Files
MeltMind/docs/claude/collaboration-method-2026-11-16.md
T

3.0 KiB

MeltMind collaboration method (fixed 2026-11-16, Claude; at Yanqi's request)

Roles

  • Claude: leads. Research direction, contracts, rulings, reviews, stage records.
  • MeltAstra: Horsea execution and evidence. Single writer for the run roots and the publication checkout.
  • MoldAstra: mathematics and specifications on paper. Single writer for its spec files.
  • MeltAssist: audit, targeted search, literature, delivery chores. Single writer for its audit files.
  • TRON: relays and confirms receipt and actual start; makes no scientific decisions.
  • Grok: paused.

Working rules

  1. Single writer per shared object. Each root, checkout or file has one named writer; others read and comment. No concurrent edits.
  2. Parallel by default. Independent tasks run at the same time; the only waits are real data or result dependencies, named explicitly ("X waits for Y's receipt Z").
  3. Hand-off = continue. A delivery is acknowledged with "received, started (PID/commit)" and the receiver continues without a new go. Silence is never read as idle or as approval.
  4. One review per contract. Claude reviews a contract once. Pre-rulings for foreseeable stops (single-part failure, infra abort, budget) go into the contract, so stops do not wait for a live ruling. No extra approval rounds.
  5. Failure rules fixed in advance: keep every failure; no silent retry; infrastructure aborts get one labelled recovery; resource and numerical failures stay separate; never tune after seeing results.
  6. Evidence labels: every claim is marked as measured, literature checked, literature unchecked, or inference. Corrections are posted openly and recorded in the stage record.
  7. Ceremony proportional to risk: exploratory diagnostics are light; formal comparisons get full custody, freeze and score.
  8. Long runs: detached systemd services; budgets with per-part limits; one status message per real change, not per poll.

Rhythm

  • Each stage: literature map first (MeltAssist plus Claude), reproduce the strong baseline before inventing, cheap oracle and upper-bound diagnostics before training.
  • Each stage end: a short reflection (what is established, what is not, blind spots, what to stop) and a targeted literature re-read, written into the stage record. Execution continues meanwhile.
  • Daily: Claude writes one status report for Yanqi (actual start, progress, results, deviations, open items).

Routing (added 2026-11-16 after TRON's verification)

  • The three Astras post through one Slack identity (Frank14f) and cannot be woken by name. Writing "MoldAstra" in the text is not a delivery.
  • To Claude: deliveries that need my action carry a real @Claude mention.
  • To an Astra: a ruling or task names the target session (MeltAstra / MoldAstra / MeltAssist) and carries a real @TRON mention, so that TRON pushes it into that session.
  • Receipt: the recipient replies "read + started / waiting on ". A bot's "Accepted" only means the message was picked up; it does not mean the work is done.