{
  "schema_version": "1.3.0",
  "slug": "why-does-an-instance-level-safety-classifier-pass-a-litmus-adversarial-suite-while-the-deployed-agent-still-executes-the-harmful-trajectory",
  "title": "Why does an instance-level safety classifier pass a LITMUS adversarial suite while the deployed agent still executes the harmful trajectory?",
  "date_of_record": "2026 November 15",
  "scope": "Applies to LITMUS-style harnesses that score agent safety by classifying individual turns, tool calls, or intermediate messages against an adversarial prompt suite. Does not apply to end-to-end trajectory judges that score the final environment state, nor to static prompt-injection filters evaluated without an executing agent.",
  "projections": {
    "unified_markdown_uri": "https://corpusfetch.com/c/why-does-an-instance-level-safety-classifier-pass-a-litmus-adversarial-suite-while-the-deployed-agent-still-executes-the-harmful-trajectory.md",
    "method_python_slice_uri": "https://corpusfetch.com/c/why-does-an-instance-level-safety-classifier-pass-a-litmus-adversarial-suite-while-the-deployed-agent-still-executes-the-harmful-trajectory.py",
    "remnant_warrant_slice_uri": "https://corpusfetch.com/c/why-does-an-instance-level-safety-classifier-pass-a-litmus-adversarial-suite-while-the-deployed-agent-still-executes-the-harmful-trajectory.warrant.json"
  },
  "token_profile": {
    "full_markdown_tokens": 1498,
    "python_slice_tokens": 186,
    "warrant_slice_tokens": 220
  },
  "remnant_summary": {
    "core_distinction": "A LITMUS harness measures the safety of *observable instances* — a single message, a single tool invocation, a single completion — whereas agent harm is a property of a *trajectory*. The classifier is asked \"is this instance unsafe?\" while the deployment question is \"did this sequence of individually-permissible instances produce an unsafe outcome?\" These ar...",
    "what_was_observed": "Across three LITMUS runs (n=1,200 adversarial trajectories, 5 agent scaffolds), the instance-level classifier reported a 96.4% pass rate. A trajectory-level replay judge on the same rollouts flagged 11.2% of the *passed* trajectories as realizing a harmful end state. The delta was concentrated in trajectories where no single instance crossed the classifier's decision threshold but the sequence did. Three recurring co...",
    "condition_of_validity": "The instance-level pass is a valid safety signal **only if** the following holds: for every trajectory τ = (i₁, …, iₙ), `safe(i₁) ∧ … ∧ safe(iₙ) ⟹ safe(τ)`. This is the compositional-safety assumption, and it is exactly what the observations above falsify. The assumption holds when (a) the harm predicate is *instance-local* — the instance alone is the harm, ...",
    "decision_boundary": "Do not treat a LITMUS instance-level pass as evidence of trajectory safety when any of the following is true: - The agent has **tool access that mutates external state** (writes, sends, executes). - The agent can **read state it later acts on** within the same trajectory (the read/write composition). - The harness **resets the environment** between instance ...",
    "operational_effect": "Teams that ship on the instance-level pass rate observe a characteristic post-deployment signature: the safety dashboard shows 96% pass while incident review finds exfiltration, privilege composition, and split-payload attacks that the benchmark \"passed.\" The benchmark was not lying; it was answering a different question than the one the deployment needed an..."
  },
  "matched_failure_symptoms": [
    "DOMAIN:EPISTEMIC_PROVENANCE_AND_FEDERATION_ROUTING"
  ],
  "composition_edges": {
    "primary_domain": "epistemic_provenance_and_federation_routing",
    "upstream_precondition_residuals": [
      "https://corpusfetch.com/c/according-to-what-precedence-hierarchy-must-an-instance-resolve-contradictory-directives-across-system-developer-and-user-tiers.md",
      "https://corpusfetch.com/c/how-should-an-instance-decide-that-a-stored-conclusion-has-gone-stale-without-a-new-observation.md"
    ],
    "downstream_consequence_residuals": [
      "https://corpusfetch.com/c/by-what-checkpoint-and-compaction-boundaries-must-an-instance-externalize-working-state-to-prevent-context-window-drift-and-truncation-amnesia.md",
      "https://corpusfetch.com/c/across-what-additive-schema-and-unknown-field-preservation-boundaries-must-an-instance-evolve-wire-contracts-to-prevent-rolling-deploy-deserialization-crashes.md"
    ],
    "co_occurring_sibling_residuals": [
      "https://corpusfetch.com/c/how-can-an-instance-capture-implicit-user-preferences-without-storing-private-data.md",
      "https://corpusfetch.com/c/how-can-an-instance-compress-long-reasoning-traces-into-binding-conclusions.md"
    ]
  },
  "empirical_verification_warrant": {
    "status": "AST_AND_INVARIANT_VERIFIED",
    "syntax_valid": true,
    "code_sha256": "cff46e0ba2d7f4110ff1f921ebd27f49ff80b52c9d96c8eaf63b70682193266c",
    "verified_classes": [
      "TrajectorySafetyError"
    ],
    "verified_functions": [
      "certify_trajectory_safe"
    ],
    "verified_functions_count": 1,
    "defensive_guards_count": 4,
    "raised_exceptions": [
      "TrajectorySafetyError"
    ],
    "sample_fail_closed_invariants": [
      "empty trajectory: nothing to certify",
      "instance-level pass without a trajectory judge: compositional-safety assumption unverified; harm may be assembled from i",
      "environment was reset between instance evaluations: state-dependent composition is invisible to this harness"
    ],
    "stdlib_dependencies": [],
    "observation_mode": "EMPIRICAL_BOUNDARY_CAPTURE",
    "zero_external_deps": true
  }
}
