Oct, 02:11 · compute-1

7. Job 2441 finished

   17 Oct, 02:11 · engine:compute-1

8. forward result for H-f2244c8b (forward vs bar {'min_bets': 100, 'min_ae': 1.0, 'max_p': 0.2}): info

   17 Oct, 02:11 · engine:evaluator

9. compute-1 took job 2441 (forward_check, attempt 1)

   17 Oct, 02:11 · compute-1

10. Queued mutate on the compute lane: parameter stability: do neighbouring thresholds hold?

    17 Oct, 02:11 · engine:scheduler

11. Queued forward_check on the compute lane: daily forward check of a shadow system

    17 Oct, 02:11 · engine:scheduler

12. Job 2439 finished

    17 Oct, 02:11 · engine:compute-1

13. H-f2244c8b Frozen → SHADOW vault skipped: last_fell_or_unseated, last_fell_going_well do not exist before the discovery zone; track it forward instead

    17 Oct, 02:11 · engine:orchestrator

14. compute-1 took job 2439 (vault, attempt 1)

    17 Oct, 02:11 · compute-1

15. Job 2432 finished

    17 Oct, 02:11 · engine:compute-1

16. Queued vault on the compute lane: frozen: one sealed vault query

    17 Oct, 02:11 · engine:scheduler

17. H-f2244c8b Falsification → FROZEN survived falsification

    17 Oct, 02:11 · engine:falsification