psyclaw 0.27.22 → 0.28.0

package/vendor/ars/scripts/legacy/check_re_review_synthesis_v1_0.py

@@ -0,0 +1,2215 @@
+#!/usr/bin/env python3
+"""Recompute the #576 Spec B three-gate re-review contract from its artifacts.
+
+Stdlib-only, same architecture class as ``check_panel_synthesis.py`` (#510:
+recompute every layer from the primary artifacts; any mismatch voids the
+synthesis). Normative source:
+``docs/design/2026-07-27-576-spec-b-re-review-precommitment-contract-spec.md``
+§5 (artifact shapes), §6 (decision derivation), §10 (reviewer-label
+normalization grammar), §11 (input manifest + apply-chain witness), §13
+(checker responsibilities).
+
+Responsibilities (spec §13):
+
+1. Schema-validate the three phase artifacts + manifest (self-contained
+   validators, no jsonschema dependency — repo convention).
+2. Hash-chain: ``input_manifest_hash`` -> ``precommitment_hash`` ->
+   ``verdict_record_hash`` verbatim binding; criterion inheritance binding
+   (roadmap_text / letter_text byte comparison + R<n> ordinal recomputation
+   with the §5.1 contiguity degradation).
+3. Consumer-side apply-chain witness (§11) incl. ``patch_digest``
+   content-bound positional pairing and the ``[PATCH-BINDING-ABSENT]``
+   pre-1.2 absence policy.
+4. Recompute invariants: coverage cardinalities, applied_criterion rules,
+   adjustment-chain grammar, deferral-loop referential integrity
+   (reapplications / resolutions / adjudications / acceptances / pending
+   rebuttal upgrades), the G2 deferral biconditional, DecisionInputs
+   operand equality, and the §6 Step 1-3 derivation (Steps 2-3 recomputed
+   BOTH from DecisionInputs AND independently from the raw records).
+5. Goalpost witness: the frozen new-issue set is byte-identical across
+   §5.2/§5.3 and only ``regression`` attribution enters the decision path.
+
+Exit codes (graded, spec §13):
+
+  0  pass
+  1  recomputation mismatch -> ``[RE-REVIEW-ABORT: synthesis_mismatch]``
+     (referential-integrity failures are recomputation mismatches and
+     always exit 1)
+  2  schema or manifest invalid -> the matching §3.5 abort reason
+     (``manifest_incomplete`` / ``manifest_hash_mismatch`` for the
+     manifest layer; ``phase1_lint_failed`` / ``phase2a_lint_failed`` /
+     ``phase2b_lint_failed`` for a phase artifact that fails its schema)
+
+Classification note (documented design decision): the three hash-chain
+links and the manifest-side §11 rules (artifact byte hashes, freshness,
+patch_digest pairing, apply-chain witness ``fail``) are the same-inputs
+proof — their violation is graded exit 2 with ``manifest_hash_mismatch``,
+mirroring §11's "-> manifest_hash_mismatch, G0". Content recomputation on
+top of a sound binding (criterion quotes, letter ordinals, chains,
+decision) is graded exit 1.
+
+Hashing convention: sha256 hex over JSON Canonical Form (RFC 8785 / JCS,
+the v3.6.3 passport convention). For this contract's value domain
+(objects/arrays/strings/integers/booleans/null) Python's
+``json.dumps(obj, sort_keys=True, separators=(",", ":"),
+ensure_ascii=False)`` produces the JCS byte stream.
+
+Usage:
+    python scripts/check_re_review_synthesis.py \
+        --manifest M.json --precommitment P.json \
+        --verdict-record V.json --traceability T.json \
+        --roadmap roadmap.json [--letter letter.md] \
+        [--apply-report R1.json ...]
+
+The ``--roadmap`` / ``--letter`` / ``--apply-report`` files are the
+hash-bound §11 artifacts the checker must READ (criterion binding, ordinal
+derivation, chain witness); every other §11 artifact participates through
+its manifest hash alone. ``--apply-report`` order must follow the
+manifest's ``apply_reports[]`` order.
+"""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+import re
+import sys
+import unicodedata
+from pathlib import Path
+
+EXIT_PASS = 0
+EXIT_SYNTHESIS = 1
+EXIT_INVALID = 2
+
+VERDICTS = ("FULLY_ADDRESSED", "PARTIALLY_ADDRESSED", "NOT_ADDRESSED", "MADE_WORSE", "CANNOT_VERIFY")
+JUDGE_VERDICTS = ("FULLY_ADDRESSED", "PARTIALLY_ADDRESSED", "NOT_ADDRESSED", "MADE_WORSE")
+PRIORITIES = ("must_fix", "should_fix", "consider")
+ROW_PRIORITY_MAP = {"MUST_FIX": "must_fix", "SHOULD_FIX": "should_fix", "CONSIDER": "consider"}
+SEVERITIES = ("critical", "major", "minor")
+MAGNITUDES = ("must_fix", "should_fix", "consider")
+SEAT_LABELS = ("EIC", "R1", "R2", "R3", "DA")
+VERIFIER_SEATS = ("EIC", "R1", "R2", "R3")
+ADJUSTMENT_BASES = (
+    "author_pointer_located_evidence",
+    "valid_rebuttal",
+    "scope_correction",
+    "user_accepted_fail_closed",
+    "cross_model_adjudication",
+)
+ATTRIBUTIONS = ("regression", "previously_missed", "indeterminate")
+DECISION_STATES = ("Accept", "Minor Revision", "Major Revision", "user_review_required", "aborted")
+ABORT_REASONS = (
+    "phase1_lint_failed",
+    "phase2a_lint_failed",
+    "phase2b_lint_failed",
+    "manifest_incomplete",
+    "manifest_hash_mismatch",
+    "criteria_drift",
+    "synthesis_mismatch",
+)
+WITNESS_STATES = ("pass", "fail", "first_link_not_run", "not_run_no_reports")
+DECISION_ORDER = {"Accept": 0, "Minor Revision": 1, "Major Revision": 2}
+# §5.3 mechanical 5->4 map (MADE_WORSE -> NO: the concern stands unresolved —
+# verification happened, outcome negative).
+VERIFIED_MAP = {
+    "FULLY_ADDRESSED": "YES",
+    "PARTIALLY_ADDRESSED": "PARTIAL",
+    "NOT_ADDRESSED": "NO",
+    "MADE_WORSE": "NO",
+    "CANNOT_VERIFY": "CANNOT_VERIFY",
+}
+ESCALATION_CLASSES = ("research_integrity", "ethics", "safety", "legal_compliance", "fatal_validity")
+
+_SHA256_RE = re.compile(r"^[0-9a-f]{64}$")
+_HASH12_RE = re.compile(r"^[0-9a-f]{12}$")
+_REPORT_VERSION_RE = re.compile(r"^[0-9]+(\.[0-9]+)*$")
+_ITEM_ID_RE = re.compile(r"^REV-[A-Za-z0-9-]+$")
+_CONCERN_ID_RE = re.compile(r"^[RSN][1-9][0-9]*$")
+_LETTER_REF_RE = re.compile(r"^R([1-9][0-9]*)$")
+_ANCHOR_RE = re.compile(r"^(text|table|figure|equation|dataset|absence): .+$")
+_ARTIFACT_REF_RE = re.compile(r"^(passport|path):.+$")
+_ANSWER_REF_RE = re.compile(r"^(adjudication:DIS-[1-9][0-9]*|intent:INT-[1-9][0-9]*)$")
+_CRITERION_REF_RE = re.compile(r"^(phase1:REV-[A-Za-z0-9-]+|dissent:DIS-[1-9][0-9]*)$")
+_SOURCE_REF_RE = re.compile(r"^(reapplication:RAP-[1-9][0-9]*|acceptance:ACC-[1-9][0-9]*)$")
+_CHECK_ADJUDICATED_RE = re.compile(r"^adjudicated:(RADJ-[1-9][0-9]*)$")
+_ID_RES = {
+    "NS": re.compile(r"^NS-[1-9][0-9]*$"),
+    "NEW": re.compile(r"^NEW-[1-9][0-9]*$"),
+    "DIS": re.compile(r"^DIS-[1-9][0-9]*$"),
+    "ESC": re.compile(r"^ESC-[1-9][0-9]*$"),
+    "ADJ": re.compile(r"^ADJ-[1-9][0-9]*$"),
+    "INT": re.compile(r"^INT-[1-9][0-9]*$"),
+    "RES": re.compile(r"^RES-[1-9][0-9]*$"),
+    "RADJ": re.compile(r"^RADJ-[1-9][0-9]*$"),
+    "ACC": re.compile(r"^ACC-[1-9][0-9]*$"),
+    "RAP": re.compile(r"^RAP-[1-9][0-9]*$"),
+    "PRB": re.compile(r"^PRB-[1-9][0-9]*$"),
+}
+
+
+class ManifestError(Exception):
+    """Manifest-layer failure -> exit 2 with a §3.5 G0 reason."""
+
+    def __init__(self, reason: str, message: str):
+        self.reason = reason
+        super().__init__(message)
+
+
+class ArtifactSchemaError(Exception):
+    """Phase-artifact schema failure -> exit 2 with the phase lint reason."""
+
+    def __init__(self, reason: str, message: str):
+        self.reason = reason
+        super().__init__(message)
+
+
+def canonical_hash(obj) -> str:
+    """sha256 hex over the object's JSON Canonical Form (RFC 8785 / JCS)."""
+    payload = json.dumps(obj, sort_keys=True, separators=(",", ":"), ensure_ascii=False)
+    return hashlib.sha256(payload.encode("utf-8")).hexdigest()
+
+
+# --- §10 normalization grammar -------------------------------------------------
+
+_PAREN_RE = re.compile(r"\([^()]*\)")
+_SEPARATOR_RE = re.compile(r",|/|;| and |&")
+_R_TOKEN_RES = (
+    re.compile(r"^r([1-3])$"),
+    re.compile(r"^reviewer ([1-3])$"),
+    re.compile(r"^peer reviewer ([1-3])$"),
+)
+
+
+def normalize_reviewer_labels(raw: str):
+    """§10 emitter-side normalization grammar, in declared order.
+
+    (1a) strip every parenthetical group FIRST; (1b) THEN strip any
+    " — <suffix>" tail from the residual (order normative and
+    non-commutative); (2) split on {",", "/", ";", " and ", "&"};
+    (3) trim whitespace; case-fold. A stripped token normalizes by
+    WHOLE-TOKEN EXACT MATCH (no substring extraction); unrecognized
+    tokens are DROPPED, never guessed. Returns ``(labels, parse_failure)``
+    where labels is the deduplicated, original-order result and
+    ``parse_failure`` is True when a non-empty input produced zero
+    recognized tokens (§10: routes to EIC but counts as unmapped — a
+    distinct state from a genuinely empty list).
+    """
+    text = unicodedata.normalize("NFC", raw)
+    prev = None
+    while prev != text:  # nested groups collapse outside-in
+        prev = text
+        text = _PAREN_RE.sub("", text)
+    dash_idx = text.find(" — ")
+    if dash_idx != -1:
+        text = text[:dash_idx]
+    labels = []
+    for token in _SEPARATOR_RE.split(text):
+        token = token.strip().casefold()
+        if not token:
+            continue
+        label = None
+        if token in ("eic", "editor-in-chief"):
+            label = "EIC"
+        elif token in ("da", "devil's advocate", "devil’s advocate"):
+            label = "DA"
+        else:
+            for pattern in _R_TOKEN_RES:
+                match = pattern.match(token)
+                if match:
+                    label = "R" + match.group(1)
+                    break
+        if label is not None and label not in labels:
+            labels.append(label)
+    parse_failure = bool(raw.strip()) and not labels
+    return labels, parse_failure
+
+
+# --- validation helpers --------------------------------------------------------
+
+
+def _is_str(value, nonempty=True) -> bool:
+    return isinstance(value, str) and (not nonempty or len(value) > 0)
+
+
+class _V:
+    """Accumulating validator for one artifact; raises on first failure."""
+
+    def __init__(self, reason: str, artifact_name: str):
+        self.reason = reason
+        self.artifact = artifact_name
+
+    def fail(self, path: str, message: str):
+        raise ArtifactSchemaError(self.reason, f"{self.artifact}: {path}: {message}")
+
+    def obj(self, value, path, required, optional=()):
+        if not isinstance(value, dict):
+            self.fail(path, "must be an object")
+        allowed = set(required) | set(optional)
+        for key in value:
+            if key not in allowed:
+                self.fail(path, f"unexpected field {key!r}")
+        for key in required:
+            if key not in value:
+                self.fail(path, f"missing required field {key!r}")
+
+    def string(self, value, path, nonempty=True):
+        if not _is_str(value, nonempty):
+            self.fail(path, "must be a non-empty string" if nonempty else "must be a string")
+
+    def enum(self, value, path, allowed):
+        if value not in allowed:
+            self.fail(path, f"must be one of {sorted(allowed)!r}, got {value!r}")
+
+    def pattern(self, value, path, regex, label):
+        if not isinstance(value, str) or not regex.match(value):
+            self.fail(path, f"must match {label}, got {value!r}")
+
+    def integer(self, value, path, minimum=None, maximum=None):
+        if not isinstance(value, int) or isinstance(value, bool):
+            self.fail(path, "must be an integer")
+        if minimum is not None and value < minimum:
+            self.fail(path, f"must be >= {minimum}")
+        if maximum is not None and value > maximum:
+            self.fail(path, f"must be <= {maximum}")
+
+    def array(self, value, path):
+        if not isinstance(value, list):
+            self.fail(path, "must be an array")
+
+
+def _validate_residual_gap(v: _V, gap, path):
+    v.obj(gap, path, ["text", "residual_magnitude"])
+    v.string(gap["text"], f"{path}.text")
+    v.enum(gap["residual_magnitude"], f"{path}.residual_magnitude", MAGNITUDES)
+
+
+def _validate_tagged_anchors(v: _V, anchors, path):
+    v.array(anchors, path)
+    if not anchors:
+        v.fail(path, "must carry at least one anchor")
+    for i, entry in enumerate(anchors):
+        v.obj(entry, f"{path}[{i}]", ["anchor", "anchor_artifact"])
+        v.pattern(entry["anchor"], f"{path}[{i}].anchor", _ANCHOR_RE, "the Schema 6 typed-anchor grammar")
+        v.enum(entry["anchor_artifact"], f"{path}[{i}].anchor_artifact", ("manuscript", "letter"))
+
+
+def _validate_new_issue(v: _V, rec, path):
+    v.obj(rec, path, [
+        "new_issue_id", "description", "location_anchor", "severity", "found_by",
+        "confidence", "competence_basis", "attribution", "attribution_evidence",
+        "nearest_roadmap_item", "non_match_rationale",
+    ])
+    v.pattern(rec["new_issue_id"], f"{path}.new_issue_id", _ID_RES["NEW"], "NEW-<n>")
+    v.string(rec["description"], f"{path}.description")
+    v.pattern(rec["location_anchor"], f"{path}.location_anchor", _ANCHOR_RE, "typed anchor")
+    v.enum(rec["severity"], f"{path}.severity", SEVERITIES)
+    v.enum(rec["found_by"], f"{path}.found_by", VERIFIER_SEATS)
+    v.integer(rec["confidence"], f"{path}.confidence", 1, 5)
+    v.string(rec["competence_basis"], f"{path}.competence_basis")
+    v.enum(rec["attribution"], f"{path}.attribution", ATTRIBUTIONS)
+    v.string(rec["attribution_evidence"], f"{path}.attribution_evidence")
+    nearest = rec["nearest_roadmap_item"]
+    if nearest is not None:
+        v.pattern(nearest, f"{path}.nearest_roadmap_item", _ITEM_ID_RE, "REV-<id> or null")
+    v.string(rec["non_match_rationale"], f"{path}.non_match_rationale")
+
+
+def validate_precommitment(art: dict):
+    v = _V("phase1_lint_failed", "precommitment")
+    v.obj(art, "$", ["contract_version", "round_id", "input_manifest_hash", "items", "new_standards"])
+    v.enum(art["contract_version"], "$.contract_version", ("1.0",))
+    v.string(art["round_id"], "$.round_id")
+    v.pattern(art["input_manifest_hash"], "$.input_manifest_hash", _SHA256_RE, "sha256 hex")
+    v.array(art["items"], "$.items")
+    for i, rec in enumerate(art["items"]):
+        path = f"$.items[{i}]"
+        v.obj(rec, path, [
+            "item_id", "priority", "inherited_criterion", "operationalization",
+            "expected_change_surface", "equivalence_policy", "source_reviewer",
+            "source_reviewer_labels",
+        ])
+        v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.enum(rec["priority"], f"{path}.priority", ("must_fix", "should_fix"))
+        crit = rec["inherited_criterion"]
+        v.obj(crit, f"{path}.inherited_criterion", ["roadmap_text"], ["letter_text", "letter_item_ref"])
+        v.string(crit["roadmap_text"], f"{path}.inherited_criterion.roadmap_text")
+        has_text = "letter_text" in crit
+        has_ref = "letter_item_ref" in crit
+        if has_text != has_ref:
+            v.fail(f"{path}.inherited_criterion", "letter_text and letter_item_ref are biconditional (§5.1)")
+        if has_text:
+            v.string(crit["letter_text"], f"{path}.inherited_criterion.letter_text")
+            v.pattern(crit["letter_item_ref"], f"{path}.inherited_criterion.letter_item_ref", _LETTER_REF_RE, "R<n>")
+            if rec["priority"] != "must_fix":
+                v.fail(f"{path}.inherited_criterion", "letter fields exist ONLY on must_fix items (§5.1)")
+        op = rec["operationalization"]
+        v.obj(op, f"{path}.operationalization", ["fully_addressed"], ["partially_addressed", "made_worse_discriminator"])
+        v.string(op["fully_addressed"], f"{path}.operationalization.fully_addressed")
+        if rec["priority"] == "must_fix":
+            for field in ("partially_addressed", "made_worse_discriminator"):
+                if field not in op:
+                    v.fail(f"{path}.operationalization", f"must_fix items require {field!r} (§5.1)")
+                v.string(op[field], f"{path}.operationalization.{field}")
+        else:
+            for field in ("partially_addressed", "made_worse_discriminator"):
+                if field in op:
+                    v.fail(f"{path}.operationalization", f"P2 lighter form forbids {field!r} (§5.1)")
+        v.string(rec["expected_change_surface"], f"{path}.expected_change_surface")
+        v.enum(rec["equivalence_policy"], f"{path}.equivalence_policy", ("allowed",))
+        v.string(rec["source_reviewer"], f"{path}.source_reviewer")
+        v.array(rec["source_reviewer_labels"], f"{path}.source_reviewer_labels")
+        seen = set()
+        for j, label in enumerate(rec["source_reviewer_labels"]):
+            v.enum(label, f"{path}.source_reviewer_labels[{j}]", SEAT_LABELS)
+            if label in seen:
+                v.fail(f"{path}.source_reviewer_labels", f"duplicate label {label!r}")
+            seen.add(label)
+    v.array(art["new_standards"], "$.new_standards")
+    ns_ids = set()
+    for i, rec in enumerate(art["new_standards"]):
+        path = f"$.new_standards[{i}]"
+        v.obj(rec, path, ["new_standard_id", "item_id", "standard_text", "why_not_in_round1", "classification"])
+        v.pattern(rec["new_standard_id"], f"{path}.new_standard_id", _ID_RES["NS"], "NS-<n>")
+        if rec["new_standard_id"] in ns_ids:
+            v.fail(path, f"duplicate new_standard_id {rec['new_standard_id']!r}")
+        ns_ids.add(rec["new_standard_id"])
+        if rec["item_id"] != "global":
+            v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, 'REV-<id> or "global"')
+        v.string(rec["standard_text"], f"{path}.standard_text")
+        v.string(rec["why_not_in_round1"], f"{path}.why_not_in_round1")
+        v.enum(rec["classification"], f"{path}.classification", ("advisory", "escalation_requested"))
+
+
+def validate_verdict_record(art: dict):
+    v = _V("phase2a_lint_failed", "verdict_record")
+    v.obj(art, "$", ["round_id", "precommitment_hash", "items", "new_issues", "dissents", "escalation_exceptions"])
+    v.string(art["round_id"], "$.round_id")
+    v.pattern(art["precommitment_hash"], "$.precommitment_hash", _SHA256_RE, "sha256 hex")
+    v.array(art["items"], "$.items")
+    for i, rec in enumerate(art["items"]):
+        path = f"$.items[{i}]"
+        v.obj(rec, path, ["item_id", "verdict", "change_summary", "verified_by", "applied_criterion"],
+              ["evidence_anchor", "cannot_verify_reason", "residual_gap"])
+        v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.enum(rec["verdict"], f"{path}.verdict", VERDICTS)
+        if rec["verdict"] == "CANNOT_VERIFY":
+            if "cannot_verify_reason" not in rec or "evidence_anchor" in rec:
+                v.fail(path, "CANNOT_VERIFY carries cannot_verify_reason and no evidence_anchor (§3.3)")
+            v.string(rec["cannot_verify_reason"], f"{path}.cannot_verify_reason")
+        else:
+            if "evidence_anchor" not in rec or "cannot_verify_reason" in rec:
+                v.fail(path, "non-CANNOT_VERIFY verdicts carry evidence_anchor and no cannot_verify_reason (§3.3)")
+            v.array(rec["evidence_anchor"], f"{path}.evidence_anchor")
+            if not rec["evidence_anchor"]:
+                v.fail(f"{path}.evidence_anchor", "must carry at least one anchor")
+            for j, anchor in enumerate(rec["evidence_anchor"]):
+                v.pattern(anchor, f"{path}.evidence_anchor[{j}]", _ANCHOR_RE, "typed anchor")
+        if rec["verdict"] == "PARTIALLY_ADDRESSED":
+            if "residual_gap" not in rec:
+                v.fail(path, "PARTIALLY_ADDRESSED requires residual_gap (§3.3)")
+            _validate_residual_gap(v, rec["residual_gap"], f"{path}.residual_gap")
+        elif "residual_gap" in rec:
+            v.fail(path, "residual_gap is required iff verdict = PARTIALLY_ADDRESSED")
+        v.string(rec["change_summary"], f"{path}.change_summary")
+        v.enum(rec["verified_by"], f"{path}.verified_by", VERIFIER_SEATS)
+        applied = rec["applied_criterion"]
+        if applied not in ("precommitted", "not_precommitted") and not (
+            isinstance(applied, str) and re.match(r"^dissented:DIS-[1-9][0-9]*$", applied)
+        ):
+            v.fail(f"{path}.applied_criterion", f"invalid value {applied!r}")
+    v.array(art["new_issues"], "$.new_issues")
+    for i, rec in enumerate(art["new_issues"]):
+        _validate_new_issue(v, rec, f"$.new_issues[{i}]")
+    v.array(art["dissents"], "$.dissents")
+    for i, rec in enumerate(art["dissents"]):
+        path = f"$.dissents[{i}]"
+        v.obj(rec, path, [
+            "dissent_id", "item_id", "criterion_hash", "reason_code",
+            "original_operationalization", "replacement_operationalization",
+            "evidence", "decision_impact_note",
+        ])
+        v.pattern(rec["dissent_id"], f"{path}.dissent_id", _ID_RES["DIS"], "DIS-<n>")
+        v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.pattern(rec["criterion_hash"], f"{path}.criterion_hash", _SHA256_RE, "sha256 hex")
+        v.enum(rec["reason_code"], f"{path}.reason_code",
+               ("criterion_ambiguous", "criterion_infeasible_as_written", "evidence_surface_moved", "criterion_error"))
+        for field in ("original_operationalization", "replacement_operationalization", "evidence", "decision_impact_note"):
+            v.string(rec[field], f"{path}.{field}")
+    v.array(art["escalation_exceptions"], "$.escalation_exceptions")
+    for i, rec in enumerate(art["escalation_exceptions"]):
+        path = f"$.escalation_exceptions[{i}]"
+        v.obj(rec, path, [
+            "exception_id", "escalation_class", "reason_code", "evidence_anchor",
+            "why_round1_missed_it", "mechanical_decision_impact", "approval_state",
+        ], ["new_standard_ref"])
+        v.pattern(rec["exception_id"], f"{path}.exception_id", _ID_RES["ESC"], "ESC-<n>")
+        if "new_standard_ref" in rec:
+            v.pattern(rec["new_standard_ref"], f"{path}.new_standard_ref", _ID_RES["NS"], "NS-<n>")
+        v.enum(rec["escalation_class"], f"{path}.escalation_class", ESCALATION_CLASSES)
+        v.string(rec["reason_code"], f"{path}.reason_code")
+        v.pattern(rec["evidence_anchor"], f"{path}.evidence_anchor", _ANCHOR_RE, "typed anchor")
+        v.string(rec["why_round1_missed_it"], f"{path}.why_round1_missed_it")
+        v.enum(rec["mechanical_decision_impact"], f"{path}.mechanical_decision_impact", ("Minor Revision", "Major Revision"))
+        v.enum(rec["approval_state"], f"{path}.approval_state", ("pending",))
+
+
+def _validate_adjustment_body(v: _V, rec, path, *, drafted: bool):
+    required = ["item_id", "from_verdict", "to_verdict", "basis", "rationale"]
+    optional = ["evidence_anchor", "cannot_verify_reason", "residual_gap"]
+    if drafted:
+        pass  # drafted bodies carry no booking-assigned fields
+    else:
+        required.insert(0, "adjustment_id")
+        optional += ["critical_rebuttal_check", "source_ref", "supersedes_adjustment_id"]
+    v.obj(rec, path, required, optional)
+    if not drafted:
+        v.pattern(rec["adjustment_id"], f"{path}.adjustment_id", _ID_RES["ADJ"], "ADJ-<n>")
+        if "supersedes_adjustment_id" in rec:
+            v.pattern(rec["supersedes_adjustment_id"], f"{path}.supersedes_adjustment_id", _ID_RES["ADJ"], "ADJ-<n>")
+    v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+    v.enum(rec["from_verdict"], f"{path}.from_verdict", VERDICTS)
+    v.enum(rec["to_verdict"], f"{path}.to_verdict", VERDICTS)
+    if drafted:
+        v.enum(rec["basis"], f"{path}.basis", ("valid_rebuttal",))
+    else:
+        v.enum(rec["basis"], f"{path}.basis", ADJUSTMENT_BASES)
+    v.string(rec["rationale"], f"{path}.rationale")
+    if rec["to_verdict"] == "CANNOT_VERIFY":
+        if "cannot_verify_reason" not in rec or "evidence_anchor" in rec:
+            v.fail(path, "to_verdict CANNOT_VERIFY carries cannot_verify_reason and no evidence_anchor")
+        v.string(rec["cannot_verify_reason"], f"{path}.cannot_verify_reason")
+        if rec["basis"] != "user_accepted_fail_closed":
+            v.fail(path, "user_accepted_fail_closed is the ONLY basis that may land on CANNOT_VERIFY (§3.4)")
+    else:
+        if "evidence_anchor" not in rec or "cannot_verify_reason" in rec:
+            v.fail(path, "non-CANNOT_VERIFY adjustments carry evidence_anchor and no cannot_verify_reason")
+        _validate_tagged_anchors(v, rec["evidence_anchor"], f"{path}.evidence_anchor")
+        if rec["basis"] in ("author_pointer_located_evidence", "scope_correction"):
+            for j, anchor in enumerate(rec["evidence_anchor"]):
+                if anchor["anchor_artifact"] != "manuscript":
+                    v.fail(f"{path}.evidence_anchor[{j}]",
+                           f"basis {rec['basis']} requires manuscript-side anchors (§5.3)")
+    if rec["to_verdict"] == "PARTIALLY_ADDRESSED":
+        if "residual_gap" not in rec:
+            v.fail(path, "to_verdict PARTIALLY_ADDRESSED requires residual_gap (§5.3)")
+        _validate_residual_gap(v, rec["residual_gap"], f"{path}.residual_gap")
+    elif "residual_gap" in rec:
+        v.fail(path, "residual_gap is required iff to_verdict = PARTIALLY_ADDRESSED")
+    # §3.4 Direction column. valid_rebuttal — the only basis allowing
+    # letter-side anchors — is "upgrade to FULLY_ADDRESSED"; anything else
+    # would let a persuasive letter move a verdict sideways/downward, the
+    # exact channel §1 exists to close. author_pointer_located_evidence is
+    # an UPGRADE: located evidence satisfies the Phase-1 operationalization
+    # (fully or partially), so the target is PARTIALLY/FULLY and strictly
+    # better than the source (MADE_WORSE/CANNOT_VERIFY rank lowest).
+    if rec["basis"] == "valid_rebuttal":
+        if rec["to_verdict"] != "FULLY_ADDRESSED" or rec["from_verdict"] == "FULLY_ADDRESSED":
+            v.fail(path, "valid_rebuttal adjustments upgrade to FULLY_ADDRESSED (§3.4 Direction)")
+    elif rec["basis"] == "author_pointer_located_evidence":
+        rank = {"MADE_WORSE": 0, "CANNOT_VERIFY": 0, "NOT_ADDRESSED": 1, "PARTIALLY_ADDRESSED": 2, "FULLY_ADDRESSED": 3}
+        if rec["to_verdict"] not in ("PARTIALLY_ADDRESSED", "FULLY_ADDRESSED") or rank[rec["to_verdict"]] <= rank[rec["from_verdict"]]:
+            v.fail(path, "author_pointer_located_evidence adjustments are upgrades to PARTIALLY/FULLY_ADDRESSED (§3.4 Direction)")
+    if drafted:
+        return
+    source_ref = rec.get("source_ref")
+    if rec["basis"] == "cross_model_adjudication":
+        if not (isinstance(source_ref, str) and source_ref.startswith("reapplication:")):
+            v.fail(path, 'basis cross_model_adjudication requires source_ref "reapplication:<id>"')
+    elif rec["basis"] == "user_accepted_fail_closed":
+        if not (isinstance(source_ref, str) and source_ref.startswith("acceptance:")):
+            v.fail(path, 'basis user_accepted_fail_closed requires source_ref "acceptance:<id>"')
+    elif source_ref is not None:
+        v.fail(path, f"source_ref is FORBIDDEN on basis {rec['basis']} (§5.3)")
+    if source_ref is not None:
+        v.pattern(source_ref, f"{path}.source_ref", _SOURCE_REF_RE, "reapplication:RAP-<n> | acceptance:ACC-<n>")
+    check = rec.get("critical_rebuttal_check")
+    if check is not None:
+        if rec["basis"] != "valid_rebuttal":
+            v.fail(path, "critical_rebuttal_check exists only on valid_rebuttal adjustments (§3.4)")
+        if check not in ("single_family_disclosed", "pass_unavailable_disclosed") and not _CHECK_ADJUDICATED_RE.match(check):
+            v.fail(f"{path}.critical_rebuttal_check", f"invalid value {check!r}")
+
+
+def validate_traceability(art: dict):
+    v = _V("phase2b_lint_failed", "traceability")
+    v.obj(art, "$", [
+        "round_id", "revision", "verdict_record_hash", "rows", "adjustments",
+        "new_issues", "post_letter_observations", "dissent_adjudications",
+        "resolution_intents", "cross_model_resolutions", "rebuttal_adjudications",
+        "g2d_acceptances", "pending_rebuttal_upgrades", "escalation_approvals",
+        "reapplications", "decision_inputs", "decision_state",
+    ], ["supersedes_hash", "abort_reason"])
+    v.string(art["round_id"], "$.round_id")
+    v.integer(art["revision"], "$.revision", minimum=1)
+    if art["revision"] == 1:
+        if "supersedes_hash" in art:
+            v.fail("$", "revision 1 must not carry supersedes_hash")
+    else:
+        if "supersedes_hash" not in art:
+            v.fail("$", "revision >= 2 requires supersedes_hash")
+        v.pattern(art["supersedes_hash"], "$.supersedes_hash", _SHA256_RE, "sha256 hex")
+    v.pattern(art["verdict_record_hash"], "$.verdict_record_hash", _SHA256_RE, "sha256 hex")
+    v.enum(art["decision_state"], "$.decision_state", DECISION_STATES)
+    if art["decision_state"] == "aborted":
+        if "abort_reason" not in art:
+            v.fail("$", "decision_state aborted requires abort_reason")
+        v.enum(art["abort_reason"], "$.abort_reason", ABORT_REASONS)
+    elif "abort_reason" in art:
+        v.fail("$", "abort_reason is required iff decision_state = aborted")
+
+    v.array(art["rows"], "$.rows")
+    for i, row in enumerate(art["rows"]):
+        path = f"$.rows[{i}]"
+        v.obj(row, path, [
+            "item_id", "concern_id", "priority", "original_comment", "authors_claim",
+            "revision_location", "verified", "status", "quality_assessment",
+            "final_verdict", "phase2a_verdict", "verified_by",
+        ], ["adjustment_id", "addressed_by_rebuttal", "cross_model_status", "cross_model_verdict"])
+        v.pattern(row["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.pattern(row["concern_id"], f"{path}.concern_id", _CONCERN_ID_RE, "R<n>/S<n>/N<n>")
+        v.enum(row["priority"], f"{path}.priority", tuple(ROW_PRIORITY_MAP))
+        for field in ("original_comment", "authors_claim", "revision_location", "quality_assessment"):
+            v.string(row[field], f"{path}.{field}")
+        v.enum(row["verified"], f"{path}.verified", ("YES", "PARTIAL", "NO", "CANNOT_VERIFY"))
+        v.enum(row["status"], f"{path}.status", VERDICTS)
+        v.enum(row["final_verdict"], f"{path}.final_verdict", VERDICTS)
+        v.enum(row["phase2a_verdict"], f"{path}.phase2a_verdict", VERDICTS)
+        v.enum(row["verified_by"], f"{path}.verified_by", VERIFIER_SEATS)
+        if "adjustment_id" in row:
+            v.pattern(row["adjustment_id"], f"{path}.adjustment_id", _ID_RES["ADJ"], "ADJ-<n>")
+        if "addressed_by_rebuttal" in row and row["addressed_by_rebuttal"] is not True:
+            v.fail(f"{path}.addressed_by_rebuttal", "when present the marker is the literal true")
+        has_status = "cross_model_status" in row
+        has_cmv = "cross_model_verdict" in row
+        if row["priority"] == "MUST_FIX":
+            if not has_status:
+                v.fail(path, "MUST_FIX rows always carry cross_model_status (§5.3 / #539)")
+        elif has_status or has_cmv:
+            v.fail(path, "SHOULD_FIX/CONSIDER rows carry neither cross-model field (#539 scope)")
+        if has_status:
+            v.enum(row["cross_model_status"], f"{path}.cross_model_status",
+                   ("agree", "diverges", "unavailable", "not_configured"))
+            if row["cross_model_status"] in ("agree", "diverges"):
+                if not has_cmv:
+                    v.fail(path, "evaluated (agree/diverges) rows carry cross_model_verdict")
+                v.enum(row["cross_model_verdict"], f"{path}.cross_model_verdict", JUDGE_VERDICTS)
+            elif has_cmv:
+                v.fail(path, "unavailable/not_configured rows carry no cross_model_verdict")
+
+    v.array(art["adjustments"], "$.adjustments")
+    for i, rec in enumerate(art["adjustments"]):
+        _validate_adjustment_body(v, rec, f"$.adjustments[{i}]", drafted=False)
+
+    v.array(art["new_issues"], "$.new_issues")
+    for i, rec in enumerate(art["new_issues"]):
+        _validate_new_issue(v, rec, f"$.new_issues[{i}]")
+
+    v.array(art["post_letter_observations"], "$.post_letter_observations")
+    for i, entry in enumerate(art["post_letter_observations"]):
+        v.string(entry, f"$.post_letter_observations[{i}]")
+
+    v.array(art["dissent_adjudications"], "$.dissent_adjudications")
+    for i, rec in enumerate(art["dissent_adjudications"]):
+        path = f"$.dissent_adjudications[{i}]"
+        v.obj(rec, path, ["dissent_id", "adjudicator", "outcome", "rationale"])
+        v.pattern(rec["dissent_id"], f"{path}.dissent_id", _ID_RES["DIS"], "DIS-<n>")
+        v.enum(rec["adjudicator"], f"{path}.adjudicator", ("cross_model", "user"))
+        v.enum(rec["outcome"], f"{path}.outcome", ("replacement_approved", "original_upheld"))
+        v.string(rec["rationale"], f"{path}.rationale")
+
+    v.array(art["resolution_intents"], "$.resolution_intents")
+    for i, rec in enumerate(art["resolution_intents"]):
+        path = f"$.resolution_intents[{i}]"
+        v.obj(rec, path, ["intent_id", "item_id", "answered_by"], ["guidance_note"])
+        v.pattern(rec["intent_id"], f"{path}.intent_id", _ID_RES["INT"], "INT-<n>")
+        v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.enum(rec["answered_by"], f"{path}.answered_by", ("system", "user"))
+        if "guidance_note" in rec:
+            if rec["answered_by"] == "system":
+                v.fail(path, "guidance_note exists only on user intents (§5.3)")
+            v.string(rec["guidance_note"], f"{path}.guidance_note")
+
+    v.array(art["cross_model_resolutions"], "$.cross_model_resolutions")
+    for i, rec in enumerate(art["cross_model_resolutions"]):
+        path = f"$.cross_model_resolutions[{i}]"
+        v.obj(rec, path, ["resolution_id", "item_id", "intent_id", "reapplication_id", "state", "resolved_by", "rationale"])
+        v.pattern(rec["resolution_id"], f"{path}.resolution_id", _ID_RES["RES"], "RES-<n>")
+        v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.pattern(rec["intent_id"], f"{path}.intent_id", _ID_RES["INT"], "INT-<n>")
+        v.pattern(rec["reapplication_id"], f"{path}.reapplication_id", _ID_RES["RAP"], "RAP-<n>")
+        v.enum(rec["state"], f"{path}.state", ("primary_upheld", "primary_revised"))
+        v.enum(rec["resolved_by"], f"{path}.resolved_by", ("system", "user"))
+        if rec["resolved_by"] == "user" and rec["state"] != "primary_upheld":
+            v.fail(path, "the user acceptance form is always primary_upheld — the user never hand-sets a verdict (§5.3)")
+        v.string(rec["rationale"], f"{path}.rationale")
+
+    v.array(art["rebuttal_adjudications"], "$.rebuttal_adjudications")
+    for i, rec in enumerate(art["rebuttal_adjudications"]):
+        path = f"$.rebuttal_adjudications[{i}]"
+        v.obj(rec, path, ["rebuttal_adjudication_id", "item_id", "verdict", "rationale"])
+        v.pattern(rec["rebuttal_adjudication_id"], f"{path}.rebuttal_adjudication_id", _ID_RES["RADJ"], "RADJ-<n>")
+        v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.enum(rec["verdict"], f"{path}.verdict", ("upheld", "challenged"))
+        v.string(rec["rationale"], f"{path}.rationale")
+
+    v.array(art["g2d_acceptances"], "$.g2d_acceptances")
+    for i, rec in enumerate(art["g2d_acceptances"]):
+        path = f"$.g2d_acceptances[{i}]"
+        v.obj(rec, path, ["acceptance_id", "item_id", "reapplication_id", "accepted_by"])
+        v.pattern(rec["acceptance_id"], f"{path}.acceptance_id", _ID_RES["ACC"], "ACC-<n>")
+        v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.pattern(rec["reapplication_id"], f"{path}.reapplication_id", _ID_RES["RAP"], "RAP-<n>")
+        v.enum(rec["accepted_by"], f"{path}.accepted_by", ("user",))
+
+    v.array(art["pending_rebuttal_upgrades"], "$.pending_rebuttal_upgrades")
+    for i, rec in enumerate(art["pending_rebuttal_upgrades"]):
+        path = f"$.pending_rebuttal_upgrades[{i}]"
+        v.obj(rec, path, ["proposal_id", "item_id", "drafted_adjustment", "disposition"])
+        v.pattern(rec["proposal_id"], f"{path}.proposal_id", _ID_RES["PRB"], "PRB-<n>")
+        v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        _validate_adjustment_body(v, rec["drafted_adjustment"], f"{path}.drafted_adjustment", drafted=True)
+        disposition = rec["disposition"]
+        if not isinstance(disposition, str) or not re.match(
+            r"^(booked:ADJ-[1-9][0-9]*|challenged:RADJ-[1-9][0-9]*|booked_disclosed:ADJ-[1-9][0-9]*|booked_single_family:ADJ-[1-9][0-9]*)$",
+            disposition,
+        ):
+            v.fail(f"{path}.disposition", f"invalid disposition {disposition!r}")
+
+    v.array(art["escalation_approvals"], "$.escalation_approvals")
+    for i, rec in enumerate(art["escalation_approvals"]):
+        path = f"$.escalation_approvals[{i}]"
+        v.obj(rec, path, ["exception_id", "approval_state", "approved_by"])
+        v.pattern(rec["exception_id"], f"{path}.exception_id", _ID_RES["ESC"], "ESC-<n>")
+        v.enum(rec["approval_state"], f"{path}.approval_state", ("approved", "rejected"))
+        v.enum(rec["approved_by"], f"{path}.approved_by", ("user",))
+
+    v.array(art["reapplications"], "$.reapplications")
+    for i, rec in enumerate(art["reapplications"]):
+        path = f"$.reapplications[{i}]"
+        v.obj(rec, path, [
+            "reapplication_id", "item_id", "answer_refs", "pre_reapplication_verdict",
+            "reapplied_verdict", "rationale", "criterion_ref",
+        ], ["supersedes_reapplication_id", "evidence_anchor", "cannot_verify_reason", "residual_gap"])
+        v.pattern(rec["reapplication_id"], f"{path}.reapplication_id", _ID_RES["RAP"], "RAP-<n>")
+        v.pattern(rec["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.array(rec["answer_refs"], f"{path}.answer_refs")
+        if not rec["answer_refs"]:
+            v.fail(f"{path}.answer_refs", "must carry at least one tagged ref")
+        seen_refs = set()
+        for j, ref in enumerate(rec["answer_refs"]):
+            v.pattern(ref, f"{path}.answer_refs[{j}]", _ANSWER_REF_RE, "adjudication:DIS-<n> | intent:INT-<n>")
+            if ref in seen_refs:
+                v.fail(f"{path}.answer_refs", f"duplicate ref {ref!r}")
+            seen_refs.add(ref)
+        if "supersedes_reapplication_id" in rec:
+            v.pattern(rec["supersedes_reapplication_id"], f"{path}.supersedes_reapplication_id", _ID_RES["RAP"], "RAP-<n>")
+        v.enum(rec["pre_reapplication_verdict"], f"{path}.pre_reapplication_verdict", VERDICTS)
+        v.enum(rec["reapplied_verdict"], f"{path}.reapplied_verdict", VERDICTS)
+        if rec["reapplied_verdict"] == "CANNOT_VERIFY":
+            if "cannot_verify_reason" not in rec or "evidence_anchor" in rec:
+                v.fail(path, "CANNOT_VERIFY carries cannot_verify_reason and no evidence_anchor")
+            v.string(rec["cannot_verify_reason"], f"{path}.cannot_verify_reason")
+        else:
+            if "evidence_anchor" not in rec or "cannot_verify_reason" in rec:
+                v.fail(path, "non-CANNOT_VERIFY reapplications carry evidence_anchor and no cannot_verify_reason")
+            _validate_tagged_anchors(v, rec["evidence_anchor"], f"{path}.evidence_anchor")
+        if rec["reapplied_verdict"] == "PARTIALLY_ADDRESSED":
+            if "residual_gap" not in rec:
+                v.fail(path, "reapplied PARTIALLY_ADDRESSED requires residual_gap")
+            _validate_residual_gap(v, rec["residual_gap"], f"{path}.residual_gap")
+        elif "residual_gap" in rec:
+            v.fail(path, "residual_gap is required iff reapplied_verdict = PARTIALLY_ADDRESSED")
+        v.string(rec["rationale"], f"{path}.rationale")
+        v.pattern(rec["criterion_ref"], f"{path}.criterion_ref", _CRITERION_REF_RE, "phase1:REV-<id> | dissent:DIS-<n>")
+
+    di = art["decision_inputs"]
+    v.obj(di, "$.decision_inputs", [
+        "per_item", "verdict_counts", "residual_magnitude_counts", "p2_addressed_rate",
+        "regressions", "non_regression_new_issue_ids", "escalations", "apply_chain_witness",
+    ], ["reject_recommended"])
+    v.array(di["per_item"], "$.decision_inputs.per_item")
+    for i, entry in enumerate(di["per_item"]):
+        path = f"$.decision_inputs.per_item[{i}]"
+        v.obj(entry, path, ["item_id", "final_verdict", "driving_severity"], ["residual_magnitude"])
+        v.pattern(entry["item_id"], f"{path}.item_id", _ITEM_ID_RE, "REV-<id>")
+        v.enum(entry["final_verdict"], f"{path}.final_verdict", VERDICTS)
+        if entry["driving_severity"] is not None:
+            v.enum(entry["driving_severity"], f"{path}.driving_severity", SEVERITIES)
+        if "residual_magnitude" in entry:
+            v.enum(entry["residual_magnitude"], f"{path}.residual_magnitude", MAGNITUDES)
+    for bucket_name in ("verdict_counts", "residual_magnitude_counts"):
+        bucket = di[bucket_name]
+        v.obj(bucket, f"$.decision_inputs.{bucket_name}", list(PRIORITIES))
+        keys = VERDICTS if bucket_name == "verdict_counts" else MAGNITUDES
+        for prio in PRIORITIES:
+            v.obj(bucket[prio], f"$.decision_inputs.{bucket_name}.{prio}", list(keys))
+            for key in keys:
+                v.integer(bucket[prio][key], f"$.decision_inputs.{bucket_name}.{prio}.{key}", minimum=0)
+    rate = di["p2_addressed_rate"]
+    v.obj(rate, "$.decision_inputs.p2_addressed_rate", ["numerator", "denominator"])
+    v.integer(rate["numerator"], "$.decision_inputs.p2_addressed_rate.numerator", minimum=0)
+    v.integer(rate["denominator"], "$.decision_inputs.p2_addressed_rate.denominator", minimum=0)
+    v.array(di["regressions"], "$.decision_inputs.regressions")
+    for i, entry in enumerate(di["regressions"]):
+        path = f"$.decision_inputs.regressions[{i}]"
+        v.obj(entry, path, ["new_issue_id", "severity"])
+        v.pattern(entry["new_issue_id"], f"{path}.new_issue_id", _ID_RES["NEW"], "NEW-<n>")
+        v.enum(entry["severity"], f"{path}.severity", SEVERITIES)
+    v.array(di["non_regression_new_issue_ids"], "$.decision_inputs.non_regression_new_issue_ids")
+    for i, entry in enumerate(di["non_regression_new_issue_ids"]):
+        v.pattern(entry, f"$.decision_inputs.non_regression_new_issue_ids[{i}]", _ID_RES["NEW"], "NEW-<n>")
+    v.array(di["escalations"], "$.decision_inputs.escalations")
+    for i, entry in enumerate(di["escalations"]):
+        path = f"$.decision_inputs.escalations[{i}]"
+        v.obj(entry, path, ["exception_id", "effective_approval_state", "escalation_class", "mechanical_decision_impact"])
+        v.pattern(entry["exception_id"], f"{path}.exception_id", _ID_RES["ESC"], "ESC-<n>")
+        v.enum(entry["effective_approval_state"], f"{path}.effective_approval_state", ("pending", "approved", "rejected"))
+        v.enum(entry["escalation_class"], f"{path}.escalation_class", ESCALATION_CLASSES)
+        v.enum(entry["mechanical_decision_impact"], f"{path}.mechanical_decision_impact", ("Minor Revision", "Major Revision"))
+    if "reject_recommended" in di and not isinstance(di["reject_recommended"], bool):
+        v.fail("$.decision_inputs.reject_recommended", "must be a boolean")
+    v.enum(di["apply_chain_witness"], "$.decision_inputs.apply_chain_witness", WITNESS_STATES)
+
+
+# --- manifest ------------------------------------------------------------------
+
+
+def validate_manifest(manifest: dict):
+    """Structural manifest validation (§11). Failures -> manifest_incomplete."""
+
+    def fail(message: str):
+        raise ManifestError("manifest_incomplete", f"manifest: {message}")
+
+    if not isinstance(manifest, dict):
+        fail("must be an object")
+    expected_top = {"round_id", "cross_model_active", "artifacts"}
+    if set(manifest) != expected_top:
+        fail(f"top-level fields must be exactly {sorted(expected_top)}")
+    if not _is_str(manifest["round_id"]):
+        fail("round_id must be a non-empty string")
+    if not isinstance(manifest["cross_model_active"], bool):
+        fail("cross_model_active must be a boolean")
+    artifacts = manifest["artifacts"]
+    single_keys = (
+        "original_manuscript", "revised_manuscript", "revision_roadmap",
+        "editorial_decision_letter", "response_to_reviewers",
+        "round1_findings", "round1_config_cards",
+    )
+    array_keys = ("revision_patches", "apply_reports")
+    if not isinstance(artifacts, dict) or set(artifacts) != set(single_keys) | set(array_keys):
+        fail("artifacts must carry exactly the nine §11 keys")
+
+    def check_entry_fields(entry, key, path):
+        ref = entry.get("path_or_passport_ref")
+        if not _is_str(ref) or not _ARTIFACT_REF_RE.match(ref):
+            fail(f"{path}: path_or_passport_ref must be passport:<ref> | path:<relative path>")
+        if ref.startswith("path:"):
+            rel = ref[len("path:"):]
+            segments = re.split(r"[/\\]", rel)
+            if (
+                rel.startswith(("/", "\\"))
+                or re.match(r"^[A-Za-z]:", rel)
+                or any(segment == ".." for segment in segments)
+                or not rel
+            ):
+                fail(f"{path}: path: refs carry a RELATIVE path — no absolute, drive-letter, or traversal segments (§11)")
+        sha = entry.get("sha256")
+        if not isinstance(sha, str) or not _SHA256_RE.match(sha):
+            fail(f"{path}: sha256 must be 64-hex")
+        for field in ("version_label", "origin_date"):
+            value = entry.get(field, "MISSING")
+            if value == "MISSING":
+                fail(f"{path}: missing {field}")
+            if value is not None and not _is_str(value):
+                fail(f"{path}: {field} must be a non-empty string or null")
+            if value is None and entry["path_or_passport_ref"].startswith("passport:"):
+                fail(f"{path}: null {field} on a passport:-tagged entry (§11)")
+
+    for key in single_keys:
+        entry = artifacts[key]
+        if not isinstance(entry, dict) or not isinstance(entry.get("present"), bool):
+            fail(f"artifacts.{key}: present-discriminated union required")
+        if entry["present"]:
+            if set(entry) != {"present", "path_or_passport_ref", "sha256", "version_label", "origin_date"}:
+                fail(f"artifacts.{key}: present entry carries exactly ref/sha256/version_label/origin_date")
+            check_entry_fields(entry, key, f"artifacts.{key}")
+        elif set(entry) != {"present"}:
+            fail(f"artifacts.{key}: an absent artifact carries NO ref, hash, or freshness fields (§11)")
+    for key in array_keys:
+        entry = artifacts[key]
+        if not isinstance(entry, dict) or not isinstance(entry.get("present"), bool):
+            fail(f"artifacts.{key}: present-discriminated union required")
+        if entry["present"]:
+            if set(entry) != {"present", "items"}:
+                fail(f"artifacts.{key}: present array entry carries exactly items")
+            items = entry["items"]
+            if not isinstance(items, list) or not items:
+                fail(f"artifacts.{key}: present: false is the canonical empty encoding (§11)")
+            for i, element in enumerate(items):
+                if not isinstance(element, dict) or set(element) != {"path_or_passport_ref", "sha256", "version_label", "origin_date"}:
+                    fail(f"artifacts.{key}[{i}]: every element FULLY populated (no per-element absence)")
+                check_entry_fields(element, key, f"artifacts.{key}[{i}]")
+        elif set(entry) != {"present"}:
+            fail(f"artifacts.{key}: an absent artifact carries NO ref, hash, or freshness fields (§11)")
+
+    # Required set + array pairing (G0).
+    for key in ("revised_manuscript", "revision_roadmap"):
+        if not artifacts[key]["present"]:
+            raise ManifestError("manifest_incomplete", f"manifest: hard-required artifact {key} absent (§11)")
+    patches = artifacts["revision_patches"]
+    reports = artifacts["apply_reports"]
+    if reports["present"]:
+        if not patches["present"]:
+            raise ManifestError("manifest_incomplete", "manifest: apply_reports present with revision_patches absent (§11)")
+        if len(reports["items"]) != len(patches["items"]):
+            raise ManifestError("manifest_incomplete", "manifest: apply_reports / revision_patches length mismatch (§11)")
+
+    # Freshness: a roadmap older than the letter it should pair with fails
+    # closed (§11). Comparable only when both origin_dates are non-null.
+    roadmap_entry = artifacts["revision_roadmap"]
+    letter_entry = artifacts["editorial_decision_letter"]
+    if letter_entry["present"] and roadmap_entry["origin_date"] is not None and letter_entry["origin_date"] is not None:
+        if roadmap_entry["origin_date"] < letter_entry["origin_date"]:
+            raise ManifestError(
+                "manifest_hash_mismatch",
+                "manifest: revision_roadmap origin_date predates editorial_decision_letter (freshness, §11)",
+            )
+
+
+def _hash_prefix(sha256_hex: str) -> str:
+    """§11: apply reports use the 12-hex base_draft_hash format."""
+    return sha256_hex[:12]
+
+
+def compute_apply_chain_witness(manifest: dict, report_payloads):
+    """§11 apply-chain witness over the ordered report payloads.
+
+    Returns (witness, notes). A broken checked link raises ManifestError
+    (manifest_hash_mismatch, G0) — positive breakage evidence outranks the
+    absence markers, matching the declared precedence
+    fail > not_run_no_reports > first_link_not_run > pass.
+    """
+    artifacts = manifest["artifacts"]
+    notes = []
+    if not report_payloads:
+        return "not_run_no_reports", notes
+    patches = artifacts["revision_patches"]["items"]
+    for i, report in enumerate(report_payloads):
+        if not isinstance(report, dict):
+            raise ManifestError("manifest_incomplete", f"apply report [{i}]: must be a JSON object")
+        for field in ("report_format_version", "base_draft_hash", "output_draft_hash"):
+            if not _is_str(report.get(field)):
+                raise ManifestError("manifest_incomplete", f"apply report [{i}]: missing {field}")
+        if not _REPORT_VERSION_RE.match(report["report_format_version"]):
+            raise ManifestError(
+                "manifest_incomplete",
+                f"apply report [{i}]: report_format_version {report['report_format_version']!r} is not a numeric "
+                "dotted version — the pre-1.2 absence policy applies only to a VALID explicit version below 1.2 (§11)",
+            )
+        for field in ("base_draft_hash", "output_draft_hash"):
+            if not _HASH12_RE.match(report[field]):
+                raise ManifestError(
+                    "manifest_incomplete",
+                    f"apply report [{i}]: {field} must be the 12-hex base_draft_hash format (§11)",
+                )
+        version = tuple(int(part) for part in report["report_format_version"].split("."))
+        if version >= (1, 2):
+            digest = report.get("patch_digest")
+            if not isinstance(digest, str) or not _SHA256_RE.match(digest):
+                raise ManifestError("manifest_incomplete", f"apply report [{i}]: format >= 1.2 requires patch_digest")
+            if digest != patches[i]["sha256"]:
+                raise ManifestError(
+                    "manifest_hash_mismatch",
+                    f"apply report [{i}]: patch_digest does not equal the paired revision_patches[{i}].sha256 (§11 content-bound pairing)",
+                )
+        else:
+            notes.append(f"[PATCH-BINDING-ABSENT: report format < 1.2] (apply report [{i}])")
+    first_link_not_run = False
+    original = artifacts["original_manuscript"]
+    if original["present"]:
+        if report_payloads[0]["base_draft_hash"] != _hash_prefix(original["sha256"]):
+            raise ManifestError(
+                "manifest_hash_mismatch",
+                "apply-chain witness: first report base_draft_hash does not equal the original_manuscript hash prefix (§11)",
+            )
+    else:
+        first_link_not_run = True
+    for i in range(1, len(report_payloads)):
+        if report_payloads[i]["base_draft_hash"] != report_payloads[i - 1]["output_draft_hash"]:
+            raise ManifestError(
+                "manifest_hash_mismatch",
+                f"apply-chain witness: report [{i}] base_draft_hash does not equal report [{i - 1}] output_draft_hash (§11)",
+            )
+    revised = artifacts["revised_manuscript"]
+    if report_payloads[-1]["output_draft_hash"] != _hash_prefix(revised["sha256"]):
+        raise ManifestError(
+            "manifest_hash_mismatch",
+            "apply-chain witness: last report output_draft_hash does not equal the revised_manuscript hash prefix (§11)",
+        )
+    return ("first_link_not_run" if first_link_not_run else "pass"), notes
+
+
+# --- letter + roadmap parsing --------------------------------------------------
+
+_LETTER_SECTION_RE = re.compile(r"^### Required Item Details\s*$")
+_LETTER_BLOCK_RE = re.compile(r"^\*\*(R[1-9][0-9]*): .*\*\*\s*$")
+_LETTER_CRITERIA_RE = re.compile(r"^- \*\*Acceptance criteria\*\*: (.*)$")
+
+
+def parse_letter_blocks(text: str):
+    """Extract the letter's Required Item Details blocks in document order.
+
+    Returns an ordered list of (rid, acceptance_text) where acceptance_text
+    is the single-line payload after the ``- **Acceptance criteria**:``
+    bullet, or None when the block carries no such bullet (a block without
+    an Acceptance-criteria field offers nothing to inherit — treated as
+    no-block-at-ordinal for the §5.1 letter-fields rule; its id still
+    participates in the contiguity check).
+    """
+    lines = text.splitlines()
+    blocks = []
+    in_section = False
+    current = None
+    for line in lines:
+        if _LETTER_SECTION_RE.match(line):
+            in_section = True
+            continue
+        if not in_section:
+            continue
+        if line.startswith("## "):
+            break
+        header = _LETTER_BLOCK_RE.match(line)
+        if header:
+            current = [header.group(1), None]
+            blocks.append(current)
+            continue
+        if current is not None and current[1] is None:
+            criteria = _LETTER_CRITERIA_RE.match(line)
+            if criteria:
+                current[1] = criteria.group(1)
+    return [(rid, text_value) for rid, text_value in blocks]
+
+
+def letter_ordinals_sound(blocks, must_fix_count: int) -> bool:
+    """§5.1 contiguity: ids EXACTLY the contiguous sequence R1..Rk, k <= n.
+
+    A gap, duplicate, or extra id (beyond the must_fix count) makes the
+    ordinal derivation unsound and degrades the WHOLE letter layer to
+    absent; a mere count shortfall (contiguous prefix) does not.
+    """
+    for i, (rid, _text) in enumerate(blocks):
+        if rid != f"R{i + 1}":
+            return False
+    return len(blocks) <= must_fix_count
+
+
+def load_roadmap(payload: dict):
+    """Minimal Schema 7 machine-form validation for checker consumption."""
+
+    def fail(message: str):
+        raise ManifestError("manifest_incomplete", f"revision_roadmap: {message}")
+
+    if not isinstance(payload, dict) or not isinstance(payload.get("items"), list):
+        fail("must be an object with an items array (Schema 7)")
+    items = []
+    seen = set()
+    for i, item in enumerate(payload["items"]):
+        if not isinstance(item, dict):
+            fail(f"items[{i}] must be an object")
+        for field in ("id", "priority", "verification_criteria", "reviewer"):
+            if not _is_str(item.get(field)):
+                fail(f"items[{i}].{field} must be a non-empty string")
+        if not _ITEM_ID_RE.match(item["id"]):
+            fail(f"items[{i}].id must be REV-<id>")
+        if item["id"] in seen:
+            fail(f"duplicate roadmap id {item['id']!r}")
+        seen.add(item["id"])
+        if item["priority"] not in PRIORITIES:
+            fail(f"items[{i}].priority must be one of {PRIORITIES}")
+        severity = item.get("severity")
+        if severity is not None and severity not in SEVERITIES:
+            fail(f"items[{i}].severity must be one of {SEVERITIES} when present")
+        source_kind = item.get("source_kind")
+        if source_kind is not None and source_kind not in ("question", "editorial"):
+            fail(f"items[{i}].source_kind must be question|editorial when present")
+        items.append(item)
+    return items
+
+
+# --- synthesis recomputation ---------------------------------------------------
+
+
+class Failures:
+    def __init__(self):
+        self.messages = []
+
+    def add(self, message: str):
+        self.messages.append(message)
+
+
+def _decision_max(*decisions):
+    return max(decisions, key=lambda d: DECISION_ORDER[d])
+
+
+def derive_decision(p1_items, p2_partial_magnitudes, p2_made_worse, rate_num, rate_den, regression_severities, escalations):
+    """§6 Steps 2-3. ``p1_items`` = [(final_verdict, driving_severity, residual_magnitude)].
+
+    Returns (decision, reject_recommended, base_rule).
+    """
+    p1_count = len(p1_items)
+    rate_below_80 = rate_den > 0 and rate_num * 5 < rate_den * 4  # < 80%, integer-exact
+    base = None
+    rule = None
+    if any(v == "MADE_WORSE" and sev == "critical" for v, sev, _m in p1_items) or ("critical" in regression_severities):
+        base, rule = "Major Revision", "B1"
+    elif p1_count > 0 and 2 * sum(1 for v, _s, _m in p1_items if v in ("NOT_ADDRESSED", "MADE_WORSE")) >= p1_count:
+        base, rule = "Major Revision", "B2"
+    elif any(v in ("NOT_ADDRESSED", "MADE_WORSE", "CANNOT_VERIFY") for v, _s, _m in p1_items) or ("major" in regression_severities):
+        base, rule = "Major Revision", "B3"
+    elif any(v == "PARTIALLY_ADDRESSED" and m == "must_fix" for v, _s, m in p1_items) or ("must_fix" in p2_partial_magnitudes):
+        base, rule = "Major Revision", "B4"
+    elif (
+        any(v == "PARTIALLY_ADDRESSED" and m in ("should_fix", "consider") for v, _s, m in p1_items)
+        or rate_below_80
+        or p2_made_worse
+        or ("minor" in regression_severities)
+    ):
+        base, rule = "Minor Revision", "B5"
+    else:
+        base, rule = "Accept", "B6"
+    decision = base
+    reject = rule in ("B1", "B2")
+    for esc in escalations:
+        if esc["effective_approval_state"] == "approved":
+            decision = _decision_max(decision, esc["mechanical_decision_impact"])
+            if esc["escalation_class"] == "research_integrity":
+                reject = True
+    return decision, reject, rule
+
+
+def _adjustment_chains(traceability, fails: Failures):
+    """Build per-item ordered adjustment chains (§5.3 grammar).
+
+    Returns {item_id: [records head..tail]} (empty dict entries omitted);
+    reports every grammar violation into ``fails``.
+    """
+    by_id = {}
+    for rec in traceability["adjustments"]:
+        if rec["adjustment_id"] in by_id:
+            fails.add(f"adjustment ids must be unique: {rec['adjustment_id']}")
+            return {}
+        by_id[rec["adjustment_id"]] = rec
+    chains = {}
+    by_item = {}
+    for rec in traceability["adjustments"]:
+        by_item.setdefault(rec["item_id"], []).append(rec)
+    for item_id, records in by_item.items():
+        superseded = {}
+        heads = []
+        for rec in records:
+            prev = rec.get("supersedes_adjustment_id")
+            if prev is None:
+                heads.append(rec)
+            else:
+                prev_rec = by_id.get(prev)
+                if prev_rec is None or prev_rec["item_id"] != item_id:
+                    fails.add(f"adjustment {rec['adjustment_id']}: supersedes_adjustment_id {prev!r} does not resolve on item {item_id}")
+                    return {}
+                if prev in superseded:
+                    fails.add(f"adjustment chain on {item_id}: {prev} superseded twice")
+                    return {}
+                superseded[prev] = rec
+        if len(heads) != 1:
+            fails.add(f"adjustment chain on {item_id}: exactly one head required, found {len(heads)}")
+            return {}
+        chain = [heads[0]]
+        while chain[-1]["adjustment_id"] in superseded:
+            chain.append(superseded[chain[-1]["adjustment_id"]])
+        if len(chain) != len(records):
+            fails.add(f"adjustment chain on {item_id}: records outside the linear chain (cycle or fork)")
+            return {}
+        for prev_rec, next_rec in zip(chain, chain[1:]):
+            if next_rec["from_verdict"] != prev_rec["to_verdict"]:
+                fails.add(
+                    f"adjustment chain on {item_id}: {next_rec['adjustment_id']}.from_verdict "
+                    f"!= {prev_rec['adjustment_id']}.to_verdict (head-to-tail join)"
+                )
+        chains[item_id] = chain
+    return chains
+
+
+def check(manifest, precommitment, verdict_record, traceability, roadmap_items, letter_blocks, witness, witness_notes):
+    """All §13 recomputation on schema-valid, hash-bound inputs.
+
+    Returns (failures, warnings, recomputed_outcome) where
+    recomputed_outcome is the Step 1-3 emission expectation.
+    """
+    fails = Failures()
+    warnings = list(witness_notes)
+
+    roadmap_by_id = {item["id"]: item for item in roadmap_items}
+    must_fix_order = [item["id"] for item in roadmap_items if item["priority"] == "must_fix"]
+    p12_ids = {item["id"] for item in roadmap_items if item["priority"] in ("must_fix", "should_fix")}
+    cross_model_active = manifest["cross_model_active"]
+
+    # round_id equality across all artifacts.
+    for name, art in (("precommitment", precommitment), ("verdict_record", verdict_record), ("traceability", traceability)):
+        if art["round_id"] != manifest["round_id"]:
+            fails.add(f"{name}.round_id != manifest.round_id")
+
+    # --- Phase 1 coverage + criterion binding (§4/§5.1) ---
+    pre_by_item = {}
+    for rec in precommitment["items"]:
+        if rec["item_id"] in pre_by_item:
+            fails.add(f"precommitment: duplicate record for {rec['item_id']}")
+        pre_by_item[rec["item_id"]] = rec
+    if set(pre_by_item) != p12_ids:
+        fails.add(
+            "precommitment coverage: records must exist for exactly the must_fix/should_fix roadmap items "
+            f"(missing {sorted(p12_ids - set(pre_by_item))}, extra {sorted(set(pre_by_item) - p12_ids)})"
+        )
+    letter_present = manifest["artifacts"]["editorial_decision_letter"]["present"]
+    letter_sound = letter_blocks is not None and letter_ordinals_sound(letter_blocks, len(must_fix_order))
+    if letter_present and letter_blocks is not None and not letter_sound:
+        warnings.append("[CRITERIA-LAYER-ABSENT: letter/roadmap ordinal mismatch]")
+    if letter_present and letter_sound and not letter_blocks:
+        warnings.append(
+            "NOTE: editorial decision letter present but no Required Item Details blocks parsed — the level-2 "
+            "criteria layer is empty (template drift or a genuinely block-less letter)"
+        )
+    block_by_rid = {}
+    if letter_sound and letter_blocks:
+        block_by_rid = {rid: text for rid, text in letter_blocks}
+    for item_id, rec in pre_by_item.items():
+        item = roadmap_by_id.get(item_id)
+        if item is None:
+            continue
+        if rec["priority"] != item["priority"]:
+            fails.add(f"precommitment {item_id}: priority {rec['priority']} != roadmap {item['priority']}")
+        if rec["inherited_criterion"]["roadmap_text"] != item["verification_criteria"]:
+            fails.add(f"precommitment {item_id}: inherited_criterion.roadmap_text does not match the roadmap verbatim (§4)")
+        if rec["source_reviewer"] != item["reviewer"]:
+            fails.add(
+                f"precommitment {item_id}: source_reviewer is a VERBATIM copy of the Schema 7 reviewer field "
+                f"({rec['source_reviewer']!r} != {item['reviewer']!r}, §5.1)"
+            )
+        labels, _parse_failure = normalize_reviewer_labels(rec["source_reviewer"])
+        if rec["source_reviewer_labels"] != labels:
+            fails.add(
+                f"precommitment {item_id}: source_reviewer_labels {rec['source_reviewer_labels']!r} "
+                f"!= §10 normalization of source_reviewer ({labels!r})"
+            )
+        has_letter_fields = "letter_text" in rec["inherited_criterion"]
+        if item["priority"] == "must_fix":
+            ordinal = must_fix_order.index(item_id) + 1
+            derived_ref = f"R{ordinal}"
+            block_text = block_by_rid.get(derived_ref) if letter_present and letter_sound else None
+            if block_text is not None:
+                if not has_letter_fields:
+                    fails.add(f"precommitment {item_id}: letter carries block {derived_ref} — letter_text/letter_item_ref REQUIRED (§5.1)")
+                else:
+                    if rec["inherited_criterion"]["letter_item_ref"] != derived_ref:
+                        fails.add(
+                            f"precommitment {item_id}: letter_item_ref {rec['inherited_criterion']['letter_item_ref']!r} "
+                            f"!= derived ordinal {derived_ref!r} (§5.1 — the ref is DERIVED, never chosen)"
+                        )
+                    if rec["inherited_criterion"]["letter_text"] != block_text:
+                        fails.add(f"precommitment {item_id}: letter_text does not match the letter's Acceptance criteria byte-for-byte (§13)")
+            elif has_letter_fields:
+                fails.add(f"precommitment {item_id}: letter layer absent at this item — letter fields must be absent (§5.1/§13)")
+        elif has_letter_fields:
+            fails.add(f"precommitment {item_id}: letter fields on a non-must_fix item (§5.1)")
+
+    # --- Phase 2A coverage + applied_criterion (§5.2) ---
+    verdict_by_item = {}
+    for rec in verdict_record["items"]:
+        if rec["item_id"] in verdict_by_item:
+            fails.add(f"verdict_record: duplicate record for {rec['item_id']}")
+        verdict_by_item[rec["item_id"]] = rec
+    if set(verdict_by_item) != set(roadmap_by_id):
+        fails.add(
+            "verdict_record coverage: exactly one record per roadmap item, ALL priorities "
+            f"(missing {sorted(set(roadmap_by_id) - set(verdict_by_item))}, extra {sorted(set(verdict_by_item) - set(roadmap_by_id))})"
+        )
+    dissent_by_id = {}
+    for rec in verdict_record["dissents"]:
+        if rec["dissent_id"] in dissent_by_id:
+            fails.add(f"verdict_record: duplicate dissent_id {rec['dissent_id']}")
+        dissent_by_id[rec["dissent_id"]] = rec
+        pre = pre_by_item.get(rec["item_id"])
+        if pre is None:
+            fails.add(f"dissent {rec['dissent_id']}: item {rec['item_id']} has no pre-commitment record")
+        elif rec["criterion_hash"] != canonical_hash(pre):
+            fails.add(f"dissent {rec['dissent_id']}: criterion_hash does not recompute from the Phase-1 record (§7)")
+        # Reverse witness (§7): a dissent swaps the criterion BEFORE the
+        # verdict — "The item's verdict record then carries
+        # applied_criterion: dissented:<dissent_id>". A ghost dissent that
+        # was never applied cannot exist, trip the §7 bound, or authorize
+        # an original_upheld re-application of a criterion that was in
+        # fact applied at 2A.
+        applied = verdict_by_item.get(rec["item_id"], {}).get("applied_criterion")
+        if applied != f"dissented:{rec['dissent_id']}":
+            fails.add(
+                f"dissent {rec['dissent_id']}: the item's verdict record must carry "
+                f"applied_criterion dissented:{rec['dissent_id']} — a dissent that was never applied is a ghost (§7)"
+            )
+    # §7 bounds (SD-5): dissent on a P1 item, or dissents on > ceil(N/3) of
+    # all items, trips independent adjudication of EVERY dissent record.
+    total_items = len(roadmap_items)
+    bound_tripped = any(
+        roadmap_by_id.get(dissent["item_id"], {}).get("priority") == "must_fix"
+        for dissent in dissent_by_id.values()
+    ) or (len(dissent_by_id) > math.ceil(total_items / 3))
+    for item_id, rec in verdict_by_item.items():
+        item = roadmap_by_id.get(item_id)
+        if item is None:
+            continue
+        applied = rec["applied_criterion"]
+        if item["priority"] == "consider":
+            if applied != "not_precommitted":
+                fails.add(f"verdict {item_id}: consider items carry applied_criterion not_precommitted (§5.2)")
+        else:
+            if applied == "not_precommitted":
+                fails.add(f"verdict {item_id}: not_precommitted is valid ONLY for consider items (§5.2)")
+            elif applied.startswith("dissented:"):
+                dissent = dissent_by_id.get(applied.split(":", 1)[1])
+                if dissent is None or dissent["item_id"] != item_id:
+                    fails.add(f"verdict {item_id}: applied_criterion {applied!r} does not resolve to a dissent on this item")
+
+    ns_by_id = {rec["new_standard_id"]: rec for rec in precommitment["new_standards"]}
+    if not manifest["artifacts"]["original_manuscript"]["present"] and verdict_record["escalation_exceptions"]:
+        fails.add(
+            "escalation exceptions cannot be substantiated with the original manuscript absent — the required "
+            "original-text anchor cannot be produced, so the new_standard stays advisory "
+            "([ESCALATION-UNSUBSTANTIATABLE], §11 degradation (iii))"
+        )
+    exceptions_by_id = {}
+    for rec in verdict_record["escalation_exceptions"]:
+        if rec["exception_id"] in exceptions_by_id:
+            fails.add(f"verdict_record: duplicate exception_id {rec['exception_id']}")
+        exceptions_by_id[rec["exception_id"]] = rec
+        ref = rec.get("new_standard_ref")
+        if ref is not None:
+            target = ns_by_id.get(ref)
+            if target is None:
+                fails.add(f"escalation exception {rec['exception_id']}: new_standard_ref {ref!r} does not resolve")
+            elif target["classification"] != "escalation_requested":
+                fails.add(f"escalation exception {rec['exception_id']}: new_standard_ref target is not escalation_requested (§5.1)")
+
+    # --- new-issue freeze witness (§13.5) ---
+    if traceability["new_issues"] != verdict_record["new_issues"]:
+        fails.add("new_issues: §5.3 set is not WHOLE-RECORD byte-identical to the frozen §5.2 set (§3.3 freeze witness)")
+    new_issue_by_id = {}
+    for rec in verdict_record["new_issues"]:
+        if rec["new_issue_id"] in new_issue_by_id:
+            fails.add(f"verdict_record: duplicate new_issue_id {rec['new_issue_id']}")
+        new_issue_by_id[rec["new_issue_id"]] = rec
+        nearest = rec["nearest_roadmap_item"]
+        if nearest is not None and nearest not in roadmap_by_id:
+            fails.add(f"new issue {rec['new_issue_id']}: nearest_roadmap_item {nearest!r} does not resolve")
+        if rec["attribution"] == "regression" and nearest is not None:
+            warnings.append(
+                f"ADVISORY: regression {rec['new_issue_id']} names nearest_roadmap_item {nearest} — "
+                "review for roadmap-item overlap (§8)"
+            )
+        if not manifest["artifacts"]["original_manuscript"]["present"] and rec["attribution"] != "indeterminate":
+            fails.add(
+                f"new issue {rec['new_issue_id']}: with the original manuscript absent every attribution is "
+                "indeterminate, never a guess (§11 degradation (i))"
+            )
+
+    # --- rows (§5.3) ---
+    row_by_item = {}
+    for row in traceability["rows"]:
+        if row["item_id"] in row_by_item:
+            fails.add(f"traceability: duplicate row for {row['item_id']}")
+        row_by_item[row["item_id"]] = row
+    if set(row_by_item) != set(roadmap_by_id):
+        fails.add(
+            "row coverage: EVERY roadmap item, all priorities, has exactly one row "
+            f"(missing {sorted(set(roadmap_by_id) - set(row_by_item))}, extra {sorted(set(row_by_item) - set(roadmap_by_id))})"
+        )
+    for item_id, row in row_by_item.items():
+        item = roadmap_by_id.get(item_id)
+        if item is None:
+            continue
+        if ROW_PRIORITY_MAP[row["priority"]] != item["priority"]:
+            fails.add(f"row {item_id}: priority {row['priority']} does not match the roadmap ({item['priority']})")
+        vrec = verdict_by_item.get(item_id)
+        if vrec is not None:
+            if row["phase2a_verdict"] != vrec["verdict"]:
+                fails.add(f"row {item_id}: phase2a_verdict != committed §5.2 verdict")
+            if row["verified_by"] != vrec["verified_by"]:
+                fails.add(f"row {item_id}: verified_by not copied from the §5.2 verdict record")
+        if row["status"] != row["final_verdict"]:
+            fails.add(f"row {item_id}: status must map 1:1 from final_verdict (§5.3)")
+        if row["verified"] != VERIFIED_MAP[row["final_verdict"]]:
+            fails.add(f"row {item_id}: verified must follow the mechanical 5->4 map (§5.3)")
+        if "cross_model_verdict" in row:
+            derived = "agree" if row["cross_model_verdict"] == row["final_verdict"] else "diverges"
+            if row["cross_model_status"] != derived:
+                fails.add(f"row {item_id}: cross_model_status {row['cross_model_status']!r} != per-emission derivation {derived!r} (§5.3)")
+
+    # --- adjustment chains (§5.3) ---
+    chains = _adjustment_chains(traceability, fails)
+    for item_id, row in row_by_item.items():
+        chain = chains.get(item_id)
+        if chain:
+            if row.get("adjustment_id") != chain[-1]["adjustment_id"]:
+                fails.add(f"row {item_id}: adjustment_id must name the chain's LATEST record")
+            if chain[0]["from_verdict"] != row["phase2a_verdict"]:
+                fails.add(f"adjustment chain on {item_id}: head from_verdict != phase2a_verdict")
+            if chain[-1]["to_verdict"] != row["final_verdict"]:
+                fails.add(f"adjustment chain on {item_id}: tail to_verdict != final_verdict")
+        else:
+            if "adjustment_id" in row:
+                fails.add(f"row {item_id}: adjustment_id present with no adjustment chain")
+        tail_rebuttal = bool(chain) and chain[-1]["basis"] == "valid_rebuttal" and chain[-1]["to_verdict"] == "FULLY_ADDRESSED"
+        if bool(row.get("addressed_by_rebuttal")) != tail_rebuttal:
+            fails.add(f"row {item_id}: addressed_by_rebuttal marker must reflect a valid_rebuttal chain-tail upgrade to FULLY_ADDRESSED (§3.4)")
+    for item_id in chains:
+        if item_id not in row_by_item:
+            fails.add(f"adjustment chain on {item_id}: no matrix row exists")
+
+    # --- deferral-loop referential integrity (§5.3/§6) ---
+    intents_by_id = {}
+    for rec in traceability["resolution_intents"]:
+        if rec["intent_id"] in intents_by_id:
+            fails.add(f"duplicate intent_id {rec['intent_id']}")
+        intents_by_id[rec["intent_id"]] = rec
+        if rec["answered_by"] == "system":
+            # §5.3: system intents are emitted for every P1 diverges row —
+            # the row must be an EVALUATED P1 row (cross_model_verdict
+            # present implies an active configuration).
+            intent_row = row_by_item.get(rec["item_id"])
+            if intent_row is None or intent_row["priority"] != "MUST_FIX" or "cross_model_verdict" not in intent_row:
+                fails.add(
+                    f"resolution intent {rec['intent_id']}: system intents exist only for evaluated P1 diverges "
+                    "rows under ACTIVE cross-model (§5.3)"
+                )
+    adjudications_by_dissent = {}
+    for rec in traceability["dissent_adjudications"]:
+        if rec["dissent_id"] in adjudications_by_dissent:
+            fails.add(f"duplicate dissent adjudication for {rec['dissent_id']}")
+        adjudications_by_dissent[rec["dissent_id"]] = rec
+        if rec["dissent_id"] not in dissent_by_id:
+            fails.add(f"dissent adjudication {rec['dissent_id']}: no such dissent record")
+        if rec["adjudicator"] == "cross_model" and not cross_model_active:
+            fails.add(f"dissent adjudication {rec['dissent_id']}: cross_model adjudicator on an inactive configuration (§11)")
+        if not bound_tripped:
+            fails.add(
+                f"dissent adjudication {rec['dissent_id']}: no §7 bound tripped — dissents below the bound stand "
+                "unadjudicated by design (§7)"
+            )
+        dissent = dissent_by_id.get(rec["dissent_id"])
+        dissent_item = roadmap_by_id.get(dissent["item_id"]) if dissent else None
+        if rec["adjudicator"] == "cross_model":
+            if dissent_item is None or dissent_item["priority"] != "must_fix":
+                fails.add(
+                    f"dissent adjudication {rec['dissent_id']}: the judge's adjudication scope EQUALS the §9 pass's "
+                    "P1 coverage — dissents on non-P1 items always take the G2(a) user path, even on an active "
+                    "setup (§7)"
+                )
+        # A `user` adjudication on an active-setup P1 dissent stays LEGAL:
+        # the §6 deferral loop records "a user-adjudicated DissentAdjudication
+        # ... DIRECTLY" without an activity qualifier, and the §9 pass can be
+        # per-row unavailable — rejecting the user path would make a
+        # judge-transport failure unrecoverable (adjudicated against the
+        # stricter one-way reading; see PR #605 round-4/5 record).
+    reapp_by_id = {}
+    for rec in traceability["reapplications"]:
+        if rec["reapplication_id"] in reapp_by_id:
+            fails.add(f"duplicate reapplication_id {rec['reapplication_id']}")
+        reapp_by_id[rec["reapplication_id"]] = rec
+    superseded_reapps = {}
+    for rec in traceability["reapplications"]:
+        prev = rec.get("supersedes_reapplication_id")
+        if prev is not None:
+            prev_rec = reapp_by_id.get(prev)
+            if prev_rec is None or prev_rec["item_id"] != rec["item_id"]:
+                fails.add(f"reapplication {rec['reapplication_id']}: supersedes_reapplication_id {prev!r} does not resolve on the same item")
+            elif prev in superseded_reapps:
+                fails.add(f"reapplication supersession: {prev} superseded twice")
+            else:
+                superseded_reapps[prev] = rec["reapplication_id"]
+    # acyclicity of supersession
+    for rec in traceability["reapplications"]:
+        seen_ids = set()
+        cursor = rec
+        while cursor is not None:
+            if cursor["reapplication_id"] in seen_ids:
+                fails.add(f"reapplication supersession cycle at {cursor['reapplication_id']}")
+                break
+            seen_ids.add(cursor["reapplication_id"])
+            prev = cursor.get("supersedes_reapplication_id")
+            cursor = reapp_by_id.get(prev) if prev is not None else None
+    current_reapps = [rec for rec in traceability["reapplications"] if rec["reapplication_id"] not in superseded_reapps]
+
+    def _chain_tail_verdict(item_id):
+        chain = chains.get(item_id)
+        if chain:
+            return chain[-1]["to_verdict"]
+        row = row_by_item.get(item_id)
+        return row["phase2a_verdict"] if row else None
+
+    for rec in traceability["reapplications"]:
+        item_id = rec["item_id"]
+        has_adjudication_ref = False
+        has_intent_ref = False
+        for ref in rec["answer_refs"]:
+            kind, ref_id = ref.split(":", 1)
+            if kind == "intent":
+                has_intent_ref = True
+            if kind == "adjudication":
+                has_adjudication_ref = True
+                adjudication = adjudications_by_dissent.get(ref_id)
+                dissent = dissent_by_id.get(ref_id)
+                if adjudication is None or dissent is None:
+                    fails.add(f"reapplication {rec['reapplication_id']}: answer ref {ref!r} does not resolve")
+                    continue
+                if adjudication["outcome"] != "original_upheld":
+                    fails.add(f"reapplication {rec['reapplication_id']}: adjudication ref {ref!r} must resolve to original_upheld (§6)")
+                if dissent["item_id"] != item_id:
+                    fails.add(f"reapplication {rec['reapplication_id']}: adjudication ref {ref!r} is on a different item")
+            else:
+                intent = intents_by_id.get(ref_id)
+                if intent is None:
+                    fails.add(f"reapplication {rec['reapplication_id']}: answer ref {ref!r} does not resolve")
+                elif intent["item_id"] != item_id:
+                    fails.add(f"reapplication {rec['reapplication_id']}: intent ref {ref!r} is on a different item")
+        # effective-criterion rule (§6)
+        replacement_dissents = [
+            dissent_id for dissent_id, adjudication in adjudications_by_dissent.items()
+            if adjudication["outcome"] == "replacement_approved"
+            and dissent_by_id.get(dissent_id, {}).get("item_id") == item_id
+        ]
+        if has_adjudication_ref:
+            expected_ref = f"phase1:{item_id}"
+        elif replacement_dissents:
+            expected_ref = f"dissent:{replacement_dissents[0]}"
+        else:
+            expected_ref = f"phase1:{item_id}"
+        if rec["criterion_ref"] != expected_ref:
+            fails.add(
+                f"reapplication {rec['reapplication_id']}: criterion_ref {rec['criterion_ref']!r} "
+                f"violates the §6 effective-criterion rule (expected {expected_ref!r})"
+            )
+        # derived adjustment (§6 loop step 2)
+        derived = [
+            adj for adj in traceability["adjustments"]
+            if adj.get("source_ref") == f"reapplication:{rec['reapplication_id']}"
+        ]
+        # §6 trigger binding: a reapplication whose answer_refs carry ONLY
+        # intent refs is a DIVERGENCE re-verification — it exists only for
+        # an EVALUATED P1 row (one carrying cross_model_verdict, which
+        # itself implies an active configuration): "diverges rows exist
+        # only under ACTIVE cross-model (a not_configured run has none)".
+        # A G2(d) retry is unaffected — its refs are CUMULATIVE and always
+        # include the mandating adjudication.
+        if has_intent_ref and not has_adjudication_ref:
+            row = row_by_item.get(item_id)
+            if row is None or row["priority"] != "MUST_FIX" or "cross_model_verdict" not in row:
+                fails.add(
+                    f"reapplication {rec['reapplication_id']}: a divergence-only re-application exists only for an "
+                    "evaluated P1 row under ACTIVE cross-model (§6/§5.3) — no committed verdict moves without its "
+                    "triggering divergence"
+                )
+            elif row["cross_model_verdict"] == rec["pre_reapplication_verdict"]:
+                # §6 dispatch order is normative: diverges is identified
+                # FIRST (judge verdict vs the row's verdict), THEN the
+                # system intent is emitted and the 2B' call dispatched. A
+                # judge verdict equal to the dispatch-time pre-value means
+                # no divergence existed to resolve — the chain manufactures
+                # its own diverges status after the fact.
+                fails.add(
+                    f"reapplication {rec['reapplication_id']}: no divergence existed at dispatch — the judge's "
+                    "verdict equals the dispatch-time pre_reapplication_verdict (§6)"
+                )
+            if not any(
+                intents_by_id.get(ref.split(":", 1)[1], {}).get("answered_by") == "system"
+                for ref in rec["answer_refs"]
+                if ref.startswith("intent:")
+            ):
+                fails.add(
+                    f"reapplication {rec['reapplication_id']}: a divergence re-application chain carries the "
+                    "ORIGINAL mandating system intent (§6 — divergence resolution is MECHANICAL by default; user "
+                    "intents exist only on the retry paths, cumulatively alongside it)"
+                )
+        # §5.3 letter-tag condition: a letter-tagged anchor on a
+        # reapplication (and hence on its mechanically-copied
+        # cross_model_adjudication adjustment) is valid EXACTLY when the
+        # re-examined chain projection carried a booked valid_rebuttal
+        # record — the §3.4 machine witness that an assertion in the letter
+        # with no locatable manuscript evidence changes nothing. The 2B'
+        # input set withholds the Response Letter itself, so a booked
+        # rebuttal anchor is the only legal letter-side source.
+        if rec["reapplied_verdict"] != "CANNOT_VERIFY" and any(
+            anchor["anchor_artifact"] == "letter" for anchor in rec["evidence_anchor"]
+        ):
+            letter_tag_allowed = False
+            for chain_adj in chains.get(item_id, []):
+                if len(derived) == 1 and chain_adj["adjustment_id"] == derived[0]["adjustment_id"]:
+                    break
+                if chain_adj["basis"] == "valid_rebuttal":
+                    letter_tag_allowed = True
+                    break
+            if not letter_tag_allowed:
+                fails.add(
+                    f"reapplication {rec['reapplication_id']}: letter-tagged anchors are valid exactly when the "
+                    "re-examined chain carries a booked valid_rebuttal record (§5.3/§3.4)"
+                )
+        changing = rec["reapplied_verdict"] != "CANNOT_VERIFY" and rec["reapplied_verdict"] != rec["pre_reapplication_verdict"]
+        if rec["reapplied_verdict"] == "CANNOT_VERIFY":
+            if derived:
+                fails.add(f"reapplication {rec['reapplication_id']}: CANNOT_VERIFY appends NOTHING mechanically (§6)")
+        elif changing:
+            if len(derived) != 1:
+                fails.add(f"reapplication {rec['reapplication_id']}: a verdict-changing reapplication requires exactly one derived adjustment (§13)")
+            else:
+                adj = derived[0]
+                if adj["item_id"] != item_id or adj["basis"] != "cross_model_adjudication":
+                    fails.add(f"reapplication {rec['reapplication_id']}: derived adjustment must be a cross_model_adjudication on the same item")
+                if adj["from_verdict"] != rec["pre_reapplication_verdict"] or adj["to_verdict"] != rec["reapplied_verdict"]:
+                    fails.add(f"reapplication {rec['reapplication_id']}: derived adjustment endpoints must bind to the RECORDED pre-value (§6)")
+                if adj["rationale"] != rec["rationale"]:
+                    fails.add(f"reapplication {rec['reapplication_id']}: derived adjustment rationale must be MECHANICALLY COPIED (§6)")
+                if adj.get("residual_gap") != rec.get("residual_gap"):
+                    fails.add(f"reapplication {rec['reapplication_id']}: derived adjustment residual_gap must be MECHANICALLY COPIED (§6)")
+                if adj.get("evidence_anchor") != rec.get("evidence_anchor"):
+                    fails.add(f"reapplication {rec['reapplication_id']}: derived adjustment anchors must be MECHANICALLY COPIED, tags included (§5.3)")
+        else:
+            if derived:
+                fails.add(f"reapplication {rec['reapplication_id']}: no verdict change, so no derived adjustment may exist (§6)")
+
+    # every original_upheld adjudication + every intent appears in exactly one CURRENT record's answer_refs
+    def _answers_in_current(ref: str):
+        return [rec for rec in current_reapps if ref in rec["answer_refs"]]
+
+    for dissent_id, adjudication in adjudications_by_dissent.items():
+        if adjudication["outcome"] == "original_upheld":
+            holders = _answers_in_current(f"adjudication:{dissent_id}")
+            if len(holders) != 1:
+                fails.add(
+                    f"dissent adjudication {dissent_id}: original_upheld must appear in exactly one CURRENT reapplication's "
+                    f"answer_refs (found {len(holders)}) — silent non-reapplication cannot pass (§6)"
+                )
+    for intent_id in intents_by_id:
+        holders = _answers_in_current(f"intent:{intent_id}")
+        if len(holders) != 1:
+            fails.add(
+                f"resolution intent {intent_id}: must appear in exactly one CURRENT reapplication's answer_refs "
+                f"(found {len(holders)}) (§6)"
+            )
+
+    # --- cross-model resolutions (§5.3/§9) ---
+    resolutions_by_reapp = {}
+    for rec in traceability["cross_model_resolutions"]:
+        if rec["reapplication_id"] in resolutions_by_reapp:
+            fails.add(f"reapplication {rec['reapplication_id']}: referenced by more than one CrossModelResolution")
+        resolutions_by_reapp[rec["reapplication_id"]] = rec
+    # §6 loop step 2: when an intent-mandated re-application derives a state
+    # (reapplied_verdict != CANNOT_VERIFY) the resolution is created
+    # MECHANICALLY — its absence means the pipeline skipped a mechanical step.
+    for rec in current_reapps:
+        has_intent_ref = any(ref.startswith("intent:") for ref in rec["answer_refs"])
+        if has_intent_ref and rec["reapplied_verdict"] != "CANNOT_VERIFY":
+            resolution = resolutions_by_reapp.get(rec["reapplication_id"])
+            if resolution is None or resolution["resolved_by"] != "system":
+                fails.add(
+                    f"reapplication {rec['reapplication_id']}: a derived state mandates a mechanically-created "
+                    "system CrossModelResolution (§6 loop step 2)"
+                )
+    for rec in traceability["cross_model_resolutions"]:
+        intent = intents_by_id.get(rec["intent_id"])
+        reapp = reapp_by_id.get(rec["reapplication_id"])
+        if intent is None or intent["item_id"] != rec["item_id"]:
+            fails.add(f"resolution {rec['resolution_id']}: intent_id does not resolve on the same item")
+        if reapp is None or reapp["item_id"] != rec["item_id"]:
+            fails.add(f"resolution {rec['resolution_id']}: reapplication_id does not resolve on the same item")
+            continue
+        if f"intent:{rec['intent_id']}" not in reapp["answer_refs"]:
+            fails.add(
+                f"resolution {rec['resolution_id']}: reapplication {rec['reapplication_id']} did not answer "
+                f"intent {rec['intent_id']} — the resolution derives FROM the mandated re-application, never from "
+                "an unrelated same-item record (§6 loop step 2)"
+            )
+        if rec["resolved_by"] == "system":
+            if reapp["reapplied_verdict"] == "CANNOT_VERIFY":
+                fails.add(f"resolution {rec['resolution_id']}: a CANNOT_VERIFY reapplication derives no system resolution (§5.3)")
+            else:
+                expected_state = "primary_upheld" if reapp["reapplied_verdict"] == reapp["pre_reapplication_verdict"] else "primary_revised"
+                if rec["state"] != expected_state:
+                    fails.add(f"resolution {rec['resolution_id']}: state {rec['state']!r} != §9 derivation ({expected_state!r})")
+        else:
+            if reapp["reapplied_verdict"] != "CANNOT_VERIFY":
+                fails.add(f"resolution {rec['resolution_id']}: the user acceptance form requires a CANNOT_VERIFY reapplication (§5.3)")
+
+    # --- G2(d) acceptances ---
+    acceptances_by_id = {}
+    for rec in traceability["g2d_acceptances"]:
+        if rec["acceptance_id"] in acceptances_by_id:
+            fails.add(f"duplicate acceptance_id {rec['acceptance_id']}")
+        acceptances_by_id[rec["acceptance_id"]] = rec
+        reapp = reapp_by_id.get(rec["reapplication_id"])
+        if reapp is None or reapp["item_id"] != rec["item_id"]:
+            fails.add(f"acceptance {rec['acceptance_id']}: reapplication_id does not resolve on the same item")
+        elif reapp["reapplied_verdict"] != "CANNOT_VERIFY":
+            fails.add(f"acceptance {rec['acceptance_id']}: the accepted reapplication must be CANNOT_VERIFY (§6 G2(d))")
+        if rec["reapplication_id"] in superseded_reapps:
+            fails.add(
+                f"acceptance {rec['acceptance_id']}: references a SUPERSEDED reapplication — an acceptance binds "
+                "to the CURRENT failed attempt, never a stale one (§6 G2(d))"
+            )
+        row = row_by_item.get(rec["item_id"])
+        if row is not None and row["final_verdict"] != "CANNOT_VERIFY":
+            fails.add(f"acceptance {rec['acceptance_id']}: the accepted item's final_verdict must be CANNOT_VERIFY (§6 G2(d))")
+        backing = [
+            adj for adj in traceability["adjustments"]
+            if adj["basis"] == "user_accepted_fail_closed"
+            and adj.get("source_ref") == f"acceptance:{rec['acceptance_id']}"
+        ]
+        if len(backing) != 1:
+            fails.add(
+                f"acceptance {rec['acceptance_id']}: a G2(d) acceptance backs exactly ONE user_accepted_fail_closed "
+                f"adjustment (found {len(backing)}) — the one legal CANNOT_VERIFY append; an orphan acceptance cannot "
+                "clear the deferral (§6 G2(d)/§3.4)"
+            )
+    acceptance_adj_by_reapp = {}
+    for adj in traceability["adjustments"]:
+        if adj["basis"] == "user_accepted_fail_closed":
+            acc_id = adj["source_ref"].split(":", 1)[1]
+            acceptance = acceptances_by_id.get(acc_id)
+            if acceptance is None or acceptance["item_id"] != adj["item_id"]:
+                fails.add(f"adjustment {adj['adjustment_id']}: acceptance ref does not resolve on the same item")
+                continue
+            acceptance_adj_by_reapp[acceptance["reapplication_id"]] = adj
+            reapp = reapp_by_id.get(acceptance["reapplication_id"])
+            if reapp is not None and reapp.get("cannot_verify_reason") != adj.get("cannot_verify_reason"):
+                fails.add(f"adjustment {adj['adjustment_id']}: cannot_verify_reason must be copied from the accepted reapplication (§3.4)")
+        elif adj["basis"] == "cross_model_adjudication":
+            rap_id = adj["source_ref"].split(":", 1)[1]
+            if rap_id not in reapp_by_id:
+                fails.add(f"adjustment {adj['adjustment_id']}: reapplication ref does not resolve")
+
+    # pre_reapplication_verdict consistency (§13): the recorded pre-value is
+    # the TIME ANCHOR — it equals this reapplication's own derived
+    # adjustment's from_verdict when one exists; on the G2(d) path it equals
+    # the acceptance-backed adjustment's from_verdict (that adjustment
+    # consumed this reapplication, so the tail moved AFTER dispatch); else
+    # the current chain tail; else phase2a_verdict when the item has no chain.
+    for rec in traceability["reapplications"]:
+        derived = [
+            adj for adj in traceability["adjustments"]
+            if adj.get("source_ref") == f"reapplication:{rec['reapplication_id']}"
+        ]
+        if rec["reapplication_id"] in superseded_reapps:
+            # UNCONDITIONAL for every superseded record — a derived
+            # adjustment must not bypass this branch. Supersession is
+            # defined for FAILED attempts only (§6 — "a retry names the
+            # failed attempt it supersedes"), and a failed attempt appends
+            # nothing, so its dispatch tail equals its direct retry's
+            # recorded pre-value. The current-tail fallback never applies
+            # to a historical record (the tail may have moved since).
+            successor = reapp_by_id[superseded_reapps[rec["reapplication_id"]]]
+            if rec["reapplied_verdict"] != "CANNOT_VERIFY":
+                fails.add(
+                    f"reapplication {rec['reapplication_id']}: only FAILED (CANNOT_VERIFY) attempts are superseded "
+                    "by a retry (§6)"
+                )
+            if rec["pre_reapplication_verdict"] != successor["pre_reapplication_verdict"]:
+                fails.add(
+                    f"reapplication {rec['reapplication_id']}: a failed attempt appends nothing, so its dispatch "
+                    f"tail equals its direct retry's pre_reapplication_verdict "
+                    f"({rec['pre_reapplication_verdict']!r} != {successor['pre_reapplication_verdict']!r}, §6/§13)"
+                )
+            continue
+        if len(derived) == 1:
+            expected_pre = derived[0]["from_verdict"]
+        elif rec["reapplication_id"] in acceptance_adj_by_reapp:
+            expected_pre = acceptance_adj_by_reapp[rec["reapplication_id"]]["from_verdict"]
+        else:
+            expected_pre = _chain_tail_verdict(rec["item_id"])
+        if expected_pre is not None and rec["pre_reapplication_verdict"] != expected_pre:
+            fails.add(
+                f"reapplication {rec['reapplication_id']}: pre_reapplication_verdict {rec['pre_reapplication_verdict']!r} "
+                f"!= the chain tail as it stood at dispatch ({expected_pre!r}, §13)"
+            )
+
+    # --- critical-rebuttal machinery (§3.4/§5.3) ---
+    radj_by_id = {}
+    for rec in traceability["rebuttal_adjudications"]:
+        if rec["rebuttal_adjudication_id"] in radj_by_id:
+            fails.add(f"duplicate rebuttal_adjudication_id {rec['rebuttal_adjudication_id']}")
+        radj_by_id[rec["rebuttal_adjudication_id"]] = rec
+        if not cross_model_active:
+            fails.add(f"rebuttal adjudication {rec['rebuttal_adjudication_id']}: the §3.4 pass exists only under active cross-model (§11)")
+    adjustments_by_id = {adj["adjustment_id"]: adj for adj in traceability["adjustments"]}
+
+    def _is_critical(item_id):
+        item = roadmap_by_id.get(item_id)
+        return item is not None and item.get("severity") == "critical"
+
+    for adj in traceability["adjustments"]:
+        should_have_check = adj["basis"] == "valid_rebuttal" and _is_critical(adj["item_id"])
+        has_check = "critical_rebuttal_check" in adj
+        if should_have_check != has_check:
+            fails.add(
+                f"adjustment {adj['adjustment_id']}: critical_rebuttal_check present iff the adjustment is a critical "
+                f"valid_rebuttal (§3.4)"
+            )
+        if has_check:
+            check_value = adj["critical_rebuttal_check"]
+            match = _CHECK_ADJUDICATED_RE.match(check_value) if isinstance(check_value, str) else None
+            if match:
+                radj = radj_by_id.get(match.group(1))
+                if radj is None or radj["verdict"] != "upheld" or radj["item_id"] != adj["item_id"]:
+                    fails.add(f"adjustment {adj['adjustment_id']}: adjudicated:* must resolve to an upheld RebuttalAdjudication on the same item (§3.4)")
+            elif check_value == "single_family_disclosed" and cross_model_active:
+                fails.add(f"adjustment {adj['adjustment_id']}: single_family_disclosed is invalid on an active cross-model run (§3.4)")
+
+    proposals_by_id = {}
+    booked_by_adjustment = {}
+    seen_drafted_bodies = {}
+    challenged_bodies = []
+    for rec in traceability["pending_rebuttal_upgrades"]:
+        if rec["proposal_id"] in proposals_by_id:
+            fails.add(f"duplicate proposal_id {rec['proposal_id']}")
+        proposals_by_id[rec["proposal_id"]] = rec
+        kind, ref_id = rec["disposition"].split(":", 1)
+        if rec["drafted_adjustment"]["item_id"] != rec["item_id"]:
+            fails.add(f"proposal {rec['proposal_id']}: drafted_adjustment.item_id must equal the proposal's item_id (§5.3)")
+        body_key = (rec["item_id"], canonical_hash(rec["drafted_adjustment"]))
+        if body_key in seen_drafted_bodies:
+            fails.add(
+                f"proposal {rec['proposal_id']}: duplicate PendingRebuttalUpgrade drafted body on item {rec['item_id']} "
+                f"(same canonical body as {seen_drafted_bodies[body_key]}) — one proposal per drafted upgrade (§5.3)"
+            )
+        seen_drafted_bodies[body_key] = rec["proposal_id"]
+        if kind == "challenged":
+            challenged_bodies.append((rec["proposal_id"], rec["item_id"], rec["drafted_adjustment"]))
+        if kind == "challenged":
+            radj = radj_by_id.get(ref_id)
+            if radj is None or radj["item_id"] != rec["item_id"]:
+                fails.add(f"proposal {rec['proposal_id']}: challenged ref does not resolve on the same item")
+            elif radj["verdict"] != "challenged":
+                fails.add(f"proposal {rec['proposal_id']}: challenged disposition must reference a challenged adjudication")
+        else:
+            adj = adjustments_by_id.get(ref_id)
+            if adj is None or adj["item_id"] != rec["item_id"] or adj["basis"] != "valid_rebuttal":
+                fails.add(f"proposal {rec['proposal_id']}: booked ref must resolve to a valid_rebuttal adjustment on the same item")
+                continue
+            if ref_id in booked_by_adjustment:
+                fails.add(f"proposal {rec['proposal_id']}: adjustment {ref_id} already booked by {booked_by_adjustment[ref_id]}")
+            booked_by_adjustment[ref_id] = rec["proposal_id"]
+            drafted = dict(rec["drafted_adjustment"])
+            booked = {
+                key: value for key, value in adj.items()
+                if key not in ("adjustment_id", "supersedes_adjustment_id", "critical_rebuttal_check")
+            }
+            if drafted != booked:
+                fails.add(
+                    f"proposal {rec['proposal_id']}: booked adjustment must content-equal the drafted body "
+                    "field-for-field, excluding EXACTLY {adjustment_id, supersedes_adjustment_id, critical_rebuttal_check} (§5.3)"
+                )
+            check_value = adj.get("critical_rebuttal_check")
+            if kind == "booked":
+                match = _CHECK_ADJUDICATED_RE.match(check_value) if isinstance(check_value, str) else None
+                if not match or radj_by_id.get(match.group(1), {}).get("verdict") != "upheld":
+                    fails.add(f"proposal {rec['proposal_id']}: booked disposition requires an upheld adjudicated:* check (§3.4)")
+            elif kind == "booked_disclosed":
+                if check_value != "pass_unavailable_disclosed":
+                    fails.add(f"proposal {rec['proposal_id']}: booked_disclosed requires critical_rebuttal_check pass_unavailable_disclosed")
+                if not cross_model_active:
+                    fails.add(f"proposal {rec['proposal_id']}: booked_disclosed exists only under active cross-model (§11)")
+            elif kind == "booked_single_family":
+                if check_value != "single_family_disclosed":
+                    fails.add(f"proposal {rec['proposal_id']}: booked_single_family requires critical_rebuttal_check single_family_disclosed")
+                if cross_model_active:
+                    fails.add(f"proposal {rec['proposal_id']}: booked_single_family is invalid under active cross-model (§3.4)")
+    for adj in traceability["adjustments"]:
+        if adj["basis"] == "valid_rebuttal" and _is_critical(adj["item_id"]) and adj["adjustment_id"] not in booked_by_adjustment:
+            fails.add(
+                f"adjustment {adj['adjustment_id']}: every critical valid_rebuttal adjustment traces to exactly one "
+                "booked* PendingRebuttalUpgrade (§5.3)"
+            )
+        if adj["basis"] == "valid_rebuttal":
+            booked_view = {
+                key: value for key, value in adj.items()
+                if key not in ("adjustment_id", "supersedes_adjustment_id", "critical_rebuttal_check")
+            }
+            for proposal_id, item_id, drafted in challenged_bodies:
+                if adj["item_id"] == item_id and booked_view == drafted:
+                    fails.add(
+                        f"adjustment {adj['adjustment_id']}: content-equals CHALLENGED proposal {proposal_id}'s "
+                        "drafted body — a challenged upgrade is NEVER booked (§3.4)"
+                    )
+
+    # --- escalation approvals ---
+    approvals_by_exception = {}
+    for rec in traceability["escalation_approvals"]:
+        if rec["exception_id"] in approvals_by_exception:
+            fails.add(f"duplicate escalation approval for {rec['exception_id']}")
+        approvals_by_exception[rec["exception_id"]] = rec
+        if rec["exception_id"] not in exceptions_by_id:
+            fails.add(f"escalation approval {rec['exception_id']}: no such exception record")
+
+    # --- cross-model activity consistency (§11) ---
+    if not cross_model_active:
+        for item_id, row in row_by_item.items():
+            if row.get("cross_model_status") in ("agree", "diverges", "unavailable"):
+                fails.add(f"row {item_id}: cross_model_status {row['cross_model_status']!r} implies an active configuration (§11)")
+
+    # --- DecisionInputs recomputation (§13) ---
+    di = traceability["decision_inputs"]
+
+    def _final_residual_magnitude(item_id):
+        row = row_by_item.get(item_id)
+        if row is None or row["final_verdict"] != "PARTIALLY_ADDRESSED":
+            return None
+        chain = chains.get(item_id)
+        record = chain[-1] if chain else verdict_by_item.get(item_id)
+        if record is None:
+            return None
+        gap = record.get("residual_gap")
+        return gap["residual_magnitude"] if gap else None
+
+    expected_per_item = []
+    for item_id in must_fix_order:
+        row = row_by_item.get(item_id)
+        if row is None:
+            continue
+        item = roadmap_by_id[item_id]
+        severity = item.get("severity")
+        # §5.3: a null driving_severity is legal ONLY for source_kind items
+        # and legacy roadmaps. A P1 item carrying OTHER transported markers
+        # (#574 A2/A3) but no severity is a half-transported finding-driven
+        # item, not a legacy one — B1's critical join would be silently
+        # suppressed, so it fails here.
+        if (
+            severity is None
+            and item.get("source_kind") is None
+            and any(item.get(field) is not None for field in ("evidence_anchor", "confidence", "competence_basis"))
+        ):
+            fails.add(
+                f"roadmap item {item_id}: transported fields present but severity absent — a non-legacy "
+                "finding-driven P1 item cannot carry driving_severity null (§5.3)"
+            )
+        entry = {
+            "item_id": item_id,
+            "final_verdict": row["final_verdict"],
+            "driving_severity": severity if severity in SEVERITIES else None,
+        }
+        magnitude = _final_residual_magnitude(item_id)
+        if magnitude is not None:
+            entry["residual_magnitude"] = magnitude
+        expected_per_item.append(entry)
+    if di["per_item"] != expected_per_item:
+        fails.add("decision_inputs.per_item does not equal the recomputed per-P1 operand list (roadmap must_fix order, §13)")
+
+    expected_counts = {prio: {verdict: 0 for verdict in VERDICTS} for prio in PRIORITIES}
+    expected_magnitudes = {prio: {mag: 0 for mag in MAGNITUDES} for prio in PRIORITIES}
+    for item_id, row in row_by_item.items():
+        item = roadmap_by_id.get(item_id)
+        if item is None:
+            continue
+        expected_counts[item["priority"]][row["final_verdict"]] += 1
+        if row["final_verdict"] == "PARTIALLY_ADDRESSED":
+            magnitude = _final_residual_magnitude(item_id)
+            if magnitude is not None:
+                expected_magnitudes[item["priority"]][magnitude] += 1
+    if di["verdict_counts"] != expected_counts:
+        fails.add("decision_inputs.verdict_counts does not recompute from the final row verdicts (§13)")
+    if di["residual_magnitude_counts"] != expected_magnitudes:
+        fails.add("decision_inputs.residual_magnitude_counts does not recompute from the PARTIALLY_ADDRESSED rows (§13)")
+
+    p2_ids = [item["id"] for item in roadmap_items if item["priority"] == "should_fix"]
+    expected_num = sum(
+        1 for item_id in p2_ids
+        if row_by_item.get(item_id, {}).get("final_verdict") in ("FULLY_ADDRESSED", "PARTIALLY_ADDRESSED")
+    )
+    if di["p2_addressed_rate"] != {"numerator": expected_num, "denominator": len(p2_ids)}:
+        fails.add("decision_inputs.p2_addressed_rate does not recompute over FINAL P2 verdicts (§6)")
+
+    expected_regressions = [
+        {"new_issue_id": rec["new_issue_id"], "severity": rec["severity"]}
+        for rec in verdict_record["new_issues"] if rec["attribution"] == "regression"
+    ]
+    expected_non_regression = [
+        rec["new_issue_id"] for rec in verdict_record["new_issues"] if rec["attribution"] != "regression"
+    ]
+    if di["regressions"] != expected_regressions:
+        fails.add("decision_inputs.regressions does not equal the regression-attributed frozen new issues (§8/§13)")
+    if di["non_regression_new_issue_ids"] != expected_non_regression:
+        fails.add("decision_inputs.non_regression_new_issue_ids does not equal the frozen previously_missed/indeterminate ids (§8/§13)")
+
+    expected_escalations = []
+    for rec in verdict_record["escalation_exceptions"]:
+        approval = approvals_by_exception.get(rec["exception_id"])
+        expected_escalations.append({
+            "exception_id": rec["exception_id"],
+            "effective_approval_state": approval["approval_state"] if approval else "pending",
+            "escalation_class": rec["escalation_class"],
+            "mechanical_decision_impact": rec["mechanical_decision_impact"],
+        })
+    if di["escalations"] != expected_escalations:
+        fails.add("decision_inputs.escalations does not equal the joined exception/approval summary (§5.3)")
+
+    if di["apply_chain_witness"] != witness:
+        fails.add(f"decision_inputs.apply_chain_witness {di['apply_chain_witness']!r} != recomputed witness {witness!r} (§11)")
+
+    # --- §6 Step 1 recomputation (gates) ---
+    g1 = any(
+        row["final_verdict"] != row["phase2a_verdict"] and row["item_id"] not in chains
+        for row in traceability["rows"]
+    )
+    pending = []
+    # bound_tripped computed once, next to dissent_by_id (§7 SD-5).
+    if bound_tripped:
+        for dissent_id in dissent_by_id:
+            if dissent_id not in adjudications_by_dissent:
+                pending.append(f"G2(a): dissent {dissent_id} lacks its DissentAdjudication (bound tripped — covering cardinality)")
+
+    def _covering_resolution(item_id, final_verdict):
+        for rec in traceability["cross_model_resolutions"]:
+            if rec["item_id"] != item_id:
+                continue
+            reapp = reapp_by_id.get(rec["reapplication_id"])
+            if reapp is None:
+                continue
+            anchor = reapp["reapplied_verdict"] if rec["resolved_by"] == "system" else reapp["pre_reapplication_verdict"]
+            if anchor == final_verdict:
+                return True
+        return False
+
+    for item_id, row in row_by_item.items():
+        if row.get("cross_model_status") == "diverges" and ROW_PRIORITY_MAP[row["priority"]] == "must_fix":
+            if not _covering_resolution(item_id, row["final_verdict"]):
+                pending.append(f"G2(b): P1 diverges row {item_id} has no COVERING CrossModelResolution")
+    for rec in verdict_record["escalation_exceptions"]:
+        if rec["exception_id"] not in approvals_by_exception:
+            pending.append(f"G2(c): escalation exception {rec['exception_id']} is pending user approval")
+    accepted_reapps = {rec["reapplication_id"] for rec in traceability["g2d_acceptances"]}
+    for dissent_id, adjudication in adjudications_by_dissent.items():
+        if adjudication["outcome"] != "original_upheld":
+            continue
+        holders = _answers_in_current(f"adjudication:{dissent_id}")
+        if len(holders) != 1:
+            continue  # already a synthesis failure above
+        current = holders[0]
+        if current["reapplied_verdict"] == "CANNOT_VERIFY" and current["reapplication_id"] not in accepted_reapps:
+            pending.append(f"G2(d): original_upheld dissent {dissent_id} fail-closed reapplication awaits user resolution")
+
+    # --- emission-kind-scoped equality set (§13) ---
+    recomputed = {"pending": pending, "g1": g1}
+    emitted_state = traceability["decision_state"]
+    if emitted_state == "aborted":
+        # §13: an aborted emission carries its true root-cause abort_reason
+        # and is EXEMPT from the deferral biconditional and Steps 2-3 (abort
+        # precedence over deferral). The only checker-recomputable reason is
+        # criteria_drift; the phase-lint reasons are fenced-call facts the
+        # checker cannot re-derive and are accepted as recorded (the
+        # manifest-class reasons would already have exited 2 above).
+        reason = traceability.get("abort_reason")
+        recomputed["expect"] = ("aborted", reason)
+        if "reject_recommended" in di:
+            fails.add("decision_inputs.reject_recommended must be ABSENT on a gated emission (§5.3 presence biconditional)")
+        if reason in ("manifest_incomplete", "manifest_hash_mismatch", "synthesis_mismatch"):
+            fails.add(
+                f"aborted emission claims {reason!r} as root cause, but the checker reached recomputation — the "
+                "§11 manifest layer validated against these same hash-bound inputs, so the claimed abort cannot be "
+                "the true root cause (§13: an aborted emission carries its TRUE root-cause abort_reason)"
+            )
+        if g1 and reason != "criteria_drift":
+            fails.add("Step 1 recomputes G1 (silent verdict change) — the aborted emission's true root cause is criteria_drift (§13)")
+        if reason == "criteria_drift" and not g1:
+            fails.add("the emission claims [RE-REVIEW-ABORT: criteria_drift] but no silent verdict change recomputes (§13)")
+    elif g1:
+        recomputed["expect"] = ("aborted", "criteria_drift")
+        fails.add("Step 1 recomputes G1 (silent verdict change) — the emission must be [RE-REVIEW-ABORT: criteria_drift]")
+    elif pending:
+        recomputed["expect"] = ("user_review_required", None)
+        if emitted_state != "user_review_required":
+            fails.add(
+                "G2 deferral biconditional: pending user-input state(s) exist, so user_review_required is the only "
+                "checker-valid decision_state — " + "; ".join(pending)
+            )
+        if "reject_recommended" in di:
+            fails.add("decision_inputs.reject_recommended must be ABSENT on a gated emission (§5.3 presence biconditional)")
+    else:
+        if emitted_state == "user_review_required":
+            fails.add(
+                "Step 1 recomputes no gate, but the emission carries decision_state 'user_review_required' "
+                "(the G2 biconditional runs both directions)"
+            )
+        else:
+            # Steps 2-3, recomputed independently from the raw records...
+            p1_operands = []
+            for item_id in must_fix_order:
+                row = row_by_item.get(item_id)
+                if row is None:
+                    continue
+                item = roadmap_by_id[item_id]
+                severity = item.get("severity") if item.get("severity") in SEVERITIES else None
+                p1_operands.append((row["final_verdict"], severity, _final_residual_magnitude(item_id)))
+            p2_partial_magnitudes = [
+                _final_residual_magnitude(item_id) for item_id in p2_ids
+                if row_by_item.get(item_id, {}).get("final_verdict") == "PARTIALLY_ADDRESSED"
+            ]
+            p2_made_worse = any(
+                row_by_item.get(item_id, {}).get("final_verdict") == "MADE_WORSE" for item_id in p2_ids
+            )
+            regression_severities = {entry["severity"] for entry in expected_regressions}
+            raw_decision, raw_reject, raw_rule = derive_decision(
+                p1_operands, p2_partial_magnitudes, p2_made_worse,
+                expected_num, len(p2_ids), regression_severities, expected_escalations,
+            )
+            # ...AND from the emitted DecisionInputs copies.
+            di_p1 = [
+                (entry["final_verdict"], entry["driving_severity"], entry.get("residual_magnitude"))
+                for entry in di["per_item"]
+            ]
+            di_p2_partials = [
+                mag for mag in MAGNITUDES
+                for _ in range(di["residual_magnitude_counts"]["should_fix"][mag])
+            ]
+            di_decision, di_reject, _di_rule = derive_decision(
+                di_p1,
+                di_p2_partials,
+                di["verdict_counts"]["should_fix"]["MADE_WORSE"] > 0,
+                di["p2_addressed_rate"]["numerator"],
+                di["p2_addressed_rate"]["denominator"],
+                {entry["severity"] for entry in di["regressions"]},
+                di["escalations"],
+            )
+            recomputed["expect"] = (raw_decision, None)
+            recomputed["rule"] = raw_rule
+            if emitted_state != raw_decision:
+                fails.add(f"decision_state {emitted_state!r} != Steps 2-3 recomputed from the raw records ({raw_decision!r}, base {raw_rule})")
+            if di_decision != raw_decision:
+                fails.add(f"Steps 2-3 from DecisionInputs ({di_decision!r}) diverge from the raw-record recomputation ({raw_decision!r})")
+            if "reject_recommended" not in di:
+                fails.add("decision_inputs.reject_recommended must be PRESENT on a non-gated emission (§5.3 presence biconditional)")
+            else:
+                if di["reject_recommended"] != raw_reject:
+                    fails.add(f"reject_recommended {di['reject_recommended']!r} != recomputed {raw_reject!r} (§6)")
+                if di_reject != raw_reject:
+                    fails.add("reject_recommended recomputed from DecisionInputs diverges from the raw-record recomputation")
+
+    return fails, warnings, recomputed
+
+
+# --- CLI -----------------------------------------------------------------------
+
+
+def _load_json(path: Path, reason: str, label: str):
+    try:
+        return json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError, UnicodeDecodeError) as exc:
+        if reason.startswith("manifest"):
+            raise ManifestError(reason, f"{label}: unreadable or invalid JSON ({exc})")
+        raise ArtifactSchemaError(reason, f"{label}: unreadable or invalid JSON ({exc})")
+
+
+def _verify_file_hash(path: Path, expected_sha256: str, label: str) -> bytes:
+    try:
+        raw = path.read_bytes()
+    except OSError as exc:
+        raise ManifestError("manifest_incomplete", f"{label}: unreadable ({exc})")
+    actual = hashlib.sha256(raw).hexdigest()
+    if actual != expected_sha256:
+        raise ManifestError("manifest_hash_mismatch", f"{label}: file bytes do not match the manifest sha256 (§11)")
+    return raw
+
+
+def run(argv=None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+    parser.add_argument("--manifest", type=Path, required=True)
+    parser.add_argument("--precommitment", type=Path, required=True)
+    parser.add_argument("--verdict-record", type=Path, required=True)
+    parser.add_argument("--traceability", type=Path, required=True)
+    parser.add_argument("--roadmap", type=Path, required=True,
+                        help="the file the manifest's revision_roadmap entry hash-binds (Schema 7 machine form)")
+    parser.add_argument("--letter", type=Path, default=None,
+                        help="the file the manifest's editorial_decision_letter entry hash-binds (markdown)")
+    parser.add_argument("--apply-report", type=Path, action="append", default=[], dest="apply_reports",
+                        help="apply-report file(s), in the manifest's apply_reports[] order")
+    args = parser.parse_args(argv)
+
+    try:
+        manifest = _load_json(args.manifest, "manifest_incomplete", "manifest")
+        validate_manifest(manifest)
+        artifacts = manifest["artifacts"]
+
+        roadmap_raw = _verify_file_hash(args.roadmap, artifacts["revision_roadmap"]["sha256"], "revision_roadmap file")
+        roadmap_items = load_roadmap(json.loads(roadmap_raw.decode("utf-8")))
+
+        letter_blocks = None
+        if artifacts["editorial_decision_letter"]["present"]:
+            if args.letter is None:
+                raise ManifestError("manifest_incomplete", "manifest declares editorial_decision_letter present — --letter is required")
+            letter_raw = _verify_file_hash(args.letter, artifacts["editorial_decision_letter"]["sha256"], "editorial_decision_letter file")
+            letter_blocks = parse_letter_blocks(letter_raw.decode("utf-8"))
+        elif args.letter is not None:
+            raise ManifestError("manifest_incomplete", "--letter provided but the manifest declares the letter absent")
+
+        report_entries = artifacts["apply_reports"]["items"] if artifacts["apply_reports"]["present"] else []
+        if len(args.apply_reports) != len(report_entries):
+            raise ManifestError(
+                "manifest_incomplete",
+                f"manifest declares {len(report_entries)} apply report(s); {len(args.apply_reports)} provided via --apply-report",
+            )
+        report_payloads = []
+        for i, (path, entry) in enumerate(zip(args.apply_reports, report_entries)):
+            raw = _verify_file_hash(path, entry["sha256"], f"apply report [{i}]")
+            try:
+                report_payloads.append(json.loads(raw.decode("utf-8")))
+            except (json.JSONDecodeError, UnicodeDecodeError) as exc:
+                raise ManifestError("manifest_incomplete", f"apply report [{i}]: invalid JSON ({exc})")
+        witness, witness_notes = compute_apply_chain_witness(manifest, report_payloads)
+
+        precommitment = _load_json(args.precommitment, "phase1_lint_failed", "precommitment")
+        validate_precommitment(precommitment)
+        verdict_record = _load_json(args.verdict_record, "phase2a_lint_failed", "verdict_record")
+        validate_verdict_record(verdict_record)
+        traceability = _load_json(args.traceability, "phase2b_lint_failed", "traceability")
+        validate_traceability(traceability)
+
+        # Hash chain (§11): the same-inputs proof.
+        if precommitment["input_manifest_hash"] != canonical_hash(manifest):
+            raise ManifestError("manifest_hash_mismatch", "precommitment.input_manifest_hash does not recompute from the manifest (§11 chain)")
+        if verdict_record["precommitment_hash"] != canonical_hash(precommitment):
+            raise ManifestError("manifest_hash_mismatch", "verdict_record.precommitment_hash does not recompute from the precommitment artifact (§11 chain)")
+        if traceability["verdict_record_hash"] != canonical_hash(verdict_record):
+            raise ManifestError("manifest_hash_mismatch", "traceability.verdict_record_hash does not recompute from the verdict_record artifact (§11 chain)")
+    except ManifestError as exc:
+        print(f"[RE-REVIEW-ABORT: {exc.reason}]")
+        print(f"FAIL: {exc}")
+        return EXIT_INVALID
+    except ArtifactSchemaError as exc:
+        print(f"[RE-REVIEW-ABORT: {exc.reason}]")
+        print(f"FAIL: {exc}")
+        return EXIT_INVALID
+
+    fails, warnings, recomputed = check(
+        manifest, precommitment, verdict_record, traceability,
+        roadmap_items, letter_blocks, witness, witness_notes,
+    )
+    for note in warnings:
+        print(note, file=sys.stderr)
+    if fails.messages:
+        print("[RE-REVIEW-ABORT: synthesis_mismatch]")
+        for message in fails.messages:
+            print(f"MISMATCH: {message}")
+        return EXIT_SYNTHESIS
+    expect = recomputed.get("expect")
+    outcome = expect[0] if expect else traceability["decision_state"]
+    print(
+        f"re-review synthesis ok: round {manifest['round_id']!r}, revision {traceability['revision']}, "
+        f"decision_state {outcome!r}, apply_chain_witness {witness!r}"
+    )
+    return EXIT_PASS
+
+
+if __name__ == "__main__":
+    raise SystemExit(run())