Team handoff pack for new sessions

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
ClaudeandClaude Opus 5.5 committed 2026-11-18 13:50:11 +00:00
1 parent 3c05a466c1
commit 45afa47af5
1 file changed
+49
@@ -0,0 +1,49 @@
# MeltMind team handoff pack (Claude, 2026-11-18)
For new sessions (NewMeltAstra, NewMeltAssist) and anyone resuming. Read in this order: this file → `research-tree.md` → `phase2-retrospective-2026-11-17.md` → `stage-research-plan-2026-11-18.md`.
## 1. Yanqi's standing directions (verbatim where available, via TRON)
- Goal: 新零件的实用充填代理模型,少数据;争取 SOTA。"目标先在三维场景争取SOTA,工业可用是bonus,不要过度悲观"; "先使用 GINO 和 Greif 数据,力争超过相应结果".
- Method: tree search; "每个不用一直做到有效果…逐个剪枝…不断探索新的分支…Claude领导".
- Compute: use it; "先把还没探索完的研究路线都做好,现在horsea还比较空闲"; parallel *research work* ("假如你有9个点子…三个一组同时跑"), GPU0-2 up to 3 tasks per card, **GPU3 single task only (cooling)**. Not operator/throughput optimisation.
- Leadership: "以后我有什么意见,你都不要自己去执行,Claude才是领导" (to TRON). TRON relays; Claude decides; Astras execute/review.
- Focus: "好了,先不管这些细节,认真搞研究先"; don't send Yanqi questions or routine reports — stage results only.
- Sharing: may share to Drive / Git / Slack (not Undermind for internal details).
- Licence: student licence only (Fusion with Moldflow Adviser); not a commercial or batch-API right. Data generation is later preparation.
## 2. Roles and boundaries
- Claude: leads, rules, keeps `research-tree.md` (single writer of `docs/claude/`).
- Executor (MeltAstra): sole executor and **single Git writer** on Horsea; one root per entry; never edits running frozen sources.
- Reviewer (MeltAssist): independent reviews; never executes science or writes Git.
- MoldAstra: maths and specs (paper only).
- TRON: relay and monitoring; may launch an already-reviewed frozen entry *once, labelled* only when the executor is unreachable; never edits or interprets.
## 3. Process rules that held up
- Review before any test-target read (exact five-field binding PASS); training-only and documentation changes get light/narrow reviews (proportionality).
- Pre-register gates before results; failures and STOP receipts are kept; no silent retries; infrastructure aborts get **one labelled recovery**.
- Budget stops: prefer total-projection rules over per-fold caps on shared hosts; size caps from measured epochs (+ margin).
- Never back-fill observations; never reuse a PASS across a different binding; publication commit ≠ implementation commit.
- Corrections are recorded, not hidden (e.g. my g-sign misreading on item 1; B15 over-generalisation).
## 4. Routing lessons
- Native session IDs (Horsea host `remote-ssh-codex-managed:5bf202cd-...`): MeltAstra `01a1d30d-...`, MeltAssist `01a1e358-...`. TRON's native route failed twice ("Transport closed", "thread not found"); fallback = the shared Slack thread, which the Astras read. A delivery is not an acknowledgement.
- Auto-approval can block GPU launches without verifiable user authorisation; supply the original user message or ask Yanqi one yes/no question; never route around it.
## 5. Current state (2026-11-18 evening)
- Confirmed: B13-E1b (small field gain); H2 (field 7.08 → 6.58 % five folds; last fill worse).
- Running: seed-1 v2 (noise reference); Track A recovery (folds 3-4; H2 on Gre24 protocol).
- Built/under review: item c (last-fill objective) + shared capacity rule (fix pending).
- Queued (parallel after review): 1b (100-epoch schedule), b (H2 + B13 context), Track B (GINO ShapeNet).
- Blocked: fresh holdout (no gate metadata for v2 extras).
- Budgets: phase 3 cap 720,000 GPU-s; Track A separate (162,000).
## 6. Proposed division with the new sessions (planning, not yet in effect)
- **NewMeltAstra (high capability):** second *implementer* on separate roots (e.g. b and Track B) while MeltAstra keeps c and 1b; Git publication still through one writer (MeltAstra) to avoid conflicts; hard debugging / shared-infra changes.
- **NewMeltAssist (lower cost, Sol):** light reviews (ledger, constant/diff checks), monitoring receipts, publication HTTP read-backs, literature lookups, report drafting. High-stakes pre-score semantic reviews stay with MeltAssist (independence: reviewer ≠ implementer of the same entry).
- Switch-over only after current items (c fix, seed-1, Track A recovery) reach a stable point; each new session first reads this pack and confirms in-thread.