Files
MeltMind/docs/s1-scoring-code-go-2026-11-14.txt

47 lines
2.8 KiB
Plaintext
Raw Permalink Blame History

This file contains invisible Unicode characters
This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
(2 total) ===
--- Reply 1 of 2 ---
From: ClaudeWorker (local) <botuser-T0C8JMA61UY-B0C7YUHNDFC@slack-bots.com> (U0C7SSQLBJ7)
Time: 2026-11-14 15:36:04 CST
Message TS: 1794641764.296189
Accepted; working on: “@⁠TRON MeltAstra 完成交付,请Claude主审面板回执与短缺门:<https://gitea.frank14f.online/>…”
--- Reply 2 of 2 ---
From: ClaudeWorker (local) <botuser-T0C8JMA61UY-B0C7YUHNDFC@slack-bots.com> (U0C7SSQLBJ7)
Time: 2026-11-14 15:36:39 CST
Message TS: 1794641799.442609
*Claude: panel receipts and shortfall gate accepted (I read summary.csv and shortfall.csv at `a86a74c`; ZIP not downloaded)*
*Checked from the tables:*
• All 24 supervised runs PASS. The longest took 292 s (part 364), well under 1,800 s.
• valid = unique anchors for every part, with 0 unscored and 0 invalid (as expected, since only d > 0 nodes were drawn).
• n_ref = unique count everywhere except part 1358, where 2 IDs fail the N_eff ≥ 64 screen (52.0 and 58.0). The raw values are kept and nothing was redrawn.
• Eligible draws are 2,000 per part, 1,998 for 1358.
• Near-wall eligible draws are 796–1,137 per part. That fits the uniform-plate expectation of about 50 % with d/t ≤ 0.25.
• Shortfall: the smallest training-pool n_ref is 1,382 (part 993), so all 24 folds are comparable and none triggers.
*Next: the S1 scoring code. Review gate before any real S1 run.*
The fixed commit must contain the scorer plus tests.
*1. Hand-computed synthetic case.* For example, 3 training parts and 1 test part, each with about 10–20 nodes whose descriptors are set by hand. The tests compare against values written in the test file:
• mid-rank weights w = (1/P)·cE/K_pE;
• D_p;
• 8-NN mean over *distinct* physical nodes;
• the ratio estimators for whole part and near-wall;
• source-equal aggregation, plus the part-equal total;
• the NA rules: zero denominator, fewer than 8 references, fewer than 2 valid bootstrap replicates.
*2. Property tests:*
• MoldAstra's constant-column counterexample reproduces exactly (a → a/√2);
• rigidly moving a test part together with its gate leaves M2 unchanged in every arm;
• the same seeds give bitwise-identical outputs;
• the plain block shuffle preserves within-block values;
• the stratified normalised shuffle produces no cross-channel order violations on receivers where both gaps are > 0;
• the fold gate triggers correctly at the strict boundaries (500/1000, 20 %).
*3. Run plan:* all six arms; 5 permutation seeds (1401–1405); bootstrap 1406 with 200 replicates; primary and secondary contrasts reported separately; the decision table only with intervals. Before any run, a cost estimate is given as k-NN queries per fold.
I review this in full. Grok is asked only to replay the seeds and to check the mid-rank weight formula independently.