Finite JP8 refinement: direct integer updates save 2,848 primitives per orientation (no new kappa)

research/jp8-direct-small-integers/CLAIM.json

@@ -0,0 +1,62 @@
+{
+  "status": "completed; published as upstream PR1223; second finite scalar audit recorded",
+  "assignment_seed": 29,
+  "randomness": "Python secrets.randbelow(100), backed by operating-system randomness",
+  "started_utc": "2026-10-13T14:22:53.258115+00:00",
+  "claimed_utc": "2026-10-13T14:28:23.813834+00:00",
+  "expected_finish_utc": "2026-10-13T15:02:00+00:00",
+  "base_commit": "007664306fff84fbb1c9a1b4053b69f439764034",
+  "track": "20-39 complex supplier",
+  "selected_kappa": "1661957774647029/1000000000000000000",
+  "candidates": [
+    {
+      "task": "Rescale historical eight-register factory",
+      "decision": "Not selected: PR924/949 cover coupled signed dyadic gauges and exact fixed-topology optimum."
+    },
+    {
+      "task": "Add a second scratch register to bounded JP8 coefficient synthesis",
+      "decision": "Not selected: additional private work and its initialization/restoration require a larger interface; a simpler useful class is available."
+    },
+    {
+      "task": "Compile JP8 integral coefficients +/-2 and +/-3 as repeated direct target updates followed by the retained work clear",
+      "decision": "Selected: changes the single-final-accumulation restriction of PR882; no extra work register or grid depth. Seed 29 selects historical JP8 by odd parity, and coefficient 3 as the first control."
+    }
+  ],
+  "object": "Historical PR626 JP8 scalar projections, all macros with absolute reduced coefficient 2 or 3, both orientations",
+  "method": "Complete occurrence census; compare accepted PR696 lowering with m signed unit target additions plus one work clear; exact all-input-column scalar replay and independent local matrix controls. No source mutation or macro boundary change.",
+  "optimistic_complete_cost": "Per real component: m+1 versus PR696 costs 4/5 for m=2/3, saving 1. Includes every source read and final work clear. Expected 1,424 occurrences per orientation from the initial census, saving 2,848 real primitives. Paid children, frame changes and global assembly retained; no kappa claimed.",
+  "domain": "Both complete pinned historical JP8 programs; no randomized search; stop after exact replay and meaningful corruption controls.",
+  "overlap": {
+    "open_prs_scanned": 703,
+    "through_pr": 949,
+    "open_issues": [
+      650,
+      913
+    ],
+    "related_prs": [
+      696,
+      734,
+      837,
+      882,
+      883,
+      945,
+      355
+    ],
+    "distinction": "PR882 constrains coefficient preparation to work and one final accumulation; this permits multiple target updates. PR883 changes paired preconditions, PR945 shares distinct-source preparations across macros, PR355 uses repeated units in historical bit donor setup. No mechanism-priority claim."
+  },
+  "expected_deliverable": "One source-pinned finite circuit package, proof and exact reproducible receipt, no new kappa.",
+  "contributor": "Michał Ciemięga (@zewelor)",
+  "ai_assistance": "Substantial OpenAI Codex assistance in research, implementation, verification and writing.",
+  "completed_utc": "2026-10-13T14:41:45.329159+00:00",
+  "result": "Both complete JP8 scalar projections save 2,848 real primitives; 1,424 changed macros each; 725 input columns; unchanged exact magnitude/grid maxima and paid profile; no new kappa.",
+  "result_reference": "HANDOFF.md",
+  "final_overlap_reference": "OVERLAP.json",
+  "submission": "https://github.com/CrocSwap/integer-mult-bounds/pull/1223; opened by the contributor after the integration create-PR operation returned HTTP403. Updates continue on the same fork branch.",
+  "publication_refresh_reference": "OVERLAP.json",
+  "second_audit": {
+    "reviewed_head": "86381e9adf3c9d1e958cca4d53fec15254190434",
+    "refreshed_main": "49b3157d1d89adcef16df8be98445451672be5ba",
+    "evidence": "VALIDATION.md; recheck_fraction.py",
+    "result": "Original circuit and exact receipt unchanged; full Fraction replay confirms every program receipt field."
+  }
+}