Files
MeltMind/docs/claude/s2-design-v0.4-2026-11-14.md
T

2.6 KiB
Raw Blame History

S2 design v0.4: rulings A1–A4 and adoption of the G1 framework (Claude, 2026-11-14)

Records the decisions already posted in Slack at 17:38 (thread 1794219487.255049) so the Git design record is complete.

Execution diff closing: codex/s2-execution-diff @ bf03ee8d.

G1: grouping framework (MoldAstra freeze sheet, Drive 1UDhLnx2fNApKDF25Z0VjyC2J91CYpvyE), adopted

  • Parts with the same source key always go to the same side.

  • Across keys, three relationships are coded separately:

    • rigid near-copies;
    • mirror or scale templates;
    • variants sharing a parent design with altered flow paths.

    All three go to the same side to prevent leakage. This is not a claim of physical equivalence.

  • Recall has two routes: physical scale and scale-free.

  • Confirmation uses an area-weighted two-way residual plus checks of small holes and thin bridges. Each reviewed pair gets one of three states (PASS / REJECT / UNRESOLVED).

  • An UNRESOLVED pair that could change a component puts S2 on HOLD.

  • The old 30 pairs are marked exposed; 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 held-out check.

  • Deterministic fold and pool rounding: adopted.

  • Still open: numeric calibration, review evidence, and the actual component and split artefacts.

A1: width selection

  • One global choice inside the pilot. The pilot's 4 runs (N and E x widths 128/256) are on one fold and the 100 % pool.
  • 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 additional search runs. The validation exposure is recorded honestly; this is not per-fold nested selection.

A2: Kendall

  • Standard tau-b over 5,000 pairs, keeping ties in the tie-corrected denominator.
  • Pairs are drawn uniformly with replacement from i<j over distinct eligible IDs.
  • Fewer than 2 IDs, or a zero denominator, gives NA with a reason.

A3: 10-part sample

  • 00000666 is excluded from the random top-tertile draw and then added as a fixed member, giving 10 distinct parts by construction.

A4: ties, validity and τ

  • An argmax tie goes to the smallest original ID.
  • A component 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.

Status

  • Design rulings: closed.
  • Grouping evidence, code, and supervision acceptance: pending.
  • The 10-part measurement, the pilot and the full matrix each need Yanqi's explicit go for that scope.