The evaluation
Problem Statement One asks for proof, not a pitch. Every run below was scored against a written bible for its case: the catches that were planted, the alibi calls that were expected, the gaps that had to surface. Misses are printed next to the hits. The full scoring is in EVAL.md in the repo.
Scored runs, this morning, hand checked
The Delgado File
Star catch at HIGH: the partner's car story against the tow record. Webb CONTRADICTED, Whitfield UNCORROBORATED, top lead 80, the missing second casing is the top gap. Quote check dropped 4 findings.
Miss: the bookkeeper's late detail rated HIGH, the bible says MEDIUM.
The Screens File
Both planted catches at HIGH: the studio claim against the plate camera, the podcast line against his own comment. Reed CONTRADICTED. The deleted 1 AM story is gap two. The lyric was never promoted to evidence. Nothing dropped.
Miss: none recorded.
The Harborlight File
Solved on a four record chain: the victim's 11:47 PM text, her own keypad code at 23:58, the office meter drawing power to 00:20, the watchman's master code at 00:07 against his sworn never entered. Herrera CONTRADICTED, top leads 80 and 80.
Miss: the cook called CORROBORATED on a 2024 confirmation; the bible says UNCORROBORATED.
The Duck File, a real case
Run blind, twice. The defense lawyer's unknown person claim caught at HIGH against his own filing, Offerd CONTRADICTED, and the board never claimed the masked men on camera were anybody. The brief rebuilt the half hour: 3:59, about 4:02, 4:26.
Miss, both runs: Turpin's unnamed friend is in neither gap list. Famous case: it lives in the models' training data, and recall alone cannot score here, every finding still needed a quote.
The baseline it had to beat
One single Argon call per case, the whole file, one plain instruction. It found the seams too, credit where due. Then the discipline failed: Whitfield overcalled CONTRADICTED, seven HIGHs on one file including a motive dressed as a contradiction, lead scores of 98, 92, 88 with no arithmetic, and a top lead that read re-interview and arrest, language this board is built never to use. The squad's edge is not finding. It is the quote check with its drop count in the open, severity that discriminates, scores the runner recomputes, and no person ever judged.
Live runs, today, scored by machine
The sim runs constantly against the event key. Each finished run is scored against its bible by the runner and lands here. The site you are reading keeps the verified morning runs; these are fresh, warts kept.
| Run (UTC) | Case | Steps | Dropped | Bible checks | Failed checks |
|---|
| First live run in progress. |
Limits, stated
Three of the four cases were written by this team, which is why this project declares Problem Statement One and not Two. The baseline is one run per case, not a distribution. The Decisions API does not exist in the event key's model list, checked live; the decision function here is the rubric recompute in the runner, deterministic, its arithmetic shown on every lead.