Gallops research lab
Gallops is a research lab that runs itself. Claude agents propose betting hypotheses, each with a reason the market should miss it. A deterministic engine measures them, a hostile agent tries to break the survivors, and the few left face a sealed holdout and then races that had not been run when they were frozen. No model ever judges whether something makes money. Everything below is the lab as it is right now.
Loading the lab…
Running
Last heartbeat
Available
57 agent jobs done
The compute lane keeps going when Claude is limited.
Work done by the lab since midnight (UTC). Imported history is not counted here.
investigate (ai lane)started
5 observation(s) waiting, top interest 0.55
Part of thread H126-lightly-raced-four-year-old-returnsai-1 · attempt 1
0.60
overlay · passes: A/E 1.228 on 621 bets, p 0.0201, periods [1.15, 1.29], control 0.958 ·
overlay · passes: A/E 1.211 on 571 bets, p 0.0328, periods [1.21, 1.21], control 0.961 ·
seasonal · vault SUPPORTIVE: A/E 1.101 on 698 bets, p 0.2014 against BH threshold 0.003226 over 31 vault tests ·
Live: checks for new events every 10 seconds
ai-1 took job 1362 (investigate, attempt 1)
· ai-1
Job 1370 finished
· engine:ai-1
Queued evaluate on the compute lane: discovery run of the broad rule
· engine:scheduler
H-503a093f Proposed → RESEARCHdesigned pre-spec written and hashed
· agent:designer@v1
agent run 57
· agent:designer@v1
ai-1 took job 1370 (design, attempt 1)
· ai-1
Job 1369 finished
· engine:ai-1
Queued evaluate on the compute lane: discovery run of the broad rule
· engine:scheduler
H-5ef5773d Proposed → RESEARCHdesigned pre-spec written and hashed
· agent:designer@v1
agent run 56
· agent:designer@v1
ai-1 took job 1369 (design, attempt 1)
· ai-1
Job 1361 finished
· engine:ai-1
thread reflection IQ140a-wide-trip-with-figures
· agent:investigator@v1
Queued design on the ai lane: write the pre-spec
· engine:scheduler
New idea H-503a093f: Wide and prominent last time: covered extra ground while doing the pace work (rival-explanation)
· agent:investigator@v1
Queued design on the ai lane: write the pre-spec
· engine:scheduler
New idea H-5ef5773d: Wide trip in a small field last time should pay less if lost ground is the cause (test-mechanism)
· agent:investigator@v1
question answered Q-b8b1b52c
· agent:investigator@v1
New question: Are the 16/1+ winners that lift the big-field and sharp-track subsets (6 wins each, A/E about 2.0) largely the same horses? (thread)
· agent:investigator@v1
New question: Within the full parent base on the Flat, does A/E differ between runners drawn in the low half and the high half today? (thread)
· agent:investigator@v1
agent run 55
· agent:investigator@v1
ai-1 took job 1361 (investigate, attempt 1)
· ai-1
Job 1355 finished
· engine:ai-1
Queued attack on the compute lane: attack: Handicaps only
· engine:scheduler
Queued attack on the compute lane: attack: 8/1 or shorter (removes the 16/1+ and 10-14/1 bands doing the lifting)
· engine:scheduler
Queued attack on the compute lane: attack: Beaten-lengths threshold tightened to 4
· engine:scheduler
Queued attack on the compute lane: attack: Sharp turf tracks only (separates track shape from all-weather surface)
· engine:scheduler
Queued attack on the compute lane: attack: Same sharp-track close finish but NOT wide
· engine:scheduler
Queued attack on the compute lane: attack: Like-for-like galloping control (last_track_sharpness = 0)
· engine:scheduler
agent run 54
· agent:falsifier@v1
Job 1359 finished
· engine:compute-1
forward result for H-b05a33c3 (forward vs bar {'min_bets': 100, 'min_ae': 1.0, 'max_p': 0.2}): info
· engine:evaluator
compute-1 took job 1359 (forward_check, attempt 1)
· compute-1
Job 1358 finished
· engine:compute-1
forward result for H-568b7703 (forward vs bar {'min_bets': 100, 'min_ae': 1.05, 'max_p': 0.2}): info
· engine:evaluator
compute-1 took job 1358 (forward_check, attempt 1)
· compute-1
Job 1357 finished
· engine:compute-1
forward result for H-4e7cc5f4 (forward vs bar {'min_bets': 100, 'min_ae': 1.05, 'max_p': 0.2}): info
· engine:evaluator
compute-1 took job 1357 (forward_check, attempt 1)
· compute-1
Queued investigate on the ai lane: 5 observation(s) waiting, top interest 0.55
· engine:scheduler
1,163 tests run, ever
31 sealed holdout queries · 2 this month · 38 of 40 left this month
Test enough ideas and some look good by luck alone: at 5%, one in twenty. So the lab counts every test it has ever run, including 940 from the archive, lets each frozen idea query the sealed holdout only once, and judges those queries together with false-discovery control (Benjamini-Hochberg over every vault test so far). A pass there is still only PROVISIONAL: only races run after an idea was frozen can make it VALIDATED. See the vault.
H-edbf6921 (Wide trip in a big field last time loses more ground, so pays more) passed discovery: A/E 1.23 on 621 bets (p 0.02), same-race control 0.96.
result0.60IQ140a: wide trip with figures
H-1eed6021 (Like-for-like: ground-on-bends test inside the full parent base (sharp) passed discovery: A/E 1.21 on 571 bets (p 0.033), same-race control 0.96.
result0.60IQ140a: wide trip with figures
H-b05a33c3 (Lightly raced four-year-old returners on the all-weather: is the edge ) met the sealed vault supportive: A/E 1.1 on 698 bets (p 0.2014, threshold 0.0032).
result0.55waitingH126: lightly raced four year old returns
H-b05a33c3 (Lightly raced four-year-old returners on the all-weather: is the edge ) survived falsification: 4 of 5 attacks left it standing; it is frozen for its one vault query.
result0.55waitingH126: lightly raced four year old returns
H-568b7703 (Lightly raced four-year-old returners whose mark was raised: improvers) survived falsification: 1 of 1 attacks left it standing; it is frozen for its one vault query.
result0.55waitingH126: lightly raced four year old returns
H-b05a33c3 (Lightly raced four-year-old returners on the all-weather: is the edge ) passed discovery as a niche: A/E 1.18 on 140 bets (p 0.268), same-race control 0.97.
result0.50waitingH126: lightly raced four year old returns
Each bar counts the hypotheses that were ever at that stage; “here” is how many sit there now. Work from before the lab existed was imported where it ended: the published systems are LIVE because a human published them, not because the lab validated them. 277 have been killed along the way, and are kept: a failure is knowledge too. Hover a stage for what it means.
seasonal · survived falsification ·
seasonal · vault skipped: rating_change do not exist before the discovery zone; track it forward instead ·
seasonal · survived falsification ·
seasonal · niche: thin but right way, A/E 1.179 on 140 bets (p 0.268), control 0.97 ·
seasonal · niche: thin but right way, A/E 1.376 on 41 bets (p 0.231), control 0.939 ·
overlay · falsified: IRE A/E 0.9 ·
seasonal · fails: A/E 1.04, needs 1.05; p 0.4051, needs 0.05 or lower (not a niche pass: A/E 1.04 below 1.05) ·
seasonal · fails: A/E 0.953, needs 1.1; p 0.606, needs 0.05 or lower; periods [1.04, 0.86], each needs 1.0; A/E 0.953 against same-race control 1.01, needs 0.03 above (not a niche pass: A/E 0.953 below 1.1; p 0.6059551900052518 above 0.5; does not beat the same-race control 1.01; a period points the other way [1.04, 0.86]) ·
seasonal · falsified: driven by one price band (up to 14/1): the rest score 1.00 ·
seasonal · fails: p 0.3532, needs 0.05 or lower; periods [0.61, 1.16], each needs 1.0 (not a niche pass: a period points the other way [0.61, 1.16]) ·
seasonal · fails: 89 bets, needs 300; p 0.332 (lower tail), needs 0.05; periods [1.1], each needs 0.99 or lower ·
overlay · fails: 41 bets, needs 300; A/E 0.973, needs 0.95 or lower; p 0.57 (lower tail), needs 0.05; periods [], each needs 0.99 or lower ·
age · fails: A/E 0.974, needs 1.05; p 0.6117, needs 0.05 or lower; periods [1.0, 0.88], each needs 1.0; A/E 0.974 against same-race control 1.16, needs 0.03 above (not a niche pass: A/E 0.974 below 1.05; p 0.6116703107013322 above 0.5; does not beat the same-race control 1.16; a period points the other way [1.0, 0.88]) ·
Discovery grades (green, amber, red) are measurements on the data an idea was found on. They never mean a system is validated.