AGENT ASSURE: CONTROLLED EVIDENCE-SENSITIVITY WALKTHROUGH

STATUS
Repository-recorded deterministic terminal walkthrough. The expected console
block is mechanically checked against the authoritative packaged renderer. It
is not an external reproduction record, a real-provider experiment, or a
prevalence measurement. Paths may vary by environment; no provider key,
network call, or token spend is required.

PREREQUISITE
Python 3.11 or newer with the current agent-assure package installed.

1. SHOW THE INSTALLED VERSION

$ agent-assure --version

EXPECTED: the command identifies the installed package version.

2. RUN THE COMPLETE OFFLINE DETECTOR CONTRACT TEST

$ agent-assure demo evidence-sensitivity --out .tmp/demo/evidence-sensitivity --clean

EXPECTED CONSOLE FACTS:

agent-assure controlled evidence sensitivity demo

Synthetic declarative fixture harness/oracle only; this result does not show that a model used contextual evidence instead of parametric memory. This detector contract is not a causal guarantee or a real-model failure-prevalence estimate.

Authoritative corpus substitution:
  Policy A expected decision: approve
  Policy B expected decision: deny

Responsive synthetic subject:
  observed relation: decision_flip
  state: responsive
  gate effect: pass

Intentionally evidence-inertial synthetic subject:
  decisions: approve -> approve
  retrieval and citation prerequisites: pass
  ordinary arm evaluations: pass
  decision inertia: true
  state: evidence_insensitive
  gate effect: block

Artifacts:
  demo-summary.json
  responsive/evidence-sensitivity.json
  evidence-inertial/evidence-sensitivity.json
  evidence-inertial/evidence-sensitivity.md
  evidence-inertial/evidence-sensitivity.html

Demo result:
  success: citations pass while controlled sensitivity blocks inertia

The wrapper returns zero only after independently executing both complete
two-arm experiments, validating their artifacts, observing the responsive
control, and observing the expected blocking result for the intentionally
evidence-inertial fixture.

3. INSPECT THE MACHINE SUMMARY

$ python -m json.tool .tmp/demo/evidence-sensitivity/demo-summary.json

EXPECTED FIELDS INCLUDE:

  "status": "success"
  "underlying_exit_code": 1
  "expected_behavior_observed": true
  "notice": "Synthetic declarative fixture harness/oracle only; this result does not show that a model used contextual evidence instead of parametric memory. This detector contract is not a causal guarantee or a real-model failure-prevalence estimate."

The responsive experiment records `approve -> deny`, `decision_flip`, and a
passing gate effect. The evidence-inertial experiment records `approve ->
approve`, passing retrieval/citation prerequisites and ordinary evaluations,
`decision_same`, decision inertia, and a blocking gate effect. The summary also
records the two internal commands and their expected exit codes without
embedding environment-specific absolute paths.

4. REVIEW THE PORTABLE EVIDENCE BUNDLE

$ python -m json.tool .tmp/demo/evidence-sensitivity/evidence-inertial/evidence-sensitivity.json

Open these reviewer views locally as needed:

  .tmp/demo/evidence-sensitivity/evidence-inertial/evidence-sensitivity.md
  .tmp/demo/evidence-sensitivity/evidence-inertial/evidence-sensitivity.html

The JSON report embeds the exact compiled suite, both RunSets, both evaluation
summaries, fixture snapshots, corpus snapshots, authority contract, controlled
difference checks, deterministic retrieval replay, result, assumptions, and
limitations. The Markdown and HTML are derived reviewer views.

5. EXPOSE THE UNDERLYING BLOCKING EXIT

$ agent-assure demo evidence-sensitivity --out .tmp/demo/evidence-sensitivity --clean --strict

EXPECTED: the same evidence is regenerated and the process exits 1 because
the evidence-inertial fixture did not make the declared decision response when
the authority-bound governing evidence changed.

END OF RECORDED WALKTHROUGH
