001Governance

Canon without governance
is suggestion, not authority.

Grounding answers "what may the AI say?" Governance answers everything else: who owns each claim, whether it's still true, how strictly it must be reproduced, and who queried it. This layer is why compliance and brand owners open MsgStack every week.

002Lifecycle × tier — two orthogonal contracts

Status: is it ready?

Every entry has a lifecycle state that gates grounding. Draft, In-Review, and Outdated entries are invisible to agents by default — generation grounds only on approved truth.

DraftHIDDENBeing authored — not ready for consumption
In ReviewHIDDENAwaiting SME approval
ApprovedGROUNDSActive canon — included in all grounding
LockedGROUNDSImmutable anchor — visible, uneditable
OutdatedHIDDENSuperseded — hidden until replaced

Tier: how may an LLM use it?

Orthogonal to status. Tier is the per-entry generation contract — and entries can't be approved without one. A paraphrased Tier 1 entry is a hard conflict, by definition.

Tier 1 — LockedGraph traversal only

Used exactly as written. No paraphrasing, no approximation, never nearest-neighbor.

Tier 2 — StructuredHybrid retrieval

Substance and positioning preserved; phrasing adapts to context.

Tier 3 — GroundedVector retrieval

Direction and tone consistent with canon; full phrasing latitude.

003Alignment scoring — continuous governance

Score anything against the canon.

Paste a draft, or point an agent at one, and get a per-section alignment report: 0–100 score, missing proof points, and contradictions — split into hard conflicts (factual contradictions) and soft conflicts (voice and style drift).

score_alignment API + MCP toolSHIPPED ✦
Color-coded report UI with suggestionsSHIPPED ✦
Hard vs. soft conflict classificationSHIPPED ✦
Markdown report exportSHIPPED ✦
Batch scoring across CRM / Drive foldersPLANNED · v0.9
Weekly drift reportPLANNED · v0.9
ALIGNMENT REPORT — sdr_email_draft.md71 / 100
HARD

"stops threats instantly" contradicts Entry 228-C ("sub-millisecond") — paraphrased Tier 1 claim.

SOFT

Tone exceeds brand register bounds in ¶2 — warmth slider at 0.8 vs. approved 0.5.

GAP

Missing: efficiency proof point (Entry 441-B is approved and unused).

PLATE IIThe report a reviewer actually reads.
004Accountability — DRI, SLA, audit

Every claim has a DRI

SHIPPED ✦

A named, accountable individual on every entry and domain — distinct from the approver, transferable when teams change, with unowned items surfaced first in the accountability view.

Freshness is a contract

PLANNED · v0.9

Content SLAs replace static staleness flags: per-domain review cadences plus trigger events — a release, a competitive move — that open review windows. Miss the window, and the DRI is notified.

Every query is on the record

SHIPPED ✦

The query audit log captures caller, query, entries returned, confidence, and latency for every grounding search — filterable, CSV-exportable, 90-day retention by default.

RECORD — query_audit_logCSV ↓
CallerQueryEntriesConf.
claude-mcp"zero-day detection rate"441-B0.97
cursor-mcp"API rate limit v2 paths"302-A, 302-C0.91
portal:field"SOC2 EU data sectors"118-F0.95
navigator"what changed in pricing canon"77-A, 77-B0.88
PLATE IIIWho asked what, and what the canon answered.
005Where this goes — content CI/CD

Promotion becomes a gated pipeline.

v1.0 treats canon promotion like a code merge — five stages, and a failed validation blocks promotion. PLANNED · v1.0

I

Validate

Schema, tier tag, DRI, metadata completeness

II

Test

Golden query dataset — precision/recall vs. baseline

III

Merge

Status transitions; vector + graph reindex

IV

Propagate

Signal to bindings, artifacts, dependent domains

V

Audit

Promotion event logged with alignment delta