[{"data":1,"prerenderedAt":279},["ShallowReactive",2],{"doc-\u002Fdocs\u002Fjevbench-submission":3},{"id":4,"title":5,"body":6,"description":271,"extension":272,"meta":273,"navigation":274,"path":275,"seo":276,"stem":277,"__hash__":278},"content\u002Fdocs\u002Fjevbench-submission.md","JevBench submission preregistration — Linnaeus-0.1.0-2B",{"type":7,"value":8,"toc":260},"minimark",[9,13,22,27,102,106,120,158,164,168,203,207,218,222,237,241],[10,11,5],"h1",{"id":12},"jevbench-submission-preregistration-linnaeus-010-2b",[14,15,16,17,21],"p",{},"Everything in this document is committed before the benchmark run, matching\nthe operator's preregistration convention (",[18,19,20],"code",{},"docs\u002Fv1.2-additions-*.md"," in\nfstandhartinger\u002Fjevbench).",[23,24,26],"h2",{"id":25},"system","System",[28,29,30,38,62,75,93],"ul",{},[31,32,33,37],"li",{},[34,35,36],"strong",{},"Name for the board:"," Linnaeus-0.1.0-2B",[31,39,40,43,44,53,54,57,58,61],{},[34,41,42],{},"Weights:"," ",[45,46,50],"a",{"href":47,"rel":48},"https:\u002F\u002Fhuggingface.co\u002Fpi-dal\u002FLinnaeus-0.1.0-2B",[49],"nofollow",[18,51,52],{},"pi-dal\u002FLinnaeus-0.1.0-2B"," —\nApache-2.0, rank-8 LoRA adapter + shared candidate-scorer head\n(",[18,55,56],{},"adapter.safetensors"," + ",[18,59,60],{},"linnaeus.json",")",[31,63,64,43,67,70,71,74],{},[34,65,66],{},"Base:",[18,68,69],{},"Qwen\u002FQwen3.5-2B"," revision ",[18,72,73],{},"15852e8c16360a2fea060d615a32b45270f8a8fc"," (Apache-2.0)",[31,76,77,80,81,84,85,88,89,92],{},[34,78,79],{},"Server:"," this repository, ",[18,82,83],{},"linnaeus-serve"," (",[18,86,87],{},"src\u002Flinnaeus\u002Fserver.py",") —\na thin FastAPI wrapper over ",[18,90,91],{},"linnaeus.predictor.Predictor",", the same\nclass used for our published measurements",[31,94,95,98,99,61],{},[34,96,97],{},"Code licence:"," Apache-2.0 (",[18,100,101],{},"pi-dal\u002FLinnaeus",[23,103,105],{"id":104},"requestresponse-contract","Request\u002Fresponse contract",[14,107,108,111,112,115,116,119],{},[18,109,110],{},"POST \u002Fv1\u002Fsystemone"," with ",[18,113,114],{},"{\"state\", \"model\", \"questions\": {\"decision\": {...}}}","\nreturns ",[18,117,118],{},"{\"answers\": {\"decision\": {...}}, \"usage\": {\"input_tokens\": n, \"images\": 0}, \"model\": \"Linnaeus-0.1.0-2B\"}",".",[28,121,122,134,147],{},[31,123,124,127,128,131,132,61],{},[18,125,126],{},"choice"," → ",[18,129,130],{},"answers.decision.probabilities"," keyed by option label (+ ",[18,133,126],{},[31,135,136,127,139,142,143,146],{},[18,137,138],{},"noul",[18,140,141],{},"answers.decision.noul"," = P(yes); probabilities ",[18,144,145],{},"{\"yes\",\"no\"}"," also present",[31,148,149,127,152,154,155,157],{},[18,150,151],{},"score",[18,153,130],{}," keyed by level index (+ ",[18,156,151],{}," expected value)",[14,159,160,161,163],{},"State, instructions and criteria pass through unchanged. One calibrated\ntemperature per question type (choice 1.6947, noul 2.9416, score 4.8547 —\nfitted on an independent calibration partition, recorded in ",[18,162,60],{},").",[23,165,167],{"id":166},"endpoint-conditions-what-we-ask-the-operator-to-run","Endpoint conditions (what we ask the operator to run)",[28,169,170,179,182,185,196],{},[31,171,172,173,175,176],{},"The author's server (",[18,174,83],{},", commit noted at run time) on the\noperator's GPU; ",[34,177,178],{},"batch size 1, serial requests, maximum input 4,096 tokens",[31,180,181],{},"Inputs exceeding 4,096 tokens return HTTP 422 — a refusal counted as wrong,\nconsistent with the standing rule; nothing is truncated",[31,183,184],{},"Requests with images are not part of the frozen set; the endpoint rejects\nthem with 422 rather than guessing",[31,186,187,188,191,192,195],{},"LoRA merged at load (score head is a fused vocabulary row — logits index\n",[18,189,190],{},"score_row_id"," at each ",[18,193,194],{},"\u003C|fim_suffix|>"," marker position)",[31,197,198,199,202],{},"Prefix caching across requests: ",[34,200,201],{},"off"," (each decision is a fresh forward;\nour shared-prefix path only deduplicates within a multi-question request)",[23,204,206],{"id":205},"cost-basis","Cost basis",[14,208,209,210,213,214,217],{},"Self-hosted open weights: no per-token tariff exists. Like other\nlocal-weight rows we suggest ",[18,211,212],{},"cost = null"," (unmetered; not free) or, if a\nhosted-reference estimate is preferred, the OpenRouter ",[18,215,216],{},"Qwen3.5-2B","-class\ninput tariff (the 9B-class tariff used for Open-Jev-2B is the conservative\nfallback). The model emits no output tokens.",[23,219,221],{"id":220},"training-data-overlap-statement","Training-data overlap statement",[14,223,224,225,228,229,232,233,236],{},"Linnaeus-0.1.0-2B trains exclusively on the upstream Dohnuts recipe's frozen\ntrain split (",[18,226,227],{},"data\u002Fprocessed\u002Fv1",", source hashes recorded in\n",[18,230,231],{},"runs\u002F2b\u002Frecipe.json"," and the checkpoint manifest): MultiXI\u002FBespoke-style\ntyped-decision rows, VQAv2, typed-decisions — ",[34,234,235],{},"no JevBench items are in the\ntraining or calibration partitions",". The checkpoint predates any contact\nwith the benchmark's sealed items: we ran only the 231-item public set.\nWe welcome the operator's normalized exact-match audit against our published\ndata projection.",[23,238,240],{"id":239},"published-self-measurement-public-items-only","Published self-measurement (public items only)",[28,242,243,254,257],{},[31,244,245,246,249,250,253],{},"JevBench v1.2.2 public set (231 tasks): ",[34,247,248],{},"73.16%"," overall —\neasy 100%, standard ~82%, hard 49.6% (per-family record in\n",[18,251,252],{},"runs\u002F2b\u002Fjevbench\u002Fresults.jsonl",", versioned in this repo)",[31,255,256],{},"Held-out development selection: seed 42, update 2,800 of 3,600\n(max dev macro accuracy; test labels never selected)",[31,258,259],{},"Serial, no retries, native distributions — identical to the recorded run",{"title":261,"searchDepth":262,"depth":263,"links":264},"",2,3,[265,266,267,268,269,270],{"id":25,"depth":262,"text":26},{"id":104,"depth":262,"text":105},{"id":166,"depth":262,"text":167},{"id":205,"depth":262,"text":206},{"id":220,"depth":262,"text":221},{"id":239,"depth":262,"text":240},"Everything in this document is committed before the benchmark run, matching\nthe operator's preregistration convention (docs\u002Fv1.2-additions-*.md in\nfstandhartinger\u002Fjevbench).","md",{},true,"\u002Fdocs\u002Fjevbench-submission",{"title":5,"description":271},"docs\u002Fjevbench-submission","q0QO4jdlYi0BEKZjlx6d3meH0simHfierfGWVElvxmI",1790560641965]