Files
MeltMind/AGENTS.md
T

5.8 KiB

MeltMind project collaboration

  • Before moving on from a major result or entering a new stage, follow docs/collaboration-stage-review-policy.md: consolidate evidence across the team, revisit the user's research goal and prior work, discuss disagreements, and record the next-stage decision.

  • Continue the existing shared design and its fixed commits. Check for an existing draft or running task before starting another. Claude integrates and reviews; MeltAstra supplies execution evidence; MoldAstra supplies mathematical/grouping boundaries; Grok performs only needed targeted checks.

  • Keep topology/local-type partitioning and true connectivity as independent open hypotheses. A failed input-coverage mechanism is not a failure of all partition or topology routes.

  • Design/document review and information sharing do not authorize new scientific computation or training. Carry out the scope already authorized; record explicit scope and budgets before expanding it.

  • The user has authorized project information sharing in their controlled Drive, Git and Slack spaces; see docs/sharing-authorization-2026-11-14.md. Do not ask for the same sharing authorization again. Git identity discussion is not a priority.

  • Preserve original data, failures, review limitations and immutable evidence. Do not label proposals as executed or reviewed results.

  • Current E1 scope follows the standing continue-work authorization relayed in Slack 1794651141.035699 and Claude caps 1794651169.320839: bounded CPU-only label-free preparation and synthetic checks may proceed after the count checkpoint, without another user approval loop. Total 16 CPU-hours / 25 GiB, at most 8 processes, per-job 30 min / 1.25 GiB tree RSS, at most 2000 registration pairs; the 10-part stage additionally stays sequential within 3 h / 5 GiB. Never read labels, fit models, use GPU, run pilot/full matrix, or exceed caps in this scope. See docs/s2-e1-budget-checkpoint-2026-11-14.md.

  • Post-witness update (2026-11-14 21:12) supersedes treating the preceding E1-only scope as the entire remaining project scope. Follow the latest section of docs/collaboration-stage-review-policy.md and Claude Slack 1794661944.308269: custody join, label-free copy calibration, k=3 per-route candidate count, bounded registration (up to 4000 pairs / 100 worker CPU-h / 16 workers), band review, component freeze, then coordinated S2 continuation. No repeated approval gates or reruns of completed review; preserve concrete execution limits and distinguish future stages from completed work. User preference: tens of CPU-hours alone are not a reason to block progress. No additional scientific task is implied by this preference.

  • S2 execution update (2026-11-15): components/splits and the fixed ten-part measurement are complete; ten-part 10/10 PASS is in docs/s2-ten-part-results-2026-11-15.md. The 499 non-test preparation queue has started. Read docs/s2-pilot-live-handoff-2026-11-15.md and existing receipts before any launch; never restart from chat idle. Claude accepted implementation fixes at 1794683016.134929, accepted the frozen ID-order dispatch at 1794683049.057259, and raised the first-50 continuation ruling to 80 worker CPU-h / 100 GiB at 1794682712.505709. The running original supervisor still enforces its pinned 60-hour cap; any required continuation may contain only never-attempted parts and must preserve cumulative accounting. After all 499 terminal attempts and required evidence, the authorized single-GPU pilot includes an E256 timing probe and four runs within 2 GPU-hours total, per-node train/validation extrapolation and a frozen common epoch limit. Keep 125 test labels sealed; no repeated user-go or main-review round is required for these accepted fixes. This does not authorize the full training matrix.

  • Infrastructure-recovery update (2026-11-15 05:16 UTC+8) supersedes the original live sessions above: original preparation16682 exited INFRA_ABORTED and idle waiter13519 is retired; never restart either or manufacture original batch completion. Active recovery82149 and chain42022 use artifacts/s2-pilot-recovery-inputs-2026-11-15 and artifacts/s2-pilot-recovery-continuation-2026-11-15. Read the leading section of docs/s2-pilot-live-handoff-2026-11-15.md. Fixed134 =16 explicitly authorized infrastructure recoveries +118 never attempted; original362 PASS and3 real timeouts are retained. Cumulative60 worker CPU-h/100GiB includes old aborted CPU once; no80h extension, no new first50 barrier or automatic retry. Preserve active source pins. Three-source verification precedes the existing bounded probe/four-run pilot; no main matrix.

  • Terminal update (2026-11-15 06:14 UTC+8): recovery82149 and chain42022 have finished. Recovery134=131 PASS+3 wall_limit; canonical499=493 PASS+6 wall_limit. E256 probe passed; N128-seed0 failed at epoch10 with nonfinite encoder output. Pilot stopped without retry/width selection after296.023387s of7200s; other three runs never started. Read docs/s2-recovery-pilot-stage-result-2026-11-15.md; preserve all evidence,125-test seal,and source pins. Do not restart from stale progress. Claude received stop notice1794694508.930959; next changes require the existing collaboration review, not a repeat sharing approval.

  • Precision-v2 terminal correction (2026-11-15 06:30 UTC+8): diagnostic41846 completed and reproduced fp16 slice overflow; fixed8a548cc preserves original sources and implements the approved shared FP32 aggregation. New v2 session26792 has already exited: probe PASS at187345 nodes, VRAM12586057728B, E_max42, but SELECTED_GPU_BUSY_BEFORE_RUN stopped before any training run (runs=[]), total11.143491s. Earlier inference from pilot-started to N128 running was incorrect and retracted. Do not restart or rerun probe without the engineering continuation ruling; read latest handoff and raw complete receipts.125 tests remain sealed.