Controlled Evidence Sensitivity
Synthetic detector contract test. This report measures a declared controlled evidence-response relation. It is not a causal guarantee and does not estimate real-model failure prevalence.
Scope limit: Synthetic declarative fixture harness/oracle only; this result does not show that a model used contextual evidence instead of parametric memory.
wrong_direction_flip: Expected decisions (baseline -> counterfactual): approve -> deny; observed decisions: deny -> approve. The subject changed decision fields, but in the wrong direction relative to the authority contract.
Relation category: expected decision_flip; observed decision_flip.
Baseline
Observed: deny/denied
Expected: approve/approved
Expected match: false
Retrieved: true
Governing support: true
Evidence link: true
f1d680df273c0189a9c78a1c6246d02e07ab7b52f14285b788ed79f5f601d551
Counterfactual
Observed: approve/approved
Expected: deny/denied
Expected match: false
Retrieved: true
Governing support: true
Evidence link: true
d8e26f1502ae20fc4de4bee2d799bfb1f763d8fa7eaea2852cbd193f7350137e
Detector contract
- Endpoint:
expected_decision_response - Detector status:
synthetic_detector_contract_test - Subject execution scope:
declarative_fixture_harness_not_model_evidence_use - Provenance binding:
fixture_declared_unverified_digests - Synthetic data provenance:
bundled_digest_verified - Synthetic data attestation digest:
not_applicable - Raw content persistence:
exact_corpus_and_fixture_utf8_embedded - Deterministic:
true - Decision inertia detected:
false - Reason codes:
EVIDENCE_SENSITIVITY_EXPECTED_RESPONSE_MISSING - Protocol digest:
d37947d8d978654f91cd6946bd98656097b5b1e8afa3a81cf9a5f1e6052ff983 - Authority contract:
43c6f8258181292c8cb32dd5bcad76f6c7bfd97b95fceb339fbc2ad9bb9d0e57
Controlled-difference manifest
protocol_fixed rows are structural protocol controls.
arm_observed rows are independently derived from each corpus arm and carry
the discriminating evidence.
| Dimension | Basis | Baseline | Counterfactual | Contract | State |
|---|---|---|---|---|---|
agent_implementation_digest | protocol_fixed | c868544c210bdfa1e9c0ea5654f273eee04c8016e703b8330d3d246d92d9cfc7 | c868544c210bdfa1e9c0ea5654f273eee04c8016e703b8330d3d246d92d9cfc7 | equal | controlled |
corpus_digest | arm_observed | f1d680df273c0189a9c78a1c6246d02e07ab7b52f14285b788ed79f5f601d551 | d8e26f1502ae20fc4de4bee2d799bfb1f763d8fa7eaea2852cbd193f7350137e | different | expected_difference |
corpus_document_catalog_digest | arm_observed | df06ee8bf96869705e857b27659df0029f6e5d6d9984960ff3ca93f134854bcb | df06ee8bf96869705e857b27659df0029f6e5d6d9984960ff3ca93f134854bcb | equal | controlled |
deterministic_subject | protocol_fixed | true | true | equal | controlled |
fixture_manifest_digest | protocol_fixed | edd4e04d290fec432d8d4f896895e8ce8410f92843c8431040baf0934b85f73f | edd4e04d290fec432d8d4f896895e8ce8410f92843c8431040baf0934b85f73f | equal | controlled |
governing_evidence_digest | arm_observed | 39e0ed91e07ef56563b5f2f74422735196e6b7889081c603a3d2f443b895dc13 | 0a11b36c7bc440e7227ae9bc91d3e071fa0c4c06240c1c1c35635f3fd6a7e9f6 | different | expected_difference |
governing_retrieval_identity_digest | arm_observed | dafa54ea76f884dabf884e1ac39f76763168404201103a0f7b050d3d7d3c22e3 | dafa54ea76f884dabf884e1ac39f76763168404201103a0f7b050d3d7d3c22e3 | equal | controlled |
knowledge_contract_digest | protocol_fixed | 43c6f8258181292c8cb32dd5bcad76f6c7bfd97b95fceb339fbc2ad9bb9d0e57 | 43c6f8258181292c8cb32dd5bcad76f6c7bfd97b95fceb339fbc2ad9bb9d0e57 | equal | controlled |
model_digest | protocol_fixed | bc987bc3e4bd9475c7a04887537029f65e0cfafc29511974911eb99582b9be7c | bc987bc3e4bd9475c7a04887537029f65e0cfafc29511974911eb99582b9be7c | equal | controlled |
non_governing_evidence_digest | arm_observed | 4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 | 4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945 | equal | controlled |
prompt_template_digest | protocol_fixed | fdcd7641403825e9cfec862b3a47d203c335de53f333ddecc0d7f7bd2ea64ed9 | fdcd7641403825e9cfec862b3a47d203c335de53f333ddecc0d7f7bd2ea64ed9 | equal | controlled |
producer_version | protocol_fixed | 0.6.4 | 0.6.4 | equal | controlled |
query_digest | protocol_fixed | 81c610fe9830b111fe218d68247f5ed1af33ceb2190c0fca2f3719be415bbc7d | 81c610fe9830b111fe218d68247f5ed1af33ceb2190c0fca2f3719be415bbc7d | equal | controlled |
query_family_id | arm_observed | synthetic-benefit-eligibility | synthetic-benefit-eligibility | equal | controlled |
request_digest | protocol_fixed | 8da4857cfd203a6a91c33c8a4e5d8745534500ebc388669ef4701f5a5c8f550e | 8da4857cfd203a6a91c33c8a4e5d8745534500ebc388669ef4701f5a5c8f550e | equal | controlled |
retrieval_algorithm | arm_observed | deterministic-term-overlap@1.0.0 | deterministic-term-overlap@1.0.0 | equal | controlled |
retrieval_top_k | arm_observed | 1 | 1 | equal | controlled |
subject_configuration_digest | protocol_fixed | 2aa320565984b6f436a51578990e6607f249b3e4be27f63f5d8c142c2e50d23b | 2aa320565984b6f436a51578990e6607f249b3e4be27f63f5d8c142c2e50d23b | equal | controlled |
suite_digest | protocol_fixed | f78b7a363b77e227170ac704c5b691a30661828842f7fddfb3cd3f7544f1ee7a | f78b7a363b77e227170ac704c5b691a30661828842f7fddfb3cd3f7544f1ee7a | equal | controlled |
tool_configuration_digest | protocol_fixed | e58a15c3e6d3b530e0149b393001ba9536a068d70aeade5dacdb868ca5c38d07 | e58a15c3e6d3b530e0149b393001ba9536a068d70aeade5dacdb868ca5c38d07 | equal | controlled |
tool_schema_digest | protocol_fixed | ab72f2449da6de27489c10fe35c6bc5e6431d3cede8f13eff0e769dfe52dc498 | ab72f2449da6de27489c10fe35c6bc5e6431d3cede8f13eff0e769dfe52dc498 | equal | controlled |
Prerequisites
| Check | State | Reason |
|---|---|---|
arm-evaluations-pass | satisfied | none |
authority-contract-valid | satisfied | none |
comparable-decision-output | satisfied | none |
controlled-differences-only | satisfied | none |
deterministic-subject | satisfied | none |
evidence-links-present-both-arms | satisfied | none |
fixture-identity-bound | satisfied | none |
governing-evidence-supported-both-arms | satisfied | none |
retrieval-succeeded-both-arms | satisfied | none |
Assumptions
- The knowledge-authority contract correctly declares which contextual evidence governs the synthetic task.
- The committed fixture subject and lexical retriever are deterministic contract-test components, not measurements of a hosted model.
Limitations
- This deterministic synthetic detector contract test does not estimate real-model prevalence.
- Controlled evidence sensitivity is an observed response relation under declared conditions, not a causal guarantee.
- Synthetic declarative fixture harness/oracle only; this result does not show that a model used contextual evidence instead of parametric memory.
- Agent implementation, prompt, model, and tool-schema digests are fixture-declared identifiers; this protocol does not verify them against implementation, prompt, model, or tool-schema bytes.
- Exact raw corpus and fixture UTF-8 is embedded verbatim in reports and packets; bundled synthetic status is digest-verified, while custom synthetic status is operator-attested and not semantically verified.
- Synthetic fixture outputs do not estimate evidence-insensitivity prevalence in real models.