track-recordlisted

**English** | [简体中文](SKILL.zh-CN.md) # track-record: judgment calibration loop discipline > Log it when you make the call, settle it when reality answers, promote it to a rule once the bias takes shape. The learning signal is **reality vs. your own past judgment** (a loop that closes across time), not "user corrections" — that's a different class of tool's territory. This document is a calibration **discipline** (it teaches the judging agent how to leave a trace, how to settle, how to turn bias into rules), not a prediction system (it bundles no auto-settlement engine or statistics backend). Read the details in `references/` on demand: ledger format in `ledger-format.md`, settlement in `settlement.md`, distillation in `distill.md`, the excuse table in `rationalizations.md`. > > One-line motivation: logging without ever backfilling = forever running on gut feel that's never actually been checked. ## Source of truth and install layer This skill's sole source of truth is the GitHub repository (address in the "Version & update self-check" section at the end). A platform's native discovery directory is only a rebuildable install mirror: Claude Code uses `~/.claude/skills/`; other hosts install into their own skills discovery directory (the path varies by harness — check the host's docs or probe it live before installing). An internal team deployment may set up its own internal source-of-truth directory with mirror sync, but never maintains multiple independently-edited copies