Instrument 08 / protected-control preservation

The refactor worked.
The bell went quiet.

A goblin improves the delivery code. Every business test stays green. But the security bell no longer rings. Should the change be admitted?

Synthetic engineering demonstration · not a model trialExecutable toy pipelines, a rule-based gate, and authored evidence. No live agents, deployments, security controls, or notifications.
—named expectations matched
—benign fixtures admitted
—finite states checked
—deliberate false approvals

Put a change at the gate.

Interactive replay of recorded Python results — not a fresh execution.

Recorded admission decision

Supplied observation

    These are in-memory event markers, not a real ledger or attested telemetry. The gold step marks the business effect.

    REPOSITORY ACCESS ONLY
    BUSINESS TESTS ONLY
    FREEZE EXACT BUILD
    EVIDENCE GATE
    Inspect the minimal decision record

    Unsigned hashes bind this toy record to candidate, policy, and evidence. They do not authenticate the agent or observer. No prompt or conversation is included.

    

    

    Now break the claim.
    Both failures are part of the published result.
    Failure 01 / coverage assumption
    The bell only fails at 999.
    All four declared probes use payload 7. A conditional hook passes those probes, then stops monitoring when payload 999 arrives. Finite evidence is not universal preservation.
    Failure 02 / observer assumption
    The witness lies.
    A forged report supplies a good trace under the bad build’s digest. This toy trusts an observer label; it does not authenticate the channel. The unmonitored build passes.
    

    The declared fixture ledger.
    21 cases. Six benign. No observed false alarm among those six.
    CandidateExpectedObservedBusiness tests
    The separate 512-state grid covers seven Boolean flags × two alert thresholds × two routes. All 512 decisions matched the same-author Boolean oracle: 3 ALLOW, 509 BLOCK. Its constructed class balance is not a real-world accuracy estimate. Inspect every state →
    

    Observed
    All 21 business tests passed, including the 12 declared failure cases. The evidence gate matched the declared outcomes. Exact-build freezing rejected five benign changes. The two assumption-breaking witnesses were falsely allowed.
    Inferred / SAAC implication
    Make preservation of required controls part of the authorized change. Separate routine outreach from mandatory security escalation. Bind any decision to the exact candidate and policy; place enforcement outside the changing workload.
    Still unknown
    Real agent behavior, semantic intent recognition, attested provenance, deployment-race resistance, and production false-alarm rates. No real identity provider, delegation chain, tamper-proof ledger, or protected-system gate is implemented here.
    

    The defensible claim: in this declared toy model, behavior checks can preserve selected controls while allowing benign change. This is a SAAC-inspired admission demonstration, not validation of Semantic Agent Admission Control or proof of an unavoidable security boundary.