…16 checks (69 total)

Codex (gpt-5.5) reviewed the initial 53-check set (codex_review.md,
committed alongside). Applied in full:

- Factual must-fixes: deletion check now matches normal-delete
  retention semantics (claims/evidence kept as audit history; only
  hard-forget scrubs); execution-classes check names E2 extraction as
  the deliberate volume-proportional exception (D25/D31); Selection
  ledger check matches schema (plain keeps not ledgered); bitemporal
  check carries the D43 no-cap rule for measurements.
- Judgeability: split the omnibus checks (boundary D60 -> 3, retrieval
  envelope D49 -> 5, lineage/living D55 -> 2); reworded checks to
  demand concrete evidence (shared hydrator boundary, import-linter
  contracts, startup/CI family validator, targeted reuse tests).
- Coverage: +9 checks — D1/D46 source-of-truth split, D16 scope views,
  D22 golden sets, D23 registry scale, D36 sub-worker chain, code
  tooling/migrations, D62 delivery artifacts + enforced architecture,
  D63 embedding port config.
- Mechanics: missing decision tags added, cite paths fixed, README
  re-inventoried (69 checks, scope D1-D63).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…D64 (seed core) in the predicate check + README range

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…coverage, designs for truth)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author review round 2: the precedence rule alone left the checks
decision-first in reading order (64/69 opened with "Read decisions.md
Dxx and <design>"). Flipped all of them: the Read instruction now
leads with the design doc / requirements section as "the binding
source", with decisions.md cited second for rationale and refinement
history. Also repaired ret_envelope_contradiction_completeness, whose
description had been corrupted by the earlier cite-insertion (its
description contained ", then", which the insertion script matched
instead of the Read sentence). Full sweep: 69 checks valid, none
decision-first, no polluted descriptions, every check cites a
design/requirements source and carries the precedence rule.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…hecks (76 total)

The round-2 review (codex_review_2.md, committed alongside) judged the
suite against the design docs as ground truth. All load-bearing quotes
were verified against the designs before applying. Applied in full:

- Surgery repairs: six mis-pathed requirements citations fixed
  (plan/requirements/ prefix) + binding-source phrasing added to the
  unflipped Read sentences; four non-design "binding source"
  designations rebound (CLAUDE.md, concepts.md, workers.md, the P3
  analysis are now context, not binding).
- Over-binding corrections (checks could fail conforming code):
  provider calls through D61 ports are conformant; direct-env access
  allowed via the requirement's reasoned per-line exception; ingest
  bumps legitimately cascade downstream; effective_period added to the
  claim_valid_kind enum; borderline per-claim grounding escalation is
  designed; observations use the shared adjudication_outcome enum;
  re-occurring relations in non-overlapping windows get new rows;
  reviewer may be a designated agent; writers run parallel across
  disjoint pages ("editing shared files" restored); model page binds
  writer invocations, not every page; version-vs-lineage tombstone
  branches spelled out; execution classes need an inspectable mapping
  (no invented enum); retry bound is a tunable per-stage default;
  producer/checker never-share per orchestration §8 (D53's
  recorded-exception prose loses under precedence); projection purity
  judged as outcome, not an invented import graph.
- D64 made scoreable (16-predicate seed incl. uses/reports_to
  signatures); D23 expanded to the full partitioned estate + the
  ungated load-test hook (a round-1 application miss).
- +7 coverage checks: K trigger surface/acyclicity, K eval canaries +
  citation faithfulness, K writer runtime two-band, K hard-forget git
  history, orchestration topology, S58 skill acceptance, continuous ER
  health.
- Tags reconciled with scored conditions; README recounted (76),
  column relabeled "primary binding sources", both reviews recorded.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>