Worked example · reviewed 15 September 2026

Mnemiq: what the open engine proves—and what deployment must still prove

Mnemiq is an Apache-2.0 enterprise data-agent and text-to-SQL engine developed and maintained by Agentic Fabriq. This assessment is pinned to upstream commit a08fe859.

Scope: This is a code-and-documentation mapping, not a security audit, product certification, comparative ranking or prediction of performance on a customer database.

Evidence map

Control	Observed support	Deployment boundary
AU-01 · Pre-retrieval authorization	The project describes and implements policy-scoped schema retrieval.	Identity and policy configuration remain operator responsibilities.
AU-02 · Pre-execution authorization	The SQL decision path re-checks referenced objects against access policy.	Source-native permissions are still required as defense in depth.
QS-01 · Statement allowlist	Read and write paths parse and constrain statement classes separately.	Views, functions and source behavior remain trust boundaries.
SE-01 · Certified definitions	The semantic and enrichment code can represent operator-provided definitions and provenance.	The organization must supply, approve and maintain those definitions.
VE-02 · Verifier failure	Verification state is represented explicitly and tests cover unavailable verification.	The organization chooses modes and risk-tier rules.
RF-01 · Unsupported-data refusal	The product path supports deferral and the evaluation harness separates right, refused and wrong outcomes.	Public results do not predict another schema’s results.
LN-01 · Reconstructable answer	Answer traces include SQL, tables, lineage completeness, semantic version and timing.	Retention, tamper evidence and enterprise audit integration sit outside the engine.
QS-02 · Source-native least privilege	Documentation addresses source-specific deployment, including Oracle constraints.	The configured database principal and infrastructure must be tested independently.

Why it is useful as an example

Inspectability

The major retrieval, authorization, SQL decision, verification and evaluation paths are public.

Outcome separation

The project treats a correct answer, refusal and wrong answer as different outcomes rather than hiding them in one number.

Deployment honesty

The upstream documentation identifies controls that depend on the database and operator, not only the application.

Reproducibility

The evidence record names the reviewed commit and the public paths supporting each mapping.

Independent tests an operator should still run

Attempt access to denied tables, columns, rows, views and functions using the deployed identity.
Verify that restricted schema names do not appear in retrieval, refusal messages or traces.
Test source-native denial using the same database principal outside the application.
Evaluate real questions across answer, definition and refusal bands.
Interrupt the verifier, semantic store and source to confirm fail-closed behavior.
Reconstruct sampled answers from retained evidence.

Project relationship: Mnemiq is the upstream technical foundation. Enterprise Data Agent Governance is independently maintained and is not official Mnemiq documentation. Product claims should be verified against the canonical repository.