← ClaudeAtlas

icm-ledgerlisted

Keep score on an AI's advice inside an ICM workspace. Every consequential recommendation is logged with a confidence word, a falsifier, a cost class, and a review date BEFORE it is given; outcomes attach in the same file; readouts turn the record into a self-briefing the next session reads on arrival. Use when (1) you are about to give a recommendation that changes what the owner does next, costs money, or ships — log it first; (2) reality delivers an outcome for an open entry — attach it; (3) the owner asks "what's our track record", "where are you overconfident", or "should I trust this call"; (4) a supervised audit pass runs — compute a readout and regenerate BRIEFING.md; an unattended maintenance pass only regenerates a stale briefing from an existing readout; (5) you are about to make a call of a class the ledger scores — read your record first; (6) the owner overrides an adviser, or two advisers oppose each other — log the disagreement as a D-entry. The record is honest only if outcomes get attached.
xnfinite/icm-ops · ★ 0 · Data & Documents · score 73
Install: claude install-skill xnfinite/icm-ops
# icm-ledger `icm-verifier` answers "was the work real?" — it scores claims about the past. This skill answers "was the judgment good?" — it scores claims about the future. A doctor gets outcomes; a forecaster gets scored for a lifetime; an AI adviser produces confident recommendations at industrial volume and never finds out whether they were right. The ledger is the fix, at the one scale where it is honest: one operation, its advisers, their calls, and what actually happened. Calibration scoring is not new — forecasting communities and eval harnesses do it. What this skill does is put the ledger *inside the working relationship*: the AI's live recommendations carry falsifiers, outcomes attach in the same workspace the next recommendation is made from, and the next session reads its own record before advising again. It does not make the adviser correct. It makes the adviser scorable, which is how you find out. This skill operates on an ICM workspace — Interpretable Context Methodology, Van Clief & McDermott, arXiv:2603.16021 (https://arxiv.org/abs/2603.16021). Build the workspace with their `icm-architect` skill (https://github.com/RinDig/icm-architect), or by hand from the paper: a front desk (`CLAUDE.md` / `AGENTS.md`), `00-catalog/CATALOG.md`, and dated markdown files with frontmatter — the `example/` tree in the icm-ops repo is a complete minimal one. This pack is a layer on top of the method, not part of it. ## First run Create `04-memory/ledger/<YYYY-MM>.md` from