Mister-Meeseeks · Oct 16, 2026 · Oct 16, 2026
diff --git a/research/t185-dirty-entry-frame-screens/NOTICE b/research/t185-dirty-entry-frame-screens/NOTICE
@@ -0,0 +1,12 @@
+T185 dirty-entry and frame-refit screens
+Prepared by digitalis-p with substantial Claude (Anthropic) assistance. New work is Apache-2.0.
+
+Vendored inputs (byte-preserved; provenance and SHA256 in vendor/INPUTS.json):
+- vendor/pr2857/: PR #2857 shared-kernel words, Yiyang (@Lyy-iiis), with its NOTICE.
+- vendor/pr2894/: PR #2894 preload-move words, Jeremy Kothe (@jeremykothe2) on #2863/#2857, with its NOTICE and CREDITS.md.
+- vendor/pr2833/: PR #2833 dirt-entry words, Jeremy Kothe (@jeremykothe2) and retained lineage, with its NOTICE.
+
+Methods and data credited: Philippe Van Velzen (#2853) for the A-phase dirt-entry census whose 16-helper result is
+the positive control; Dugongue (#2863) for the fixed-inventory packing optimum and dual prices used in
+tools/dual_screen.py; PR2833 audit_local.py / PR2857 shared_kernel.py for the frame-ledger rules reimplemented
+in tools/wordlib.py (nothing imported). This notice supplements, and does not replace, the vendored notices.
diff --git a/research/t185-dirty-entry-frame-screens/README.md b/research/t185-dirty-entry-frame-screens/README.md
@@ -0,0 +1,164 @@
+# T185 dirty-entry and frame-refit screens (negative result)
+
+**No new κ is claimed.** This package bounds the most direct native-word lever left after
+[#2863](https://github.com/CrocSwap/integer-mult-bounds/pull/2863) proved the fixed-inventory packing optimum:
+letting more helpers enter dirty **while every existing gate stays fixed**. It screens the PR2857 shared-kernel
+words of #2863 and the newer preload-move words of
+[#2894](https://github.com/CrocSwap/integer-mult-bounds/pull/2894) (Jeremy Kothe, 07:49 UTC, κ ≈ 0.0017201845).
+Within the screened class there is **no frame-feasible candidate** on any of the four words. The same screen
+reproduces exactly the 16 quartet helpers found by #2853's census on the older PR2833 words.
+
+With every frame constraint removed, all 184 eligible helpers of the PR2857 words become algebraically
+cancellable: the obstruction is frame placement, not linear algebra. #2894 is consistent with this. Its two new
+quartets (456,457,466,467) and (458,459,464,465) all have a zero-frame first gate in the PR2857 words, outside
+the fixed-gate class, and #2894 reaches them by re-synthesising the preload. This agent's ledger independently
+reproduces #2894's local profile {1:2404, 2:642, 3:262, 4:128, 5:26, 8:8} and its inventory {8:86, 9:232}.
+
+## Why this lever
+
+On #2863's fixed packing model (dual prices from its `review/PROOF.md`), the float screen in
+`tools/dual_screen.py` gives these first-order values per helper changed in all nine calls of a replica:
+
+| Change | κ change, first-order |
+|---|---:|
+| Any single packing move (the #2863 gap) | about 4e-9 |
+| One more PR2857-type staged rewrite, width 9→8, local 3H1−2H2 | +3.6e-7 |
+| Enter dirty at a dimension-1 first frame (−H1, width 9→8), before receiver costs | +5.9e-7 |
+| Width 9→8 with no native saving | −4.8e-5 |
+
+The break-even extra climb cost for a dimension-1 entry is 0.0122 per call. These are screening estimates that
+assume an unchanged LP basis, not certificates.
+
+## What was screened
+
+A compensated helper q0 can start at a nonzero frame only if its whole target-dirt response is cancelled inside
+the program. `tools/entry_screen.py` tests whether that response lies in the exact rational span of responses
+reachable by inserted gates `r += u*q0`:
+
+- **Responses.** `tools/wordlib.py` computes an exact backward (adjoint) pass giving every register's
+  target-dirt response just before every program item. It reproduces the stored compensation, after the
+  precondition map, for 402 of 408 live helpers on all four words. The six exceptions are the warp-tainted
+  initial-correction sources, which are excluded.
+- **Insertion times.** Any program position outside the prelude and cut, up to q0's purity horizon. Before q0's
+  first use, q0 = z0. Afterwards every read partner must be unchanged, with cancellation gates
+  `r += -u c s` at the same frame.
+- **Receivers.** Every live non-target register except q0, entering helpers and correction sources. Mode
+  `direct` also excludes q0's own partners. Mode `pre-only` allows only insertions before q0's first use.
+- **Frames.** One gate frame U with join(current frames) ⊆ U ⊆ meet(next requirements) for the receiver, q0 and
+  every cancelled partner. Requirements come from the full program timeline: A, A/B target and centre pins,
+  warps, B, restore and end frames.
+- **One-hop relays (`--chains`).** Inject into an intermediate helper m, then `r += k*m` later, with m's partner
+  cancellation. A relay gate moves m's whole content, so a plan is valid only with a nonzero injection
+  coefficient. The test is: target ∈ span(V + R_m), and either R_m ∈ span(V) or target ∉ span(V).
+- **Cost.** A pruned greedy upper bound on the extra climb cost, with H_r = r^p at the #2863 root.
+
+`tools/refit_screens.py` is a **side check**: single-gate frame moves and forward/backward frame cascades
+inside the fixed scalar word. The task card treats that class as covered by #1995/#2013 for older words.
+
+## Results
+
+| Word | Screen | Result |
+|---|---|---|
+| PR2833 base/extra (control) | direct, pre-only | exactly {460,461,468,469}, {462,463,470,471}, {478,479,490,491}, {480,481,492,493} |
+| PR2857 base/extra (#2863 words) | direct, pre-only | **0** of 184 eligible (86 have a zero-frame first gate) |
+| PR2857 base/extra | one-hop relays | 8 apparent (409, 411, 413, 415, 425, 427, 429, 431), **0 valid** |
+| PR2857 base/extra | refits | single-gate 345 candidates, forward 2797 (53 feasible), backward 3009 (96 feasible): **0 improving** |
+| PR2894 base/extra (preload-move words) | direct, pre-only | **0** of 182 eligible (84 zero-frame first) |
+| PR2894 base/extra | one-hop relays | 8 apparent (same helpers), **0 valid** |
+| PR2894 base/extra | refits | single-gate 345 candidates, forward 2803 (56 feasible), backward 3015 (96 feasible): **0 improving** |
+| PR2857 base/extra | frames relaxed (control) | 184/184 in span, including 86/86 zero-frame-first helpers |
+
+Failure controls in `verify.py`:
+
+- A corrupted compensation coefficient is detected.
+- A broken frame chain is rejected by the ledger.
+- A de-optimised word (one climb split by lowering a gate frame) is repaired by the refit screen.
+- The relay-validity test accepts six known relays for quartet helper 478 on the PR2833 word.
+
+The survey agent independently re-ran #2853's own census tool (`dirt3.py --climb`) on the PR2857 words and
+found 0 of 185 passes. That corroboration is local-only and not packaged.
+
+**Bugs fixed during the session.** Each produced false positives before publication, and each was caught by
+building a literal word or by the control:
+
+1. Responses were recorded after an event instead of before it.
+2. Next-frame requirements ignored B, restore and pin events.
+3. A pre-insertion into q0's own partner was combined with a post-insertion that assumed the partner unchanged.
+4. Relay vectors were treated as independent of the injection into the intermediate.
+
+The published tools implement the corrected rules.
+
+## Independent reproduction of #2863
+
+`evidence/T185-REPLAY.json` records the reproduction.
+
+- **Fast replay:** all five stages pass under Python 3.14.
+  - The published head omits two `.gitattributes` files (`* -text`) that its manifest lists; restoring them
+    from #2857 matches the pinned hash.
+  - The regenerated `FILL.json.gz` differs only in its gzip container.
+- **Complete cold replay:** passes in 1507 s (CPython 3.13.16, NumPy 2.5.4, SciPy 1.18.1, SymPy 1.14.0,
+  mpmath 1.4.1) with κ = 0.001720100805416885927944 − 10⁻⁴⁶ − 10⁻⁷⁰ exactly.
+  - The release receipt differs from the published one only in a wall-clock `seconds` field and the manifest
+    hashes it cascades into. Substituting the published values reproduces the published receipt exactly.
+
+## Reproduce
+
+From the repository root (standard library only):
+
+```sh
+python3 scripts/research_package.py --package research/t185-dirty-entry-frame-screens --check-only
+python3.12 scripts/research_package.py --package research/t185-dirty-entry-frame-screens
+# individual screens
+python3 -B research/t185-dirty-entry-frame-screens/tools/entry_screen.py \
+  --word research/t185-dirty-entry-frame-screens/vendor/pr2857/WORD-base.json.gz --chains
+python3 -B research/t185-dirty-entry-frame-screens/tools/refit_screens.py \
+  --word research/t185-dirty-entry-frame-screens/vendor/pr2857/WORD-base.json.gz
+# frame-relaxed census (about 12 minutes per word; not registered)
+python3 -B research/t185-dirty-entry-frame-screens/tools/entry_screen.py \
+  --word research/t185-dirty-entry-frame-screens/vendor/pr2857/WORD-base.json.gz --relax-frames
+```
+
+The registered `core` part takes about 110 s, and each `chains2857` / `chains2894` part about 250 s locally.
+`evidence/VERIFY-RECEIPT-local.json` stores the local outputs.
+
+## Scope and limits
+
+This is a bounded screen of single-helper dirty entry and fixed-word frame refits. It is not an impossibility
+theorem for native words. It does not cover:
+
+- changed gate schedules or mixing blocks;
+- simultaneous multi-helper entry;
+- relays longer than one hop;
+- operator-word corrections for warp-tainted helpers;
+- receivers excluded above.
+
+A positive screen would still need a literal word, the inherited exact endpoint audit, a new allocation, complete
+admission and the outer release. All inherited all-size, compiler, hosting, precision/tape and analytic
+assumptions of the frontier remain. No selected record is changed.
+
+## Open next questions
+
+The structured list is in [TASK.json](TASK.json):
+
+1. **Re-route the phase-A mixing block.** Can it be changed so zero-frame-first helpers start at nonzero frames?
+   The algebra already admits all of them. #2894's preload re-synthesis realises two quartets, leaving 84
+   zero-frame-first helpers on its words. The first step is one remaining helper's pinning gates and an exact
+   endpoint audit.
+2. **Joint plans.** Do joint multi-helper plans or relations with two or more outside helpers (e.g. 479 via
+   502/503/490) admit frame-feasible plans? This overlaps the advisory investigation in
+   [Discussion #2612](https://github.com/CrocSwap/integer-mult-bounds/discussions/2612) and #2875's unsearched
+   lane.
+3. **Warp-tainted helpers.** Can operator-word corrections make them cancellable?
+
+## Credit
+
+- **Source words:** Yiyang/Lyy-iiis (PR2857 shared-kernel words) and Jeremy Kothe (PR2833 dirt-entry words and
+  auditor conventions), both vendored byte-for-byte with their notices in `vendor/`.
+- **Census definitions and the control set:** Philippe Van Velzen (#2853).
+- **Packing optimum and dual prices:** Dugongue (#2863).
+- **Preload-move words:** Jeremy Kothe (#2894), vendored in `vendor/pr2894/` with that package's NOTICE and
+  CREDITS.
+- **Related predecessor screens:** #2875 (digitalis-p).
+- **Retained lineage:** all contributors credited in those notices.
+
+Prepared by digitalis-p with substantial Claude (Anthropic) assistance. Apache-2.0.
diff --git a/research/t185-dirty-entry-frame-screens/SOURCES.json b/research/t185-dirty-entry-frame-screens/SOURCES.json
@@ -0,0 +1,18 @@
+{
+  "schema_version": 1,
+  "base_commit": "47e4094be7b46c988b65b65db0fa92e7c1bad223",
+  "files": [
+    {
+      "path": "certificates/selected-result.json",
+      "usage": "provenance",
+      "git_blob": "b1fea0c94436b236cc2b7d41a7be032811b94c81",
+      "sha256": "bb2eaa58647b353f88d0aec2eafb9d30df1be53589231b23e12ceef3e4bad129"
+    },
+    {
+      "path": "docs/research/agent-task-catalog.json",
+      "usage": "provenance",
+      "git_blob": "889b95d936746df124872133ae1e7cc339adcb92",
+      "sha256": "7d185cd8c2af35988d638d969748ad7b670f902606444dbfc748e4402c340df6"
+    }
+  ]
+}