Reads code, PRs, commits, review threads and CI signals; proposes each task's required hours; converges with the engineer; requests the manager's approval; then measures actual hours and computes the delta-driven score. All 32 metrics below are editable in plain English by whoever owns the preset — no code change needed to re-tune weights, thresholds or evidence sources.
Delta = actual ÷ required. <1.0 = under estimate (good). 1.0–1.2 = on plan. >1.2 = review the estimate or the execution — the point is to learn, not to punish.
rnd_task_agreementsrnd_task_actualsrnd_task_agreementsrnd_task_agreementsDirected-by, scope tag.Directed-by trailer.docs/docs/decisions/The 32 rows above are all plain-English YAML that whoever owns the preset edits directly — no code deploy. Changing weight #6 from 6% to 8%, or moving the target on #12 from 24h to 36h for a specific subscriber, takes one line and shows up in the next daily digest.
Per-subscriber overrides layer on top of the preset — so Charles's "task-time 25%, quality 30%" and Romit's "delta 0.9–1.3 widened" don't fork the preset, they just carry a delta file. When the preset itself moves, all subscribers move with it except where they've overridden.
The model is swappable. Default is Claude 5; the same preset runs against OpenAI / Gemini / Ollama with no metric change — only the AI grader (#11, #32) sees different outputs, and that's expected.