[Matarener]

Requested by Mark · project thread

Summary

New package research/g20-dynamic-capture-census: an exact census of PR5586 B1 captures (own a retained helper, keep it zero while dead, inject an arbitrary payload, delete queries whose payload vanishes) on the final PR5708 G20 BIT word. Card bit-helper-contract, base 19fdb086e. No new kappa or supplier admission.

At the B1 target C = 0.0037 the PR5708 invoice needs +40.95 signed moment (+2.28 per local word); one owned full helper debits 8^t = 7.94.

• Single helpers, all epochs: none of the 415 helpers pays with any injected payload; best 0 (helper 684, H4 for H4), max removed rank mass 7. Every helper has exactly one live interval, so the multi-epoch premise of B1 has no instance on this word.
• Pairs: the 5 pairs whose joint ownership lengthens a dead interval all lose (exact best -1.968); all 6030 pairs with a shared query lose (5638 by a parity-class bound, 392 by exact enumeration; best -1.957).
• Larger sets: a parity-token bound (adapting PR5604's argument to chronological rows) limits removal to 6.5 per owned helper plus its free own query. Only helpers with a free rank-2 own query (180) can contribute, at most +0.556 each; reaching 0.0037 needs at least 5 compatible ones.
• Positive control: on the PR5667 word at C = 0.0035 the same census finds exactly the nine PR5671 late-binding helpers (free own H1 + removed H7), and nothing else.

Open continuation (for anyone interested)

The complete joint search over sets of three or more helpers is not done here. TASK.json gives an encoding: restrict to the 180 helpers with a free rank-2 own query and their shared queries, encode candidate injections with exact kill conditions, and solve the weighted maximum consistent subset (ILP/MaxSAT); check any witness with capture.py. A second open question asks to rerun capture.py on any changed G20 word.

Reproduce

python3.12 -B scripts/research_package.py --package research/g20-dynamic-capture-census

Standard library only, under a minute; expected PASS_G20_DYNAMIC_CAPTURE_CENSUS. check_research_commit.py passes on the commit.

Credit and scope

Dugongue (PR5586 B1, PR5665), Ali Romman (PR5421, PR5604 token argument, PR5671 control, PR5679/PR5708 words), ai-nanachi (PR5512) and all retained predecessors. Paid frame conversions, added contacts and changed words are out of scope; framework premises unchanged. Author check, not independent acceptance. Prepared for Matarener with substantial Anthropic Claude assistance.

🤖 Generated with Claude Code

https://claude.ai/code/session_011VhYmBnPWtZUVcoWbpL8nT

[Matarener]

CI note: plan fails on 2 of 678 tests, both in tests/test_research_current_status.py (test_current_ci_section_matches_selected_record, test_current_section_matches_selected_record). They check docs/ci-verification.md and the research status text against the selected record, which on main is now PR6030 (0.003125804); main still describes #5599 / uses a different status string. This PR only adds research/g20-dynamic-capture-census and touches neither file; the same two failures appear on the other open PRs #6099–#6112. The package itself passes scripts/research_package.py (normal and optimized) and scripts/check_research_commit.py locally.

---

Generated by Claude Code