- loop:started
- {"at":"2026-10-16T10:59:37+00:00","compute_workers":1,"kernel":"gallops.kernel","pid":7}
- digest:2026-10-15
- {"agent_runs_today":{},"at":"2026-10-15T22:21:02+00:00","by_state":{"KILLED":246,"LIVE":39,"PROMISING":129,"SHADOW":20,"VAULT_FAILED":14},"capacity":{"consecutive_failures":0,"failure_type":null,"jobs_completed":0,"last_failure":null,"last_success":null,"next_retry_at":null,"note":"initialised","provider":"claude","state":"AVAILABLE"},"date":"2026-10-15","jobs_today":{"ai:leased":1,"ai:queued":1,"compute:leased":1,"compute:queued":4},"tests_today":889,"transitions_today":{"FROZEN":34,"KILLED":246,"PROMISING":129,"SHADOW":20,"VAULT":29,"VAULT_FAILED":14}}
- digest:latest
- {"agent_runs_today":{},"at":"2026-10-16T10:02:38+00:00","by_state":{"KILLED":253,"LIVE":39,"PROMISING":129,"SHADOW":21,"VAULT_FAILED":14},"capacity":{"consecutive_failures":0,"failure_type":null,"jobs_completed":12,"last_failure":null,"last_success":"2026-10-15T22:30:17+00:00","next_retry_at":null,"note":"initialised","provider":"claude","state":"AVAILABLE"},"date":"2026-10-16","jobs_today":{"compute:leased":1,"compute:queued":2},"tests_today":1,"transitions_today":{}}
- digest:last
- "2026-10-16"
- forward:IQ131-generic-trainer-development
- {"ae":0,"at":"2026-10-15T22:21:02+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:IQ50-costly-jumping-out-of-character
- {"ae":0,"at":"2026-10-15T22:21:02+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:IQ51-costly-jumping-in-character
- {"ae":0,"at":"2026-10-15T22:21:02+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:IQ52-lost-most-ground-jumping-last-time
- {"ae":0,"at":"2026-10-15T22:21:02+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:IQ54-failed-to-finish-jumping-as-usual
- {"ae":0,"at":"2026-10-15T22:21:02+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:IQ56-failed-to-finish-good-jumper-normal-day
- {"ae":0,"at":"2026-10-15T22:21:02+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:IQ57-failed-to-finish-poor-jumper-normal-day
- {"ae":0,"at":"2026-10-15T22:21:02+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:SYS2-pulled-up-after-jumping-well
- {"ae":0,"at":"2026-10-15T22:21:03+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:SYS3-pulled-up-jumping-as-usual
- {"ae":0,"at":"2026-10-15T22:21:03+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:SYS4-costly-jumping-figures-half
- {"ae":0,"at":"2026-10-15T22:21:03+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-22ef2116
- {"ae":0,"at":"2026-10-15T22:24:56+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- librarian:last_count
- 149
- settings:effective
- {"active_hours":{"enabled":false,"end":"08:00","start":"00:00","tz":"Europe/London"},"ai_allowed":true,"ai_enabled":true,"in_active_hours":true,"max_agent_runs_per_day":300,"scout_interval_s":1800,"source":"defaults","vault_budget_monthly":40}
- vault_budget_monthly
- 40
- scan:log
- {"count":3,"date":"2026-10-16","last_at":"2026-10-16T16:05:46+00:00"}
- digest:2026-10-16
- {"agent_runs_today":{},"at":"2026-10-16T10:02:38+00:00","by_state":{"KILLED":253,"LIVE":39,"PROMISING":129,"SHADOW":21,"VAULT_FAILED":14},"capacity":{"consecutive_failures":0,"failure_type":null,"jobs_completed":12,"last_failure":null,"last_success":"2026-10-15T22:30:17+00:00","next_retry_at":null,"note":"initialised","provider":"claude","state":"AVAILABLE"},"date":"2026-10-16","jobs_today":{"compute:leased":1,"compute:queued":2},"tests_today":1,"transitions_today":{}}
- bandit:deepened
- {"H126-lightly-raced-four-year-old-returns":"2026-10-16T11:08:48+00:00","H17-jamie-snowden-novices":"2026-10-16T14:14:25+00:00","H181-proven-over-further-at-a-stiff-track":"2026-10-16T11:40:37+00:00","H19-momentum-interrupted-beaten-narrowly":"2026-10-16T13:44:22+00:00","H199-favourite-at-a-big-price-in-a-big-handicap":"2026-10-16T14:44:27+00:00","H34-won-easily-last-time":"2026-10-16T15:14:31+00:00","H98-brought-down-or-unseated":"2026-10-16T12:13:13+00:00","IQ140a-wide-trip-with-figures":"2026-10-16T10:13:21+00:00"}
- bandit:last
- {"action":"scout","explore":false,"family":"negative","reason":"exploration: a new thread. Thompson draw 0.957 (reward 7.00 over 7 hypotheses)","thread_explore":true}
- reflection:H45-young-horses-in-novice-chases
- {"agent_run_id":17,"at":"2026-10-16T10:05:52+00:00","observations":["O-4dc10e5b"],"reflection":"The thread has not yet shown that a young-chaser edge exists under the lab's yardstick. The imported H45 rule (age 5 or under in non-handicap novice chases) was GREEN at 1.095 on 435 bets. Re-measured here it gives A/E 0.963 on 204. The imported lift also came only from the later periods: the first period was 1.00 on 254 bets.\n\nTwo follow-up tests failed:\n- **Handicap chases (H-ce840402):** young horses came out at 0.967 on 521 bets. Periods were 1.01 and 0.88, and the same-race control was 1.017. GB was 0.973 and Ireland 0.933. This is not a real opposite-direction discovery, because p for the opposite direction is only 0.41. It just says youth alone, over fences, is priced right.\n- **Chase debut (H-fd37b556):** only 11 bets, so it says nothing either way. The other runners in those races came out at 1.139, though. That is a warning sign.\n\nTogether, these point away from \"the young horse is underrated\". The rival worth testing first is that non-handicap novice chases as a race type beat the field-size-corrected yardstick, and young horses only share that. The 16/1+ band in the handicap test (1.18 on 96 bets, 3 wins) is the familiar longshot artefact and is not evidence of anything. The age 4 versus age 5 split should not be run yet: slicing a rule that does not reproduce would break the method rules."}
- scan:lenses
- {"q:Q-6c05e634":"2026-10-16T10:05:12+00:00"}
- reflection:IQ140a-wide-trip-with-figures
- {"agent_run_id":95,"at":"2026-10-16T11:59:27+00:00","observations":["O-8e95e697","O-a6496079","O-ea9be132","O-ebd1cc5d","O-6d7a0835"],"reflection":"The thread has moved from a physical story to a market-reading story. The lost-ground tests had already failed: small fields scored 1.39 against 1.23 for big fields, and galloping tracks 1.23 against 1.18 for sharp ones. The idea that came out of that failure was that the wide tag leads the market to misread a close-beaten run. Its first direct test, H-c869bcd0, passed and showed a dose-response. Wide, close-beaten runners at 15.0 SP or shorter whose price drifted more than 20% since last time scored 1.47 on 260 bets (p 0.009, control 0.94). The same base where the price shortened scored 1.165 on 614. Any drift scored 1.38, and a drift of 40% or more scored 1.58 on 202. Removing class rises made it stronger (1.59 on 181, GB only). Removing the small-to-big-field step left it at 1.51. Fancied drifters at 7/1 or shorter scored 1.60. It survived 6 of 6 attacks and is frozen. Two warnings. First, on the same going group the drift subset falls to 1.245 on 120 (p 0.21), with periods 0.96 before April 2026 and 1.69 after. So much of the lift sits in drifters whose ground has changed, which is a rival story: the market may over-penalise any going switch. Second, the post-April lift shows again in most drift splits, though the shortened complement is flat (1.17/1.16). The jumps-today test (H-77a1708f) scored 1.27 on 130 (p 0.11) and was killed on 36 Irish bets (0.91). The all-weather-today test was killed the same way on 37 Irish bets (0.90). Yet the Irish slice of the drift rule is 1.90 on 34. Kills on about 35 Irish bets look like low-power noise in both directions, not evidence of a GB-only edge."}
- reflection:H126-lightly-raced-four-year-old-returns
- {"agent_run_id":190,"at":"2026-10-16T13:15:12+00:00","observations":["O-95f9a9aa","O-6d4bfc11","O-65b889c1"],"reflection":"The chain has narrowed the H126 lift step by step. First it moved from four-year-olds to all-age spring handicap returners (1.09 on 1424). Then it moved to all-weather spring returners (1.195 on 480, killed on IRE 0.449), and then to turf-to-AW returners after 150+ days in any month. That last rule, H-98ae1db9, scored 1.138 on 1213 at discovery with both periods above 1.1 and survived 5 of 5 attacks. It then failed the sealed vault outright: 0.979 on 3155, p 0.64. This is the second time the thread has shrunk badly from discovery to vault. H-5167b372 went from 1.09 to 1.03 on 6934, which counted as supportive but was far from significant.\n\nThe rival H-f370f2f4 (turf-to-AW switchers not dropping in class) was meant to be fairly priced but scored 1.08 on 3797. On its face that would say the surface switch itself is mispriced. However, it was measured on the same discovery window that produced the 1.138 the vault has just contradicted, and its CI (0.98 to 1.20) includes 1.0. The 16/1+ band scored 0.97, so it is not even the longshot pattern we saw before.\n\nWhat we now believe: every turf-to-AW refinement in this thread was found and confirmed inside one window. The only out-of-sample test said there is no edge. So the surface-switch story should be treated as unproven, probably a discovery-window artefact. It should not be sliced further. What remains of the thread is the published won-last-time four-year-old system and the shadow spring rules, which are waiting on forward data."}
- forward:H-17cdbcfa
- {"ae":0,"at":"2026-10-16T11:12:36+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-baf747c3
- {"ae":0,"at":"2026-10-16T11:12:36+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-463f02e9
- {"ae":0,"at":"2026-10-16T11:12:36+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-c8ec6922
- {"ae":0,"at":"2026-10-16T11:12:36+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-29eb96fe
- {"ae":0,"at":"2026-10-16T11:22:14+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-4e7cc5f4
- {"ae":0,"at":"2026-10-16T11:29:35+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-568b7703
- {"ae":0,"at":"2026-10-16T11:29:35+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-b05a33c3
- {"ae":0,"at":"2026-10-16T11:29:35+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-1eed6021
- {"ae":0,"at":"2026-10-16T11:35:25+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-edbf6921
- {"ae":0,"at":"2026-10-16T11:35:25+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- reflection:H181-proven-over-further-at-a-stiff-track
- {"agent_run_id":126,"at":"2026-10-16T12:24:02+00:00","observations":["O-e1a41b2a","O-934e6d95","O-b9d384ce","O-51ffda33"],"reflection":"The thread now has a consistent core, and its size keeps coming down. The original H181 stiff-track idea (1.068 on 7574) narrowed to the stiff Flat sprint, at 1.506 on 303. Capping at 14/1 and widening to any proven stamina (last run or last win 330+ yards further) gave about 1.54 in discovery. On sealed vault data it scored 1.17 on 415 for the last-run version (p 0.12) and now 1.25 on 506 for the any-stamina version. That second figure has p 0.03 against a threshold of 0.0027, after surviving 5 of 5 attacks; it was graded supportive, not validated. Two vault reads both land at about 1.2. That points to a real but modest edge, roughly 1.15 to 1.25, not the 1.5 discovery suggested. The two new near misses are not evidence either way. The dose-response test (660+ yard drop) had 2 bets (5.05) and the Irish-flag rival had 4 bets (2.77, all GB). Both pointed the predicted way, but at that size they are noise. Every rival that would separate the hill from a course quirk, a draw quirk, Ireland or the size of the drop now runs out of bets inside the stiff-sprint population. Discovery cannot slice this any further. The right next step is to let the forward record judge the frozen rules, not to write more slices of the same few hundred runners."}
- forward:H-5ef5773d
- {"ae":0,"at":"2026-10-16T11:43:53+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-503a093f
- {"ae":0,"at":"2026-10-16T11:43:53+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-dc80c356
- {"ae":0,"at":"2026-10-16T11:51:13+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-1a6f83c9
- {"ae":0,"at":"2026-10-16T11:51:13+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-cde96a9f
- {"ae":0,"at":"2026-10-16T11:51:13+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-c869bcd0
- {"ae":0,"at":"2026-10-16T11:58:15+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-4e7be756
- {"ae":0,"at":"2026-10-16T11:58:15+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- reflection:H98-brought-down-or-unseated
- {"agent_run_id":160,"at":"2026-10-16T12:49:28+00:00","observations":["O-eb56a99e","O-7f64a209"],"reflection":"This thread keeps testing the same thing: does a fall, unseat or brought-down last time hide a horse's ability at normal prices? Each test has come back no. Capped at 14/1, BD/UR runners scored 1.021 on 312. Going-well departures scored 1.099 and were killed. The not-going-well group scored 1.126, so the comment flag adds nothing. The weak-form sharpen was killed. Now fallers at 14/1 or shorter (1.07 on 389) have been killed too, because the whole lift sits in the up-to-3/1 band and the rest score 0.92 (O-7f64a209). The opposite idea, that 8/1 to 14/1 departures are over-bet, pointed the right way but proves nothing: A/E 0.82 on 280, CI 0.47 to 1.34, p 0.81, 16 wins against 19.5 expected, control 1.02 (O-eb56a99e). The earlier band figures that inspired it (fallers 0.395 on 148, quick-return 0.579 on 76) were segments picked for looking extreme. When pooled they moved back toward 1. Three killed departure rules leaned on their shortest price band (up to 3/1 or up to 7/1). These rules share most of their runners, so that is one pattern counted three times, not three confirmations. The simplest explanation is the ordinary favourite end of the jumps market in this window, not anything about departures. At normal prices, then, last-time departures look fairly priced. The only live interest is the quick-return BD/UR shadow rule (1.167 on 355), which still has no forward bets and gets its lift from 16/1+ runners. That is the only rule in the thread worth time until it has a forward record."}
- forward:H-2d115a58
- {"ae":0,"at":"2026-10-16T12:16:07+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-2b483ac5
- {"ae":0,"at":"2026-10-16T12:16:07+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-cbd60237
- {"ae":0,"at":"2026-10-16T12:23:21+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-da5bbd88
- {"ae":0,"at":"2026-10-16T12:23:21+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- reflection:H175-young-horse-in-a-marathon-chase
- {"agent_run_id":206,"at":"2026-10-16T13:41:47+00:00","observations":["O-99a444ce"],"reflection":"The thread now looks like a narrow and fragile 3m4f+ effect, not a general rule about youth and stamina. The broad marathon rule (H175) held up on handicaps at 1.18 on 584 bets. After that, every attempt to explain or widen it failed:\n- The weight-position rivals failed in the vault: 0.98 on 1378 bets and 0.90 on 1129.\n- Four positive slices were killed because the up-to-7/1 band alone carried them.\n- The hurdle version failed: 0.945 on 772 bets.\n- Widening the trip to 3m1f+ failed: 0.899 on 266 bets.\n\nThe only survivor is the slice of young horses with 6 or fewer chase runs at 3m4f+. It was supportive in the vault (1.20 on 290 bets, p 0.136) but not significant.\n\nThe new test asked whether the same profile flips sign below the 3m4f line. Young, lightly raced chasers at 3m to 3m3f scored 0.93 on 743 bets, against a same-race control of 1.06. The direction is right, but the interval (0.76 to 1.13) includes par and p is 0.78. The only striking part is the 16/1-or-bigger band: 0 wins from 107 bets. That looks like the ordinary weakness of longshots, not something specific to young stayers.\n\nSo below 3m4f the profile is roughly fairly priced or slightly over-bet. That fits the earlier finding that the effect stops sharply at a trip line rather than fading. It adds no evidence that one mechanism switches sign at 3m4f. I propose no new hypothesis. The right next step is to let the forward record judge the 3m4f+ slice of young chasers with few chase runs, rather than slicing further."}
- forward:H-5167b372
- {"ae":0,"at":"2026-10-16T12:51:34+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-98848406
- {"ae":0,"at":"2026-10-16T12:51:34+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-aa66431f
- {"ae":0,"at":"2026-10-16T13:39:25+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- reflection:H19-momentum-interrupted-beaten-narrowly
- {"agent_run_id":225,"at":"2026-10-16T13:56:14+00:00","observations":["O-874ca10d","O-af566701"],"reflection":"The parent rule (checked mid-run, beaten 2 lengths or less, Flat) still stands at A/E 1.28 on 218 bets. Both rivals have failed on large samples: narrow losers with no trouble scored 0.979 on 12,205 bets, and late closers with no trouble scored 0.969 on 5,540. So the trouble flag carries the lift. It does not stretch, though. Checked runners beaten 2 to 5 lengths scored 0.80 on 263 bets, and over jumps the rule found only 22 bets. Every sharpening dies the same way.\n\n- **Checked and recovered:** 1.59 on 31 bets, killed by 1 of 1 attacks; too thin to judge.\n- **Price did not shorten after the trouble:** 1.236 on 104 bets, broken by 2 of 5 attacks.\n- **Handicaps only:** 1.145 on 190 bets, killed by its October to April period at 0.96, after holding against all 6 attacks.\n\nThe 6-for-6 attack record makes the handicap death the most informative failure. It points to a seasonal pattern rather than a structural flaw: the lift sits in the turf summer, 1.29 after late April 2026. That is a reason for one test chosen on the mechanism (trouble on turf against trouble on all-weather), not for more slicing. The within-35-days version (1.19 on 200) is frozen and still has no forward bets. From here, the forward record of the live system will judge this thread better than further cuts of a 218-bet sample."}
- forward:H-b595a70a
- {"ae":0,"at":"2026-10-16T13:52:55+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- reflection:H17-jamie-snowden-novices
- {"agent_run_id":233,"at":"2026-10-16T14:19:40+00:00","observations":["O-5d3cd419"],"reflection":"The thread now has one live system (Snowden novices in non-handicaps: A/E 1.172 on 282 across the full history) and four children that all fail for the same reason: too few bets. The newest, bumpers, scored 1.113 on 35 and survived its one attack. It was still killed: bets up to 14/1 carried it, the rest scored 0.95, and no two periods had enough bets to compare. Read together, the children say two things. First, the one child with a usable sample (non-novice jumps, 0.975 on 187, periods 1.04 and 0.84, same-race control 1.01) says the lift is not yard-wide over obstacles: it is tied to the novice stage. Second, every slice of the novice system or its neighbours (16, 27 and 35 bets) points positive but cannot be judged in the October 2025+ window. Bumpers fit a young-horse story but are about as thin as the price and experience slices. Cutting the system finer cannot settle anything here. The forward record of the live system, taken at BSP (+8.3% ROI there, -6.7% at SP), is now the real test. More slicing would be slicing a near miss."}
- reflection:H199-favourite-at-a-big-price-in-a-big-handicap
- {"agent_run_id":253,"at":"2026-10-16T15:01:39+00:00","observations":["O-91bc292c","O-a2471f48","O-1bd420a0","O-5f33166a","O-c42ebd58","O-8a077ff1"],"reflection":"The chain keeps telling one story. Each time a front-of-market big-handicap pool is re-measured on new data or with a fixed definition, it shrinks towards par.\n\n- **Big-yard top-two (H-659f3a37):** 1.33 on 386 in discovery fell to 1.14 on 1,003 in the sealed vault (p 0.075, threshold 0.0022). It survived 5 of 6 attacks first, so it is not an artefact, but the effect is smaller than discovery said.\n- **Fixed, time-stable big-yard definition (H-2ae0cf0d):** 1.13 on 196 (p 0.30). All of the lift sits in 2025-10-01 to 2026-04-27 (1.33); after that it scores 0.92 on 91.\n- **Yard-quality rival (H-0939d0ac):** 1.10 on 162. GB scores 1.335 and Ireland 0.64, so yard quality does not explain the Irish par either.\n\nWith the second-favourite rival's vault result (1.158 down to 1.04 on 4,207), this points to a real but small lift, about 1.04 to 1.14. It sits in GB. It is concentrated in the winter half of the data and fades in the summer half. Yard size and yard quality are not the active ingredient.\n\nThe early/late split matches the jumps/flat calendar almost exactly. So the \"fade\" may be a split by code rather than a market adapting, and that has not been tested. Nothing measured here settles the open questions on the live H199 system's own full-history or GB-only behaviour, so I am closing none of them."}
- forward:H-601a8566
- {"ae":0,"at":"2026-10-16T14:49:50+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-659f3a37
- {"ae":0,"at":"2026-10-16T15:00:40+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- reflection:H34-won-easily-last-time
- {"agent_run_id":286,"at":"2026-10-16T15:39:29+00:00","observations":["O-a22c28c6"],"reflection":"The thread's core still stands as a shadow result. Easy handicap winners back within 14 days scored 1.188 on 478. Within them, the lift grows with the rise: raised 1.334 on 231, and 6 lb+ 1.47 on 131. Big-rise winners without an easy win scored 1.134 on 318, also in shadow. Every version past 14 days failed. The side branch asks why long-priced raised quick returners do so badly: 0.575 on 255, killed only on bet count. That branch has now tested its second 'visible obstacle'. The class step-up rival came out the wrong way (1.16 on 138). The trip step-up of 15% or more, predicted negative, came out at 0.67 on 50 bets. That points the predicted way, but p is 0.879, the CI runs 0.25 to 1.47, and the same-race control was 1.07. Most importantly, only 5 of the 50 were 16/1 or bigger, so a trip step-up cannot be what sits behind the 0-from-32 longshot record. So far the obstacle explanations have not held: neither class nor trip accounts for the longshot failures. The longshot weakness looks like a price-band effect within this pool, not a hidden form obstacle. Without a lot more data it cannot be told apart from noise, and slicing it further would just chase a near miss. The productive work in this thread is now the forward record of the shadow rules (H-dd3724ed, H-d9dabea2, H-ed1a0f17), not more longshot autopsies."}
- forward:H-5c714267
- {"ae":0,"at":"2026-10-16T15:19:48+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-528222df
- {"ae":0,"at":"2026-10-16T15:26:20+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-ed1a0f17
- {"ae":0,"at":"2026-10-16T15:30:12+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-dd3724ed
- {"ae":0,"at":"2026-10-16T15:34:20+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- forward:H-d9dabea2
- {"ae":0,"at":"2026-10-16T15:34:20+00:00","bets":0,"bsp_roi":null,"p_value":null,"sp_roi":null}
- reflection:H-220ad315
- {"agent_run_id":311,"at":"2026-10-16T16:03:16+00:00","observations":["O-83cf6feb"],"reflection":"The thread has narrowed the deficit down to one combination: a last-time winner that comes back within 14 days at 9/1 or bigger. Across all handicaps the winner tag alone gives a small, steady shortfall that never reaches significance: 0.879 on 2,634 bets (raised 0.843, unraised 0.817). Restricting to quick returners deepens it to 0.651 on 424, holding in both GB and IRE, and it misses p 0.05 by a hair (0.0511). The two rival explanations are ruled out. Handicap non-winners at 9/1–14/1 scored 1.029 on 15,550, so this is not a price-band effect. Quick-returning non-winners at 9/1+ scored 0.989 on 7,481 against a same-race control of 1.0, so the quick return alone does nothing. The effect needs both the winner tag and the quick return. The non-handicap replicate (O-83cf6feb) pointed the same way, 0.457 on 106 bets (control 1.03), but it has only 2 winners. It shows no shortfall at 16/1+ (1.156 on 67) and 0 wins from 39 up to 14/1, and its BSP ROI of +19.2 shows that one big-priced winner dominates the result. Non-handicaps simply don't supply enough of these runners, so the replicate neither confirms nor kills the idea. The mechanism is still unexplained. Neither the raised mark nor the quick return alone is the cause, so the most plausible remaining cause is that the market over-trusts a recent win. That predicts the deficit should be strongest when the previous win was itself a surprise."}
- reflection:H-bfcb9669
- {"agent_run_id":303,"at":"2026-10-16T15:57:19+00:00","observations":["O-8be75f13","O-4e0404bd"],"reflection":"The thread is close to exhausted. It began with chase runners that won last time despite a bottom-third Jump Index. They scored A/E 0.78 on 240 bets (p 0.067) against a same-race control of 1.07. Good-jumping chase winners scored 1.046 on 902. That gap looked like a jumping effect. Every test meant to extend it has since failed.\n\n- **Widening to placed runners:** chase runners who finished first to third with poor jumping scored 1.013 on 921. The effect disappears once the pool grows.\n- **Moving to hurdles:** the mechanism predicted a weaker version of the same negative there, because hurdles punish poor jumping less. Instead the result came out the other way: A/E 1.074 on 396, steady across periods (1.07, 1.08) and countries (GB 1.08, IRE 1.04).\n- **The sharpest case:** poor hurdle jumpers placed last time and now running over fences should have been the most exposed. They scored 0.83 on 150. That is no stronger than the parent, and it rests on Ireland alone (IRE 0.44 on 57, GB 1.03 on 93).\n\nThe original pool shows the same split (IRE 0.45, GB 0.86). So the only pockets below 1 are small Irish chase samples. The mechanism cannot explain them, and every larger pool around them sits at or above par.\n\nThe best reading is that a poor Jump Index in one winning run is not a mispriced weakness. The 0.78 is most likely noise plus an Irish chase oddity. The hurdle reversal suggests something else is going on: a horse that wins despite jumping poorly may simply have more engine than the form shows. That is a weak lead, not an edge (CI 0.86 to 1.33, with the biggest gain coming from 3 winners at 16/1+). No new hypothesis is warranted until a descriptive check shows whether the Irish split or the hurdle reversal holds up away from this exact rule."}
- reflection:H-dd2d6d12
- {"agent_run_id":296,"at":"2026-10-16T15:52:13+00:00","observations":["O-68e72211"],"reflection":"The thread tested one idea: a last-time win the figures call slow (last_top_speed_pct ≤ 0.4) is over-bet when the horse is top three in the betting next time. The result pointed the predicted way but was weak: A/E 0.961 on 1,399 bets (CI 0.86 to 1.07, p 0.25) against a same-race control of 1.01. It also wasn't consistent. GB scored 0.943 on 1,210 and IRE 1.078 on 189, and the evens-or-shorter band was 0.987 on 99. The prediction was 0.82 to 0.92, so the true effect, if there is one, is at most a few points, and that is too small to bet. There is also a rival reading from outside this thread. Close-up losers with a slow top speed (IQ150d) scored 0.967 on 9,518. A low top-speed percentile may simply be a generic, slightly negative trait that has nothing to do with winning. If so, 'slow winner' adds nothing beyond that. The lesson is that top speed relative to the field is a weak and probably well-known signal for winners in the front of the market. The figures pay where they contradict the result, and 'won but slowly' doesn't contradict it much. A sharper version of the same cause is a win handed over by tactics, such as an uncontested lead, rather than one earned on merit. That has a different observable consequence, and it is worth one test. Slicing this near miss by price or country is not."}
- published_at
- "2026-10-16T17:09:57+00:00"
