dot-iu-cutter v0.5 — Constitution Nuxt Parser Reference Implementation: Operations-First Framing

Phase: v0_5_constitution_nuxt_parser_reference_implementation_authoring · Nature: code/script authoring + test design — no capture, no seed, no DML, no dry-run · Date: 2026-05-18 · doc 1 of 5

snapshot_written: false ; source_document_insert: false ; source_document_version_insert: false
dml: none ; production_db_mutation: none ; checksum_persisted_into_registry: none
dry_run: none ; cut: none ; verify: none ; schema_change: none ; grant_revoke: none
directus_mutation: none ; deploy_restart: none ; git_commit: none
mutation: none (KB read-only grounding + 5 authoring uploads + local /tmp scratch deleted)
decision_authority: GPT / User ONLY ; self_advance: PROHIBITED

Opened per GPT ruling E1_BLOCKED_BY_PARSER_IMPLEMENTATION_DIVERGENCE__REFERENCE_IMPLEMENTATION_NEXT, selected path R2_RATIFY_REFERENCE_IMPLEMENTATION_FIRST (reviews/dot-iu-cutter-v0.5-constitution-source-snapshot-capture-e1-blocked-gpt-ruling-2026-05-18.md).

1. The operational goal this serves

north_star: a user issues one small command — "Cắt Hiến pháp" — and the system
  resolves the source, pins a deterministic version identity, cuts enacted_only,
  verifies, and returns a concise PASS/BLOCKED report; humans handle exceptions only.
  (handoffs/dot-iu-cutter-current-operating-objectives-and-principles-2026-05-18)

The Constitution source is a living Nuxt/Directus-rendered HTML page. "Cắt Hiến pháp" must bind to a reproducible source_document_version identity (content_checksum). The whole snapshot/version-policy chain (Option B captured-snapshot, write-once checksum-addressed artifact, rehash gates) exists to make that identity stable and race-free.

2. Why prose-only parser rules are insufficient (root operational defect)

defect_exposed_by_E1:
  - parser_profile nuxt-incomex-portal-constitution-v1 existed ONLY as a prose
    spec (N1..N9 + candidate_B span).
  - Two faithful readers of that prose produced DIFFERENT byte-level normalized
    content for the SAME source content:
      Codex canonical executor : 17660443…/len 17522   (ratified canonical)
      Claude prose reconstruction: 072983ac…/len 17657 (Δ +135, deterministic 3/3)
  - markers identical 19/1/1/1 -> NOT a normative document change; purely an
    executor-interpretation gap in one normalization step.
consequence: a one-command "Cắt Hiến pháp" cannot depend on WHICH agent/executor
  runs it. A version identity that changes with the executor is not an identity.
  Prose is a specification; only a byte-exact, versioned implementation (or a
  formally pinned canonical executor) can MINT a production source-version hash.

A prose rule such as "collapse blank-line runs to a single \n" is read by one executor as "one newline, zero blank lines between blocks" and by another as "one blank line between blocks". Both are defensible English; they differ by ~135 characters across the document. Identity must not hinge on that ambiguity.

3. What this phase delivers (operations-first, not mechanism-first)

deliverable: a deterministic, portable REFERENCE IMPLEMENTATION of
  parser_profile nuxt-incomex-portal-constitution-v1 such that ANY approved
  executor, given the same source bytes, yields identical:
    normalized_content / normalized_content_checksum / normalized_content_length
    / marker_counts / extraction_span_diagnostics / parser_version
purpose_in_workflow:
  - removes the executor-divergence failure class permanently
  - makes E1 snapshot capture reproducible (unblocks B5 path, still gated)
  - lets "Cắt Hiến pháp" rehash the pinned artifact and compare deterministically
scope_exclusions (this phase): no snapshot write, no seed, no DML, no dry-run,
  no CUT, no VERIFY, no schema/Directus/deploy change, no git commit, no
  self-advance to E1 capture or E2 seed.

4. PASS / BLOCKED definition for this authoring phase

PASS (reference implementation authoring):
  P1 a single deterministic implementation is authored with EVERY ambiguous
     step pinned to one explicit decision (fetch/decode, span, detag, entity
     decode, NFC, line endings, h-whitespace, v-whitespace, CHANGELOG boundary,
     marker codepoint preservation, BEGIN/END sentinel region semantics)
  P2 run read-only against the live source is deterministic across repeated
     fetches (same normalized_content_checksum)
  P3 EITHER it reproduces the ratified canonical checksum
       17660443…/17522 (then -> recommend E1 re-run with this impl),
     OR the exact divergence cause is identified and the precise fix stated,
     OR live-source drift is proven and a new-fixture-ratification path is
     recommended (never a silent canonical-checksum change)
  P4 reference implementation source is captured in KB (SSOT) so it is portable
     and ratifiable independent of any one environment

BLOCKED:
  B1 cannot obtain source read-only, OR
  B2 implementation is non-deterministic across fetches under one pinned ruleset
     (FAIL_NONDETERMINISM), OR
  B3 evidence is insufficient to either reproduce canonical or localize the
     divergence (state "insufficient evidence", do NOT guess a checksum)
forbidden_outcome (any case): inventing/altering a canonical checksum,
  freezing a non-canonical artifact, or seeding from unpinned content.

5. This phase's result (summary; detail in docs 2–5)

result: PASS (P1–P4 met). The reference implementation, run read-only against
  the live source 3/3 deterministically, REPRODUCES THE RATIFIED CANONICAL
  identity EXACTLY: 17660443e0f23e994e1807cf8e22920951a9e70c598956dbd0e752f4f5cae80c
  / length 17522 / markers ✅19 📋1 📝1 ⛔1.
divergence_cause: localized precisely to ONE step — N8 vertical-whitespace
  ("collapse blank-line runs to a single \n"). Canonical = drop ALL empty lines
  (single \n between content). Claude E1's prose reconstruction kept one blank
  line between blocks -> +135 chars / 072983ac…. Proven by controlled variant
  sweep (doc 2 / doc 4): V3 variant == canonical, V4 variant == Claude E1 output.
recommendation: GPT ratify this reference implementation as the canonical
  executor for nuxt-incomex-portal-constitution-v1, then re-run E1 capture
  using it (decision is GPT/User; not taken here). See doc 5.

6. Statement

• Operations-first framing: byte-exact implementation is required because a prose-ambiguous normalization step (N8) makes version identity executor-dependent; PASS/BLOCKED defined (QG6 framing).
• No snapshot/seed/DML/dry-run/CUT/VERIFY/schema/Directus/deploy/commit; KB read-only + 5 uploads + deleted /tmp scratch only (QG1/QG5/QG7).
• doc 1 of 5; STOP after 5 files → route GPT/User. Self-advance PROHIBITED.

Companions: algorithm-analysis, implementation-draft, test-result, authoring-report.