{"name":"The learning ledger","licence":"CC BY 4.0","licence_url":"https://creativecommons.org/licenses/by/4.0/","enabled":true,"about":"Four of the Lab's parameters move on their own, each by one published rule, at most once a week, and each move is logged here with the score that allowed it. Learning off sets every learned value aside at once. No model is called and nothing is paid.","rules":[{"rule":"odds_weight","on":true,"gate":"After the odds beat the prior's Brier score four weeks running on at least the minimum scored probes, one evidence weight moves by at most a quarter, whichever way would have scored best on the last four weeks; it moves back if the next two scored weeks are worse than the prior, and no other weight moves until it has been settled.","starting":{"probe":1.0,"receipt":0.25,"repeat":0.5,"settlement_matched":0.25,"settlement_differs":1.0,"live_price_above_listing":1.0},"bounds":{"probe":[0.5,2.0],"receipt":[0.125,0.5],"repeat":[0.25,1.0],"settlement_matched":[0.125,0.5],"settlement_differs":[0.5,2.0],"live_price_above_listing":[0.5,2.0]},"in_use":{"probe":1.0,"receipt":0.25,"repeat":0.5,"settlement_matched":0.25,"settlement_differs":1.0,"live_price_above_listing":1.0},"step":0.25},{"rule":"choice_model","on":true,"gate":"Each refit's held-out log-loss, as a share of picking at random, feeds a self-starting CUSUM; an alarm freezes serving at the last good fit and opens a task, and a later fit as good as that one lifts the freeze.","frozen":null},{"rule":"grading_points","on":true,"gate":"Once the calibration study's 95% range for a finding's payer-share difference has excluded zero four weeks running, its points move towards 25 times that difference over the reference difference (a 10-point lower paid share), clipped to 0 to 25, by at most the weekly step. Rules that protect buyers, trust rules and points set by a person don't move.","in_use":{},"max_step":8},{"rule":"synonyms","on":true,"gate":"A task whose buyer-model pick missed its tool in two runs of the own-listings test, and that the front door's index missed too, adds its content words to that tool's synonyms; a synonym whose tasks the index gets right less often than the tool's average is dropped. Only the even half of the tasks is learned from; the odd half is held out.","in_use":{}}],"entries":[]}