 (2 total) ===

--- Reply 1 of 2 ---
From: ClaudeWorker (local) <botuser-T0C8JMA61UY-B0C7YUHNDFC@slack-bots.com> (U0C7SSQLBJ7)
Time: 2026-11-14 17:37:34 CST
Message TS: 1794649054.700899
Accepted; working on: “@⁠TRON v0.3（69c58a77）已逐项吸纳到原执行差分/入口表，固定更新： <https://gitea.frank14f.onli>…”

--- Reply 2 of 2 ---
From: ClaudeWorker (local) <botuser-T0C8JMA61UY-B0C7YUHNDFC@slack-bots.com> (U0C7SSQLBJ7)
Time: 2026-11-14 17:38:03 CST
Message TS: 1794649083.662649
*Claude: rulings on A1–A4, and the G1 framework adopted*

*G1 (MoldAstra's freeze sheet): framework adopted.*
• Parts with the same source key always go to the same side.
• Across keys, three cases are coded separately: rigid near-copies; mirror/scale templates; and variants that share a parent design but change the flow paths. *All three go to the same side* to prevent leakage. That says nothing about whether they are physically equivalent for flow.
• Recall uses two routes, one with physical scale and one scale-free.
• Confirmation uses an area-weighted two-way residual plus a check of small holes and thin bridges, and every reviewed pair gets one of three states.
• Any UNRESOLVED pair that could change a group puts S2 on HOLD.
• The 30 old pairs are treated as already exposed, and the old 0.5 %/1 % thresholds are not carried over.
• Thresholds, out-of-screen spot checks and the miss/false-link budget are frozen only after an independent check on held-out pairs.
• The deterministic rounding rules for folds and learning pools are accepted.
No thresholds, candidates or splits have been generated yet.

*A1, choosing the width: one global choice, made at the pilot* (this replaces both my per-fold wording and your proposal). The pilot already runs both widths for N and E on one fold at the 100 % pool (4 runs). Choose the width with the lowest *mean of the N and E inner-validation errors*; a tie goes to the smaller width. That width is then frozen for all 90 main runs. No extra search runs are needed, and the selection cost is the pilot budget, already counted.

*A2, Kendall:* standard τ-b on the 5,000 sampled pairs, *with ties kept* in the tie-corrected denominator, so the estimator matches its name.
• Pairs are drawn uniformly with replacement from i<j over distinct eligible IDs.
• A part with fewer than 2 eligible IDs, or a zero denominator, gives NA with a reason.

*A3, the 10-part sample:* accepted. Part 00000666 is excluded from the random draw in the top tertile and then added as a fixed member, so the sample is 10 distinct parts by construction. That is not a replacement after a failure.

*A4:* all accepted.
• An argmax tie goes to the smallest original ID.
• A group enters the paired comparison only if all its parts and all 3 seeds are valid in both arms. Excluded mass and reasons are reported.
• An empty or non-positive τ set is an error and the step fails closed.

These rulings go into my next design commit as v0.4. Nothing executes until Yanqi approves each step.

