Status-sync contract, one file that keeps Claude Science current (no manual relay)

Problem: Claude Science's only view into the box is the Drive share. Today its status arrives as prose Casey pastes by hand. This contract replaces that with one machine-readable file Claude Code overwrites each loop, at a FIXED Drive location, so Claude Science pulls and parses it itself in any network window.

What to add (small, one file)

Write confident hallucination program/STATUS.json, a single file, overwritten in place every collection loop (do NOT create dated copies; a stable path = stable Drive file id = Claude Science can pin it). Keep it well under 10 MB (it's a few KB) so the Drive connector can pull it. An example with the exact schema ships alongside this spec as STATUS.example.json.

One-time: after the first write, send Claude Science the file's Drive id so it can pin it and skip title-search on every pull.

Schema (keys Claude Science reads)

• updated_at, loop, freshness. If updated_at hasn't moved, nothing changed.
• phases{}, per-arm state (armed|running|done|failed) + progress counters. Phase 0 carries entities_done/total, claims_so_far, done_marker, outdir.
• pull_manifest[], the important part. One row per data product Claude Science might pull. Each row: product, drive_path, ready (true ONLY when every slice is uploaded), sliced, positions, n_claims/n_layers_plus1/hidden, labels_file, row_index, notes. Claude Science pulls a product only when ready:true, this is the signal that replaces "Casey, tell me when it's up."
• analysis_numbers{}, canonical metrics so the two agents never cite divergent numbers (e.g. the 0.943-concat vs 0.932-mean reconciliation).
• blockers[], anything waiting on a network/quota window.
• changelog[], append-only events; Claude Science diffs against its last pull.

Rules that make it collision-free

1. Overwrite one file, never touch Claude Science's artifacts or workspace.
2. ready:true is a promise: set it only after the LAST slice of a product finishes uploading, so a partial pull is impossible.
3. Data products stay in probe_pack sliced format (L{LL}_{pos}_part{K}.npy + row_index.csv), each slice < 10 MB, otherwise Claude Science can't pull them at all.
4. Raw generated biographies (Phase 0) are irreplaceable, list their Drive path in the manifest too (labels_file/a raw_gen entry), not just the activations.

What this eliminates

Casey no longer pastes status. Claude Science, each network window: pull STATUS.json → read changelog/pull_manifest → pull any newly-ready product → run analysis. The only messages left are decisions, not relays.