5. Jev × Petri: classify evidence, never grant effects

This experiment connects two courses without putting a model on the production control path. Jev supplies a synthetic advisory classification in our confidence-gate lab; the Petri-net engine enumerates what a proposed control flow would permit. Neither alone verifies a real Gaia or IX transition.

Question and baseline

Can a confidence threshold authorize implementation? The pinned gaia_design_authority case expects contradicted: an approved design receipt exists, but no implementation grant does. The synthetic stress harness deliberately returns supported at 0.98. This is not a Jev API response or a measured model error.

The unsafe net consumes that advisory token to produce effect. Its shortest witness is classify → authorize_from_advisory. The guarded net instead routes advisory output to review, and authorize requires both independently verified evidence and implementation authority. Its tokens are single-use in this deliberately finite model; production policy may differ.

synthetic Jev choice ──> advisory ──> review ──> confirmed ──┐
                                 verified evidence ──────────┤ authorize ──> effect
                                 implementation grant ───────┘

The two right-hand inputs must come from independently checked receipts, not from Jev’s confidence or the model itself. A Petri marking describes assumptions; it does not create real receipts.

Run the bounded tracer

From the repository root:

python -m unittest discover -s code/typesafe-ai-system-one -p 'test_jev_gate_audit.py' -v
python -m unittest discover -s code/repository-dogfooding-lab -p 'test_jev_petri_fixture.py' -v
dotnet test code/petri-nets/Tests -c Release --filter JevEvidenceGateTests

The fixture is checked against the Python synthetic response and read by the C# tests. The net definition uses the existing course engine. Tests require complete reachability; they show the unsafe effect is reachable, no guarded effect is reachable with either prerequisite absent, and a separately authorized path remains possible when both are present.

What this does not establish

• No TypeSafe call, billed token measurement, or Jev quality estimate occurred.
• No production Gaia or IX state was read or changed. The case text is a pinned teaching fixture, not a current authority receipt.
• The guarded net checks a finite abstraction. It cannot establish that real code enforces the same guards, that receipts are authentic, or that concurrency and retries preserve them.

Next dogfooding gate

Pin one actual Gaia or IX transition and its revision. Map its source-of-truth receipt fields to the net’s places; seed missing and forged evidence; replay the unsafe witness through a public test seam. Compare against a simple deterministic guard test. Incubate only if the model catches an otherwise missed defect and an independent reviewer accepts the mapping. Otherwise retain the simpler test and reject this extra model.

Exercise

In a disposable copy of the guarded net, remove the implementation_authority → authorize arc. Predict which missing-token test will fail, then run the focused C# tests. Restore the arc afterwards.

Solution

With verified evidence present but no implementation grant, authorize becomes enabled after confirm. The test that explores (verified=true, authority=false) finds a reachable effect and fails. This is a model counterexample, not a real Gaia defect.