audit

# /parml:audit — price the frontier Produce a short, ranked, **graded** list of what this pipeline is leaving on the table, what each item would cost to take, and what it would break. Auditing does not change anything and does not decide anything — `/parml:plan` decides. ## First action, always Establish what already exists before reading a line of model code: ```bash ls -la ./.parml/ 2>/dev/null; git -C . log --oneline -3 2>/dev/null ``` If `findings.md` is already there, read it, say when it was written, and audit as a **delta** against it — re-confirming what has changed, not re-deriving what has not. An audit that silently re-proposes what was already rejected is noise. ## Hard rules 1. **Read-first.** Never edit project code. The only path written is `./.parml/findings.md`, and only after the findings are shown. No global or shared state is touched. 2. **Never launch training.** Print exact commands for anything longer than a few seconds and let the operator run it. Short read-only probes — a shape check, a version query, a profiler pass on a handful of steps — are fine, and should be confirmed before running if the machine is shared. 3. **Every finding carries a grade** from `references/evidence-grades.md`. An ungraded claim is not emitted. `folklore` is a respectable grade; silence about provenance is not. 4. **`analogy` and `folklore` never outrank `mechanism` or above**, whatever effect size they claim. They go in the Hypotheses block, ...