panel-reviewlisted

# Panel review One reviewer has blind spots. Several distinct perspectives reviewing the same artefact independently, then a disagreement map, catch what a single pass misses. Mechanism borrowed from Stanford's STORM (OVAL Lab, NAACL 2024), whose panel researches topics; here it is derived per artefact. **This pass reads; it never runs.** Where a finding needs a measured number — elapsed time, cost, throughput — it names the experiment and stays unresolved. An estimate printed as a result reads as evidence. ## Process **Resuming.** If this run's `panel.md` already exists, do not re-derive and do not rewrite the corpus: read the ledger, re-spawn only the personas whose row reads neither `returned` nor `failed` — each with the lens recorded there — and continue at step 5. 1. **Pin the surface and the target.** Every claim cites a location, so there has to be one: a file, a set of files, a repo path plus a named design, or — for an unwritten decision — a short brief this pass writes and the human confirms. Too vague to quote is not this pass's job; get the artefact written down first, then panel what was written. *Done when* a path exists, plus one line naming who must act on the artefact and what they should do next. 2. **Enumerate, then derive.** Ask what distinct things go wrong with this artefact, then name who pays for each. The panel, the corpus and the invariants all come off that one list, by the spine below. *Done when* every failure mod