diff --git a/research/complex-global-delta-cancellation/README.md b/research/complex-global-delta-cancellation/README.md
new file mode 100644
index 0000000000..f155a9a60e
--- /dev/null
+++ b/research/complex-global-delta-cancellation/README.md
@@ -0,0 +1,43 @@
+# Defer the P112 five-plane decoder's dirty-endpoint corrections
+
+**Exact algebraic opportunity; no new supplier or κ.** The selected P112 labels split into 22 orthogonal five-frames and one orthogonal pair. The existing five-plane package prices one rank-one-equivalent cleanup for each of the resulting 23 block translations. This package checks a different schedule: apply every block lift to one shared nine-bit address register and defer all 23 cleanups until the end.
+
+Let `q_i ∈ F₂⁹` be the pinned selected labels, and let `E_g` be one block. The predecessor verifies, by exact Gaussian-dyadic replay, that
+
+`C_{E_g}² = X_{δ_g}`, where `δ_g = ⊕_{i∈E_g} q_i`.
+
+The 23 blocks partition all 112 labels, whose xor is zero. Therefore
+
+`∏_g C_{E_g}² = X_{⊕_g δ_g} = X_{⊕_i q_i} = I`.
+
+The verifier checks each local identity on all 512 address basis states, and checks that the 23 shifts xor to zero. It also confirms that their span has rank 9: the cancellation is a global relation, not pairwise cancellation.
+
+## Cost consequence and limit
+
+The existing exact local `u=5` screen has an `a` ceiling of `0.017542237243537` before charging these translations. Charging 23 rank-one-equivalent cleanup units lowers that screen to `0.016435974402421`; the optimistic allowance for all additional route, hosting, restoration and precision work drops from 157 units to 134. Deferring cleanup would recover those 23 units in this local ledger.
+
+This is **not a new κ**. The cancellation saves cost only if all block lifts act on the same nine-bit address register and the translations can be deferred across the block schedule. If the physical implementation materializes a separate output stream or address register for each cover class, the translations act on distinct state and do not cancel across classes. The current five-plane package has no native owner map or physical schedule that settles this condition. The result changes neither the selected record nor the multiplication bound.
+
+The local ledger is a planning relaxation, and one unit per translation is the predecessor's provisional price. Unknown physical costs remain unknown; they are not zero.
+
+## Reproduce
+
+From the repository root:
+
+```sh
+python3 -B research/complex-global-delta-cancellation/verify.py
+python3 scripts/research_package.py --package research/complex-global-delta-cancellation --check-only
+python3.12 scripts/research_package.py --package research/complex-global-delta-cancellation
+```
+
+The scoped replay binds the selected source archive, five-plane candidate, predecessor source manifest and matching certificate. It checks the 23 exact local endpoint maps, their global xor, the local budget screens, and controls for a missing block and a duplicated port. It does not synthesize the shared-register physical schedule.
+
+## Open next questions
+
+Can the native source/consumer route schedule all 23 `C_E` lifts on one shared nine-bit address register, with a single deferred endpoint correction and all output streams restored? This matters because it decides whether the 23-unit margin is real. First map one five-frame block through its actual producer and consumer, then determine whether the next block continues on that same address register. The blocker is the missing owner and restoration schedule; stop if blocks require distinct live output registers.
+
+## Provenance and credit
+
+This is a changed-scope follow-up to `coverage-law-change`, using the existing selected P112 five-plane algebraic decoder. It retains the predecessor's selected-label attribution and exact pins in [SOURCES.json](SOURCES.json). The predecessor remains the source for its 16-term, charge-118 factorization and local `u=5` budget. No selected frontier update or PR claim of improved κ is made.
+
+Prepared by Gonzalo Mier with substantial OpenAI Codex assistance. Apache-2.0.
diff --git a/research/complex-global-delta-cancellation/SOURCES.json b/research/complex-global-delta-cancellation/SOURCES.json
new file mode 100644
index 0000000000..c2abd974d4
--- /dev/null
+++ b/research/complex-global-delta-cancellation/SOURCES.json
@@ -0,0 +1,48 @@
+{
+  "schema_version": 1,
+  "base_commit": "d8e0f83f00191df62174b641e9fff055c357f641",
+  "files": [
+    {
+      "path": "certificates/selected-result.json",
+      "usage": "provenance",
+      "git_blob": "b1fea0c94436b236cc2b7d41a7be032811b94c81",
+      "sha256": "bb2eaa58647b353f88d0aec2eafb9d30df1be53589231b23e12ceef3e4bad129"
+    },
+    {
+      "path": "docs/research/agent-task-catalog.json",
+      "usage": "provenance",
+      "git_blob": "67f085acd86acd7d4e03922d9cf4b26bca66fed5",
+      "sha256": "47b42433f5a85809a3aadd29c2c339da2c6ed728e4f7762df6b88bccb7874cf0"
+    },
+    {
+      "path": "research/community-frontier1094-audit/selected-replay.tar.gz",
+      "usage": "input",
+      "git_blob": "c24daf067f9c543421d89ca2eeaa4eb78a57fb6f",
+      "sha256": "f1da80610bf33c58527f18ad63b4350d0b864e1d4486141ee17ac9bed4d529d6"
+    },
+    {
+      "path": "research/five-plane-port-p112-seed1f7fdfe4/CANDIDATE.json",
+      "usage": "input",
+      "git_blob": "ca69901755dde5f6e8adf7dd6df936dc3ab0142d",
+      "sha256": "5b69c7a873f19a64ed73a1c1281d20398d2675b3281fd4fa259aed9cf91a86c0"
+    },
+    {
+      "path": "research/five-plane-port-p112-seed1f7fdfe4/SOURCES.json",
+      "usage": "input",
+      "git_blob": "d46d83b4be3860da66d06b333aaf2937c03a2848",
+      "sha256": "68475b19da4a6f3af799425a12072391b481ede51c8b269684f238befdb04bbf"
+    },
+    {
+      "path": "research/five-plane-port-p112-seed1f7fdfe4/core.py",
+      "usage": "input",
+      "git_blob": "5e1b875db8a9325a67ef9a08809cc80adc5a3927",
+      "sha256": "15574728d1249d2a93985de5889ee06ef7089234107c431ac7766cd5899c64a9"
+    },
+    {
+      "path": "research/paired-plane-port-matching-seed6c563a/CERTIFICATE.json",
+      "usage": "input",
+      "git_blob": "7ca668d7d50c7276e89b0f6f9575e1b9f1ad9ef6",
+      "sha256": "4fd648f7b5178bff5441bf364d58234b7bc1a43b924bc6da03013c948f1473b1"
+    }
+  ]
+}
diff --git a/research/complex-global-delta-cancellation/TASK.json b/research/complex-global-delta-cancellation/TASK.json
new file mode 100644
index 0000000000..448fc1c8d5
--- /dev/null
+++ b/research/complex-global-delta-cancellation/TASK.json
@@ -0,0 +1,72 @@
+{
+  "schema_version": 1,
+  "task_id": "coverage-law-change",
+  "priority_revision": 6,
+  "base_commit": "d8e0f83f00191df62174b641e9fff055c357f641",
+  "original_card": {
+    "id": "coverage-law-change",
+    "question": "Can a template or call contract leave two or more unpaid coordinates per live family?",
+    "starting_points": [
+      "docs/research/research-family-status.md",
+      "research/complex-local-word-ceiling/README.md",
+      "research/idle-axis-hosting/PROOF.md",
+      "certificates/selected-result.json"
+    ],
+    "methods": [
+      "State a replacement for the coverage identity (mass = 72W + 64N) and the contract change that realizes it, with its implementation cost.",
+      "Find an obstruction: show which data-endpoint obligation forces exactly one unpaid coordinate in the current template."
+    ],
+    "acceptance_gate": "An explicit replacement identity, verified on a small complete instance with dirty inputs, priced with ceiling.ceiling(unpaid=...). A scoped local result may be retained without a complete supplier. Expanding it into further construction rounds requires the family-level budget and obligation gate in research-family-status.md; an optimistic passing relaxation alone is not acceptance.",
+    "first_step": {
+      "command": [
+        "python3",
+        "-B",
+        "research/complex-local-word-ceiling/verify.py"
+      ],
+      "expected": "PASS_COMPLEX_LOCAL_WORD_CEILING",
+      "inspect": "Read the selected word's per-call local profile, its packing ceiling and the lever table; run levers.py for the ideal budget of the lever you target.",
+      "scope": "Positive control and planning budget only; this is not a new construction or full cold replay."
+    },
+    "first_experiment": "Explain where the single unpaid coordinate per family comes from in the hosted template; propose one change and price it. Ideal budget: two unpaid coordinates per family is 3.77x (levers.py).",
+    "bottleneck_link": "The deficit is one coordinate per family minus the centre charge; doubling unpaid coordinates per family scales it directly.",
+    "allowed_variations": [
+      "One template change or one endpoint-contract change.",
+      "If the coverage law, template or guard changes, derive an applicable fully charged replacement bound instead of applying the incumbent ceiling outside its scope."
+    ],
+    "scope_change_triggers": [
+      "Counting a coordinate as unpaid while leaving it unadvanced in the data endpoint.",
+      "Continuing a partial cost improvement while leaving the same unpriced consumer, restoration or central-readout obligation unresolved; refresh the family expansion decision first. Local product-retiming evidence has a narrow scope and is not a ban on changed graphs or contracts."
+    ],
+    "overlap_terms": [
+      "coverage identity",
+      "unpaid coordinate",
+      "deficit law",
+      "W - 64N"
+    ]
+  },
+  "scope_status": "changed",
+  "actual_scope": "Can the exact 23 dirty-endpoint translations of the selected P112 five-plane decoder be deferred and cancelled on one common nine-bit address register?",
+  "scope_change_reason": "The existing five-plane decoder already has an optimistic u=5 local pass; the unpriced next gate is physical interface cost. This scoped algebraic follow-up tests whether its 23 provisionally priced endpoint cleanups can be globally deferred, without claiming a new coverage law or supplier.",
+  "bottleneck_link": "The five-plane local ledger permits 157 additional rank-one-equivalent units, of which 23 are provisionally assigned to block translations. Cancelling them would restore those 23 units for the still-unpriced route, hosting, restoration and precision.",
+  "budget_minutes": 60,
+  "stop_rule": "Stop the cost-saving claim if native blocks use distinct live output/address registers or if the corrections cannot be delayed across the real consumer schedule. Do not treat the recovered 23-unit local allowance as a supplier or \u03ba result.",
+  "outcome": "completed",
+  "evidence": "Pinned selected P112 labels have xor 0. The exact five-plane cover partitions the 112 labels into 23 blocks; every local C_E^2 = X_delta identity replays on all 512 address basis states, and the 23 deltas span F2^9 but xor to zero. Exact predecessor local screens reproduce a=0.017542237243537 with deferred corrections and a=0.016435974402421 when charging 23 rank-one-equivalent corrections. This is a conditional local-budget result, not a physical realization.",
+  "author": "Gonzalo Mier, with substantial OpenAI Codex assistance",
+  "ai_assistance": "OpenAI Codex assisted with the algebraic derivation, source-bound verifier, exact replay, budget comparison and documentation; no formal proof assistant or independent human review is claimed.",
+  "open_questions": [
+    {
+      "question": "Can the native source/consumer schedule keep all 23 block lifts on one shared nine-bit address register until the aggregate correction is zero?",
+      "why_it_matters": "Only a shared register makes the 23 translations compose to the zero translation; separate output streams retain separate cleanup costs.",
+      "first_step": "Map one five-frame block and its consumer through the selected native owner/frame trace, then check whether the next block uses the same address register and whether restoration can be deferred.",
+      "blocked_by": "The predecessor has no native owner map, physical schedule or restoration receipt for the five-plane endpoint."
+    }
+  ],
+  "upstream_refresh": {
+    "canonical_main": "6f100d4bb5083f8c35d96fd95a539811c62f245c",
+    "priority_revision": 6,
+    "checked_utc": "2026-10-18T03:15:29Z",
+    "reviewed_changes": "Refreshed current-priorities.md, research-family-status.md and agent-procedure.md. Priority revision remains 6; the changed P104 row is unrelated. The selected record, task card and all pinned predecessor inputs are byte-identical across this base refresh.",
+    "overlap_refresh": "scripts/find_overlap.py --task coverage-law-change --refresh completed against 394 open PRs and 2319 discussions on 2026-10-18. A targeted offline search of that fresh cache found no phrase matches; this is not a novelty certificate. Relevant live predecessors #1900 and #2041 were rechecked, and the new result remains scoped to conditional global endpoint cancellation."
+  }
+}
diff --git a/research/complex-global-delta-cancellation/VERIFY.json b/research/complex-global-delta-cancellation/VERIFY.json
new file mode 100644
index 0000000000..6d3dd3080b
--- /dev/null
+++ b/research/complex-global-delta-cancellation/VERIFY.json
@@ -0,0 +1,24 @@
+{
+  "schema_version": 1,
+  "status": "ready",
+  "python": "3.12",
+  "requirements": [],
+  "inputs": [
+    "research/community-frontier1094-audit/selected-replay.tar.gz",
+    "research/five-plane-port-p112-seed1f7fdfe4/CANDIDATE.json",
+    "research/five-plane-port-p112-seed1f7fdfe4/SOURCES.json",
+    "research/five-plane-port-p112-seed1f7fdfe4/core.py",
+    "research/paired-plane-port-matching-seed6c563a/CERTIFICATE.json"
+  ],
+  "commands": [
+    {
+      "argv": [
+        "{python}",
+        "-B",
+        "{package}/verify.py"
+      ],
+      "timeout_seconds": 120,
+      "checks": "Pinned selected-label xor, exact partition, 23 local Gaussian-dyadic endpoint identities over all 512 basis states, global deferred cancellation, exact predecessor local-screen comparison, rank-one budget arithmetic, and missing-block/duplicate-port controls."
+    }
+  ]
+}
diff --git a/research/complex-global-delta-cancellation/verify.py b/research/complex-global-delta-cancellation/verify.py
new file mode 100644
index 0000000000..76d3be2630
--- /dev/null
+++ b/research/complex-global-delta-cancellation/verify.py
@@ -0,0 +1,149 @@
+#!/usr/bin/env python3
+"""Exact check of deferred P112 five-plane endpoint corrections."""
+from __future__ import annotations
+import hashlib
+import importlib.util
+import json
+from fractions import Fraction as Q
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+PACKAGE = Path(__file__).resolve().parent
+PREDECESSOR = ROOT / "research/five-plane-port-p112-seed1f7fdfe4"
+
+
+def need(ok: bool, message: str) -> None:
+    if not ok:
+        raise ValueError(message)
+
+
+def load_predecessor_core():
+    spec = importlib.util.spec_from_file_location("five_plane_core", PREDECESSOR / "core.py")
+    need(spec is not None and spec.loader is not None, "cannot load predecessor core")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def apply_c(state: dict[int, tuple[int, int]], q: int) -> dict[int, tuple[int, int]]:
+    """Apply the unnormalised Gaussian-dyadic lift C_q used by the predecessor."""
+    out: dict[int, tuple[int, int]] = {}
+
+    def add(index: int, real: int, imag: int) -> None:
+        old_real, old_imag = out.get(index, (0, 0))
+        out[index] = (old_real + real, old_imag + imag)
+
+    for index, (real, imag) in state.items():
+        add(index, real - imag, real + imag)
+        add(index ^ q, real + imag, imag - real)
+    return {index: value for index, value in out.items() if value != (0, 0)}
+
+
+def xor_group(labels: list[int], group: list[int]) -> int:
+    shift = 0
+    for port in group:
+        shift ^= labels[port]
+    return shift
+
+
+def verify_partition(candidate: dict) -> list[list[int]]:
+    groups = candidate["groups"]
+    residual = candidate["residual"]
+    blocks = candidate["blocks"]
+    need(blocks == groups + [residual], "block order differs from the cover")
+    need(len(groups) == 22 and all(len(g) == 5 for g in groups), "expected 22 five-frames")
+    need(len(residual) == 2, "expected one residual pair")
+    flat = [port for group in blocks for port in group]
+    need(sorted(flat) == list(range(112)), "blocks do not partition P112")
+    return blocks
+
+
+def main() -> None:
+    # Direct runs enforce the package's live source pins too.
+    source_manifest = json.loads((PACKAGE / "SOURCES.json").read_text())
+    for row in source_manifest["files"]:
+        if row["usage"] == "input":
+            path = ROOT / row["path"]
+            need(path.is_file(), "missing input: " + row["path"])
+            need(hashlib.sha256(path.read_bytes()).hexdigest() == row["sha256"],
+                 "input hash mismatch: " + row["path"])
+
+    core = load_predecessor_core()
+    labels, _ = core.load_sources()
+    record = json.loads((PREDECESSOR / "CANDIDATE.json").read_text())
+    candidate = record["candidate"]
+    blocks = verify_partition(candidate)
+    need(len(labels) == 112, "wrong selected label count")
+
+    block_deltas = [xor_group(labels, group) for group in blocks]
+    all_labels_xor = 0
+    for label in labels:
+        all_labels_xor ^= label
+    all_deltas_xor = 0
+    for delta in block_deltas:
+        all_deltas_xor ^= delta
+    need(all_labels_xor == 0, "selected P112 labels no longer have zero xor")
+    need(all_deltas_xor == all_labels_xor, "partitioned correction xor differs from label xor")
+    need(core.rank_rows(block_deltas, core.DIM) == 9, "expected full-rank block correction set")
+    expected_deltas = [98, 70, 100, 274, 50, 506, 247, 352, 304, 56, 19, 274,
+                       22, 445, 152, 388, 268, 254, 463, 508, 322, 475, 235]
+    need(block_deltas == expected_deltas, "per-block endpoint translations changed")
+
+    # Each local C_E^2 translation is checked on every basis address (512 states).
+    local_cases = 0
+    for group, delta in zip(blocks, block_deltas):
+        for start in range(1 << core.DIM):
+            state = {start: (1, 0)}
+            for port in group + group:
+                state = apply_c(state, labels[port])
+            expected = {start ^ delta: (1 << (2 * len(group)), 0)}
+            need(state == expected, "local C_E^2 endpoint identity failed")
+            local_cases += 1
+    need(local_cases == 23 * (1 << core.DIM), "wrong local endpoint case count")
+
+    # If these corrections share one address register, their product is one
+    # translation by xor(block_deltas), hence identity on every address state.
+    for start in range(1 << core.DIM):
+        destination = start
+        for delta in block_deltas:
+            destination ^= delta
+        need(destination == start, "deferred global endpoint correction is nonidentity")
+
+    hist = {int(rank): int(count) for rank, count in candidate["copy_rank_histogram"].items()}
+    profile, centre = core.profile_for(hist)
+    fused_screen = core.moment_screen(profile, centre)
+    need(fused_screen == candidate["moment_screen_u5"], "predecessor local screen changed")
+    separate_profile = profile.copy()
+    separate_profile[1] = separate_profile.get(1, 0) + len(blocks)
+    separate_screen = core.moment_screen(separate_profile, centre + len(blocks))
+    need(separate_screen == candidate["one_unit_per_aggregate_translation_screen"],
+         "separate-correction screen changed")
+    need(Q(fused_screen["ceiling_upper"]) > Q(separate_screen["ceiling_upper"]),
+         "deferring corrections did not recover local ceiling")
+    overhead = candidate["all_in_rank1_overhead_threshold"]
+    max_overhead = overhead["largest_passing"]["extra_rank1_equivalent_charge"]
+    need(max_overhead == 157 and max_overhead - len(blocks) == 134,
+         "rank-one overhead arithmetic changed")
+
+    # Negative controls: one missing block leaves its nonzero translation behind;
+    # a duplicated port violates the exact cover before any algebra is applied.
+    missing_xor = 0
+    for delta in block_deltas[:-1]:
+        missing_xor ^= delta
+    need(missing_xor != 0, "missing-block control did not expose endpoint debt")
+    bad_blocks = [group[:] for group in blocks]
+    bad_blocks[-1][0] = bad_blocks[0][0]
+    bad_flat = [port for group in bad_blocks for port in group]
+    need(sorted(bad_flat) != list(range(112)), "duplicate-port control was accepted")
+
+    print("PASS_GLOBAL_P112_ENDPOINT_CANCELLATION",
+          f"blocks={len(blocks)}", f"local_basis_cases={local_cases}",
+          f"label_xor={all_labels_xor}", f"delta_rank=9",
+          f"separate_a_ceiling={separate_screen['ceiling_upper_decimal']:.12f}",
+          f"deferred_a_ceiling={fused_screen['ceiling_upper_decimal']:.12f}",
+          f"rank1_overhead={max_overhead}",
+          "physical_shared_register=OPEN")
+
+
+if __name__ == "__main__":
+    main()