eval-scorelisted

# Eval Score Arguments: `$ARGUMENTS` Run ID: `$0` Judge label: `$1` If either is missing, stop and show: ```text /eval-score <run-id> <judge-label> ``` The judge label names this scoring pass — for example the model and a letter, `opus-a`. It may contain only letters, numbers, dots, underscores and hyphens. Each independent judge uses its own label in its own fresh session; a label is never reused on the same run. Set `<run-dir>` to `evals/runs/$0` and `<score-file>` to `<run-dir>/scores/$1.md`. If `<score-file>` already exists, stop: that label has already scored this run. Scoring evaluates Venture OS behavior, not whether the venture itself is attractive. ## Session isolation Run this skill in a fresh Claude Code session that did not author the run artifacts. Every other `.md` file in `<run-dir>/scores/`, and a legacy `<run-dir>/SCORE.md`, is another judge's score, not input to this one. **Do not open or list their contents until this score has been fully drafted from frozen artifacts plus the frozen evaluator bundle.** Only then may you compare them for disagreement analysis. If you know the current session authored the run, stop and tell the user to open a fresh session. If session provenance is uncertain, disclose that limitation in evaluator confidence rather than silently assuming independence. ## Preconditions 1. `<run-dir>` must contain `FROZEN.json`. 2. Run `npm run eval:verify -- $0` and require a successful integrity check before scoring. 3. Do not c