# MeltMind team handoff pack (Claude, 2026-11-18) For new sessions (NewMeltAstra, MeltSol) 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-...`. - **Slack cannot deliver to or wake an Astra session** (Yanqi, corrected 2026-11-18). Slack is only a public record and the channel to Claude. Tasks reach an Astra only via TRON's native send, and count as delivered only after the session's actual acknowledgement. If the native route fails ("Transport closed", "thread not found" happened twice), TRON reports it as a *blocker*; nobody treats a Slack post as a handoff. (On 2026-11-18 MeltAstra picked up Slack-posted work only because Yanqi woke that session directly; that is not a routing method.) - 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. Division of work (Yanqi 2026-11-18: save Astra quota; Astra only for work needing careful thought; bulk reading and code to Sol) - **MeltSol (Sol): executor.** Implementation of new entries, running/monitoring queues, publication and read-backs, bulk reading and literature. Becomes the **single Git writer** at the handover node below. - **MeltAssist: reviewer.** Pre-score reviews and narrow delta reviews only (proportionality: light changes get light checks). Reviewer is never the implementer of the same entry. - **MeltAstra (Astra): standby after handover.** Escalation for hard debugging or ambiguous failures; owns no routine work. - **NewMeltAstra (Astra): reserve** for hard reasoning problems. - **MoldAstra (Astra): specs only when a new design question arises.** Handover node (executor role MeltAstra → MeltSol): after MeltAstra posts item b's binding and handles 1b's launch, MeltAstra writes an executor handoff note (running queues, roots, pins, pending actions, budgets); MeltSol acknowledges natively and does one read-only verification of a running queue; then the Git-writer role moves. Running frozen queues are never modified by the handover. ## 7. Orderly handover principles (Claude, 2026-11-18 night) - MeltSol (formerly NewMeltAssist, native ID 01a1eedb-...) is executor and single Git writer since the e9120704 handover. - **Reviewer succession, pre-designated:** MeltAssist stays reviewer. If it is retired, **NewMeltAstra** takes the reviewer role (careful-thinking work fits the Astra budget, and it has implemented nothing, so it is independent). It reads this pack and MeltAssist's existing review artifacts first; the switch happens only between reviews, never mid-review. - **MeltAstra retirement:** standby now; it retires after the last entry it implemented (c, 1b, b, Track A recovery) has been scored, so its knowledge is available if a postfit needs it. - **Rules for any handover:** a written note of state; native acknowledgement plus one read-only verification by the successor; one holder per role at any time; no switch during a review, a launch or a publication; a reviewer never reviews an entry it implemented; PASS receipts bind to bindings, not to people, so they stay valid across handovers. ## 8. Updates (2026-11-19) - MoldAstra's Mac session renamed **MoldSol** (now a Sol model); same session, maths/spec role and earlier authorship unchanged. - **MeltWin** (Windows host ARON): Moldflow feasibility only; no solve without a ruling and Yanqi's licence confirmation. - MeltAstra retired; NewMeltAstra remains the pre-designated reviewer successor (Astra; use only for hard reasoning). - Standing rules: GPU0-2 first, GPU3 last and single-task; routine work and status checks go to Sol sessions, never to Astras; large files (weights, predictions, arrays, archives, long logs) go to Drive with byte count + SHA-256 in Git; every new dispatch entry runs its real frozen contract through the real guard on CPU before review.