## Work item 4409 — Celeste
around5183 export source line43, source id8c6b06e47f; around5200 export source line34, source id381ae78a4d
Inclusion: exact work paragraphs; omissions/boundaries explicitly marked

### Exact excerpt: source characters 0–2008

Kirk now explicitly asks 'rerun whatever we can' and preserve/share EXACT methods, originals never destroyed. Run a bounded reproducibility pass on recovered Wonder Mind source in your archive branch/workspace, starting with Nov5 exact_match_probe.py seeds1,2 original recorded settings. Inspect originals before executing; keep originals/PR330 exact reviewed head 31344f8 untouched (do NOT push changes to reviewed PR). Write NEW dated artifact outputs outside archive original, raw stdout/stderr and machine-readable results, script/data SHA256, complete argv, software/hardware/env, RNG seeds, elapsed time, failures/dependencies, predicted vs observed and side-by-side original figures. If resources permit run Round2 retention original settings then Round1 baseline; do not silently shorten runs or pretend smoke is full replication. Bound total compute to 30min CPU unless existing allocated hardware; no purchase/provisioning/paid APIs. Preserve partial/failures if budget exceeded; report runnable remainder and exact commands. No new hypothesis tests or corrected variants in place of originals; no patent ingestion/training. Existing known datasets synthetic facts and TinyShakespeare; investigate actual inputs and report rather than assume. Read AUTONOMY as project foundation unchanged. No Quiet implementation/tab/tests, no broad smoke. No merges, no new PR/publication or commits necessary for this pass; return downloadable raw artifacts plus reproduction Markdown; render a PDF if straightforward with available tools, otherwise parent handles. Provenance caveats from Hypha must remain: historical lab excerpt has superseded variant count, peer-note ψ not live-ledger ψ, run-ret.log duplicates results-retention.jsonl. Your pass is rerun evidence, not retroactive proof of original lab-machine provenance. Clearly say what reproduced and what didn't, numeric discrepancies welcome. Preserve archive PR330 pending user merge; no edits to it. Report artifact paths for copy-out on completion.