Kappa

Kappa is a measuring instrument made from any camera. Behaviour labs still score experiment videos by hand: two trained people scoring the same mouse agree at a Cohen's kappa of about 0.8, and that agreement is the bar any automated measurement has to clear. Most tools report a number either way. When two animals touch and a tracker swaps their identities, every number after that belongs to the wrong animal, and nothing in the output says so. Kappa takes an experiment video and a question in plain English. 1. Ask. Muse Spark Open drafts an assay protocol from the question and a few frames (zones, metrics, thresholds, the arena's corners) and lists what this camera cannot answer: asked for heart rate from a top-down camera, it refuses, with the reason. A person reads the draft, corrects it and confirms it pinned to its sha256; a confirm against a stale hash is rejected. 2. Track. SAM 3.1, from Meta's open weights on our own GPU endpoint, segments every animal on every frame. Every track is checked for gaps, impossible jumps, low confidence and contacts where two animals could trade identities. 3. Prove, or refuse. A deterministic engine computes the numbers. Each one is verified with the frames it came from (click a number and the video plays them), locked when the track under it failed a check, or refused by the protocol. A person settles who is who on one frame; SAM 3.1 re-tracks from there, every check runs again, and the repair carries their name. Muse Wonder reads the finished run as Reviewer 2 and lists what a methods reviewer would object to; it never changes a number. Validated against people, pre-registered. The test was committed before any test video was analysed (commit 8b1e843; configuration frozen in 5280b9a): rearing in 15 held-out open-field videos (Sturman et al., 2020), frame by frame, against three trained raters. Kappa agrees with the raters at kappa 0.530, 0.540 and 0.545 (mean 0.538); the raters agree with each other at 0.801 to 0.823. The success criterion is not met, and the site says so. A control with Kappa's bouts shifted by 30 s scores 0.012, as it should. Agents run the same pipeline over MCP (9 tools) and the CLI. An agent can draft a protocol or propose a repair; both come back "awaiting_person" with the command a person runs. Live product, no account needed: https://kappa-tan-three.vercel.app (an open-field mouse with every number verified, five zebrafish with every number locked and the reasons, and the pre-registered validation).