[1337isnot1337]

The all-native p7 controller pays full stage and idle transforms where a copied rank-one query shear suffices. Replacing those calls, charging the lost productive helpers, and removing the same multipliers from the complete physical word gives a new author-checked conditional κ = 37/12500 = 0.00296. This is 10.18% above #4416/#4433's 0.00268647 and 28.70% above the selected 0.0023 checkpoint.

The complex word has 17 pairs, 19 full native calls and 16 query shears. Its complete invoice has 11,560 independent productive fields, 72,902 paid jobs and rank surplus 800. The accompanying 16-pair BIT word supports ordinary saving 0.0034 after paid tensor completion, delayed covering and atom conversion. Its atom power changes to 0.004; the old 0.003 fails the new adapter comparison.

Both emitted physical words and their exact scalar boundary programs are supplied. A second literal replay checks every native/query/idle event against the specialized source operator. Complex Laurent signs and absolute phase contacts, rational BIT charts, all query copy/erase charges, both paid tails, effective finite compiler ranks, all nine outer rows, 54 strict predicates and four functional guards are retained. The new κ is a conditional construction claim, not a raw numerical root.

Package: research/query-thinned-controllers. PROOF.md gives the full construction and obligation ledger; RESULT.json and VALIDATION.md give arithmetic and evidence. The huge tensor/uniform words are algorithmically specified, not materialized.

Validation:

• CPython 3.12.14 metadata/pins and registered replay pass.
• Optimized CPython 3.12.14 replay passes from a Git-free export.
• Changed-source and unpaid-stage event controls reject; internal clear/erase, orientation, coefficient and outer-cost mutants reject.
• Committed-byte and source-history preflight passes at 40e22a83ed660d3d6074c4f2dbf14bfb621eee51.
• No Lean, Docker or complete historical native replay.

Hosted CI: run 39667735839 fails in the shared planning job before package replay. All four failures are in tests/test_research_current_status.py: selected decimal formatting, the historical banner/heading, and the stale current CI-navigation section. The status guide, CI guide, test module and selected record are byte-identical to canonical main 2a13eb1827. The same four failures reproduce independently on CPython 3.12.14 from those unchanged files. Hosted package replay is therefore skipped; the registered local and Git-free replays above pass.

Reproduce:

python3.12 -B scripts/research_package.py --package research/query-thinned-controllers

Sources are pinned to canonical 3a23e9d9d; submission includes refreshed main 2a13eb1827. Open PRs and Discussions were checked for overlap, with incomplete sparse-local/closed-body coverage stated. Frontier checks at 19:33, 19:51 and 20:06 UTC found no stronger κ through #4439.

Retains the reviewed source's all-column native, owned-erasure, whole-job matching, fixed-tape, precision, analytic, prime and exact-recovery premises. No unconditional theorem, independent/formal acceptance, selected-frontier update, practical speedup or global optimum is claimed. The next audit is the arbitrary-column owned-query lowering.

Credit: zxv's controller/native and higher-moment work, Jacob Alan Sussman's mixed-target flow, the #4416 paid-completion construction, and the complete credited native/BIT/outer lineage preserved in PROOF.md and unchanged source notices. Prepared for Cinna Davis with substantial OpenAI Codex assistance; no subagents. Apache-2.0.