docs(napkin-math): record PR #743 and PR #744 in 20260520 plan

[neoneye]

Summary

Updates experiments/napkin_math/docs/20260520_plan.md to reflect PR #743 (merged) and PR #744 (open, awaiting merge) — both modify compress_report_section.py for run-to-run variance handling.

• Landed-on-main section: adds PR napkin-math(compress): second-pass shifts the variance failure from emission to ranking #743 (emission-layer second pass) and PR napkin-math(compress): paraphrase-tolerant quote verification #744 (ranking-layer paraphrase-tolerant quote verification) with honest framing of what each closes and what each does not.
• Known limitations: rewrites the Compress-LLM run-to-run variance entry. Emission side is closed by napkin-math(compress): second-pass shifts the variance failure from emission to ranking #743, verification side by napkin-math(compress): paraphrase-tolerant quote verification #744. Residual modes are bucket-categorisation variance and second-pass-also-misses cases, both at the LLM's emission layer (not deterministic-post-processing-fixable).
• Phase 1 status row: now references PR napkin-math: prompt cleanup for compress + extract #737 + PR napkin-math(compress): second-pass shifts the variance failure from emission to ranking #743 with PR napkin-math(compress): paraphrase-tolerant quote verification #744 called out as the open ranking-layer addition.
• Next-likely-move: compress-variance is no longer item 1. New item 1 is bucket-categorisation discipline (the risks_and_shocks vs gates_and_thresholds filing variance observed in paperclip v53c). Proposal 141, different-LLM validation, and prompt hygiene shift down.

Test plan

• Plan doc renders cleanly in GitHub preview
• PR numbers correct (napkin-math: prompt cleanup for compress + extract #737, docs(napkin-math): proposal 141 — source-preservation audit #739, napkin-math: Phase 2 extract — source-arithmetic preservation, threshold-pairing parity, source_text truncation #740, napkin-math(compress): second-pass shifts the variance failure from emission to ranking #743, napkin-math(compress): paraphrase-tolerant quote verification #744)
• Empirical claims match PR napkin-math(compress): paraphrase-tolerant quote verification #744 artifacts (165 fallback verifies, 2/3 paperclip survival, 0 lost qv=True under 90%→100% tightening)

🤖 Generated with Claude Code