Governance model

Two axes that are never conflated. Risk class describes the consequence of an action. Trust state describes the maturity of data. The authority matrix binds them, and every control in this sandbox routes through it.

SRS section 3, governance taxonomy

Trust states, L0 to L5

A property of data. It is never raised as a side effect of some other action.

Unverified

Raw ingested content. No extraction has been attempted and nothing may cite it as evidence.

AI Extracted

Produced by the extraction model with a source anchor. Never gate-effective on its own.

Human Reviewed

An engineer has read the claim against its source anchor but has not accepted it.

Engineer Accepted

Accepted by a named engineer. Usable for downstream engineering work, not yet canonical.

Canonical

Single authoritative value selected by an Engineer Lead where claims conflicted.

Baselined

Frozen inside an approved baseline. Changing it requires a controlled change with impact analysis.

Risk classes, C0 to C5

A property of an action. It fixes who may perform it.

Read

Reads records. No state change, no audit consequence beyond access logging.

Draft

Creates or edits draft data and AI claims. The only class an AI agent may ever reach.

Propose

Submits work for review, raises open items, requests a check run.

Promotes data trust: accepting a claim, resolving a conflict, selecting a canonical fact.

Requires data at L1 or above.

Approve

Passes a workflow gate or approves a document revision for issue.

Requires data at L3 or above.

Baseline

Freezes a contractual baseline or releases to production. Highest consequence action.

Requires data at L4 or above.

Authority matrix

Section 3.4. A filled cell means the role may act at that risk class; the ceiling is the whole point.

Role	C0Read	C1Draft	C2Propose	C3Accept	C4Approve	C5Baseline
ViewerRead-only access to project records and audit history.
EngineerDrafts engineering data, reviews and accepts extraction claims, raises open items.
Engineer LeadEverything an Engineer may do, plus canonical selection on conflict and gate approval.
ManagerBaseline freeze and production release authority.
AI AgentMay only produce draft claims. Structurally incapable of accepting, approving or baselining.
Deterministic CheckerThe only actor whose verdicts are gate-effective. Proposes findings, never approves.

Trust promotions

Each promotion is an explicit action with its own risk class and role floor. There is no path from L1 to L4 that skips a person.

From	To	Action	Risk	Minimum role	Note
L0	L1	Run extraction	C1	AI Agent	Extraction produces claims with source anchors. It cannot promote further than L1.
L1	L2	Mark reviewed	C1	Engineer	Records that a human compared the claim against its anchor.
L2	L3	Accept claim	C3	Engineer	Attributes the value to a named engineer. Reversible only by a new accepted claim.
L1	L3	Accept claim directly	C3	Engineer	Permitted when the engineer reviews and accepts in one step from the workbench.
L3	L4	Select canonical fact	C3	Engineer Lead	Resolves competing accepted claims into one authoritative value.
L4	L5	Freeze in baseline	C5	Manager	Baseline membership is immutable. Later change requires impact analysis.

Try the check

This is the same function every button in the sandbox calls before it does anything.

Role

Action risk

Data trust

Permitted

Engineer is authorised for C3 Accept.

AICEED-GOV-004 (section 3.4 authority matrix)

The AI agent's ceiling is C1 and the deterministic checker's is C2. Neither can accept, approve or baseline anything. That is a structural property of the matrix rather than a rule the model is asked to follow, which is why it holds even when the model is wrong.