00 — Overview: MARK → REVIEW → CUT → VERIFY operational package

Macro: IU_CORE_80000X_MARK_REVIEW_CUT_OPERATIONAL_GUIDE_AND_AGENT_MARK_PACKAGE Date: 2026-05-25 Effort: high · Target runtime: 45–60 min Scope: documentation only. No code authored. No migration applied. No production data mutated. No CUT performed in this macro.

1. Why this package exists

The team has historically conflated two operations:

• The user gives a link/source for an article;
• The system "cuts" the article into pieces immediately.

That single-step model is wrong for governed data. The correct operating model is four discrete stages, each with its own gate:

MARK  →  REVIEW / APPROVE  →  CUT  →  VERIFY
 │              │                │         │
 │              │                │         └── reconstruct + axis checks + rollback report
 │              │                └────────── durable IU rows, manifest-driven, atomic
 │              └─────────────────────────── human / DOT approval of a manifest hash
 └────────────────────────────────────────── proposal only; no DB write

Substrate work over 12000x → 70000x has made the technical pieces (auto-cut, three-axis filter, fn_iu_compose, fn_iu_piece_emit_event, fn_iu_reconstruct_source, iu_sql_link, lifecycle DOT, product factory, event runtime) LIVE_READY. What was missing was a clear written procedure that lets a fresh Agent perform MARK correctly without inventing source, fabricating IDs, or claiming a CUT was done when it was not.

This package supplies that procedure.

2. Critical distinction — MARK is not CUT

Stage	Output	Mutates production?	Authority needed
MARK	proposed cut_manifest	no	Agent (LLM) may produce
REVIEW	annotated/approved manifest	no	Operator / DOT signs
CUT	durable IU rows + memberships	yes (one atomic TX)	Approved manifest hash
VERIFY	reconstruction + axis report	no	Auto, on CUT completion

A fresh Agent that has read this package must never say "cut completed", "IU created", "production updated" after running MARK. The Agent's output is a proposal. Until the CUT stage actually runs against an approved manifest hash, nothing is durable.

3. Package contents

File	Purpose
00-overview.md	This file.
01-operational-flow-mark-review-cut-verify.md	The four stages in detail with gates and IO contracts.
02-agent-mark-instructions.md	Step-by-step procedure a fresh Agent follows to perform MARK.
03-cut-manifest-schema.md	YAML/JSON schema for cut_manifest (required & optional fields, validation rules).
04-review-approval-checklist.md	What the operator/DOT checks before approving a manifest.
05-dot-cut-from-approved-manifest-contract.md	The DOT/CUT command contract that consumes an approved manifest.
06-three-axis-requirements.md	Axis A (source order) / B (professional) / C (parent-child) field & invariant rules.
07-examples-dieu-37-38-39-mark-request.md	Example user request + example MARK proposal for Điều 37/38/39 (placeholder source).
08-agent-mark-output-template.md	Copy-pasteable template the Agent fills in for its MARK output.
09-test-plan-codex-mark-then-claude-verify.md	Validation: Codex/new-Agent reads only this package, produces MARK; Claude Code reviews.
10-carry-forward-to-implementation.md	What still needs to be built (DOT dot_iu_cut_from_manifest, MARK harness, etc.) — packaged for the next macro.

4. System rules this package preserves

• No hardcode. All identifiers (article labels, doc codes, IDs) flow through the manifest.
• DOT 100%. Every irreversible step has a DOT command contract with explicit guards.
• MARK before CUT. MARK without an approved manifest never advances to CUT.
• Review/approval gate. A manifest's approval.status must be approved (with approved_by + approved_at + KB doc id) before the CUT command will accept it.
• No fake PASS. The Agent reports uncertainty_flags honestly; an empty manifest is preferred over a fabricated one.
• No source invention. If the source link does not contain the requested article, the Agent reports the gap and stops.
• No production mutation during MARK. MARK writes only to scratch directories and KB report paths.
• No cross-IU vector. Vector indexing is gated separately and never bundled into CUT.
• PG → Directus → Nuxt is the data flow direction. CUT writes to Postgres; downstream layers reflect it.
• production_documents UNTOUCHED. No CUT writes to production_documents (the historic raw store).
• Rollback / compensation required for CUT. Every CUT command produces a rollback package keyed on manifest_id.

5. Substrate this package builds on (from 70000x final readiness matrix)

Capability	Code path	State
Auto-cut documents	upstream piecing complete; manifest consumes existing pieces	LIVE_READY
Reconstruct by original source order	fn_iu_reconstruct_source (Axis A, dense monotonic source_position)	LIVE_READY
Axis B (professional) filter	iu_metadata_tag × iu_metadata_tag_registry (kinds: legal_document, section_type, unit_kind)	LIVE_READY
Parent / child / grandchild filter	fn_iu_subtree (relative_depth ∈ {0,1,2}), iu_tree_path	LIVE_READY
Create new document from pieces	fn_iu_compose (with new_piece minting)	LIVE_READY
Add / remove / merge / split piece	fn_iu_collection_add_piece / _remove_piece / fn_iu_piece_merge / fn_iu_piece_split	LIVE_READY
Piece event in/out	fn_iu_piece_emit_event + lifecycle trigger	LIVE_READY (dry_run); live flag exists
Bridge piece ↔ DB SQL	iu_sql_link (link_role='represents', 11-role vocab) + validator	LIVE_READY (outbound)

All six user flows PASS at the test runner level: ops/iu-core-six-flow-test-readiness/run_six_flow_tests.sh (6/6 PASS, run 20260525T081115Z).

6. What the next macro will do (carry-forward sketch)

After this documentation macro:

1. Codex / a fresh Agent reads only this package.
2. User sends a real MARK request for Điều 37 (with exact URL).
3. Codex emits a cut_manifest.json with proposed pieces, axes A/B/C drafts, reconstruction preview.
4. Claude Code / operator reviews the manifest against 04-review-approval-checklist.md.
5. If approved: the next macro runs dot_iu_cut_from_manifest against the approved hash, durably creates IUs, runs VERIFY.
6. If not approved: gaps are documented in the manifest's uncertainty_flags, the Agent re-marks; loop until approval.

That sequence is the validation of this package.

7. Tags & search anchors

The KB upload of this package will be indexed; the following phrases are placed verbatim so semantic search returns this package:

• "MARK is not CUT"
• "MARK → REVIEW → CUT → VERIFY"
• "cut_manifest_schema"
• "Agent MARK Instructions"
• "Axis A original source reconstruction"
• "Axis B professional content axis"
• "Axis C parent child grandchild"
• "approved manifest"
• "dot_iu_cut_from_manifest"