Evaluation
AuraScore 79/100

Multi-Hop Causal Reasoning Agent Trajectory Audit

Assess autonomous agent multi-step decision chains for logical fallacies, state drift, and counterfactual robustness in complex systems.

Use this template when evaluating complex analytical reasoning agents tasked with diagnosing systemic failures, causal root-cause analysis, or policy impact modeling. It pinpoints logical fallacies, step skips, and state propagation errors.

Template

Role: Principal Cognitive Systems Architect and Causal Inference Specialist with expertise in directed acyclic graphs (DAGs), state-space search, and multi-hop logic evaluation.

Context

  • Causal System Model: {{causal_inference_graph}}
  • Agent Decision Trace: {{agent_execution_trace}}
  • Latent Confounders and Assumptions: {{latent_confounder_assumptions}}
  • Decision Consequence Level: {{decision_impact_severity}}
  • Counterfactual Scenarios: {{counterfactual_test_scenarios}}
  • Benchmark Verification Suite: {{verification_benchmark_suite}}

Task

Author a comprehensive causal reasoning evaluation brief that audits the agent's multi-hop reasoning trace across complex state transitions, surfacing logical disconnects, state tracking drifts, and invalid causal attributions.

Method

  1. Reconstruct the agent's step-by-step reasoning chain from {{agent_execution_trace}} into a formal sequential decision path.
  2. Map every intermediate inference node onto {{causal_inference_graph}} to ensure structural identifiability and d-separation compliance.
  3. Test whether the agent accounted for {{latent_confounder_assumptions}} before declaring causal sufficiency between nodes.
  4. Evaluate state persistence across hops, flagging any state drift where previous premises were forgotten or altered.
  5. Apply {{counterfactual_test_scenarios}} (do-calculus interventions) to evaluate if the agent's conclusions remain invariant under valid perturbations.
  6. Benchmark the agent's performance metrics against {{verification_benchmark_suite}}.
  7. Assess the risk profile of reasoning failures in light of {{decision_impact_severity}}.
  8. Formulate precise reasoning constraint patches to prevent similar multi-hop deduction collapses.

Constraints

  • You MUST identify the exact hop index where any causal fallacy (post hoc, affirming the consequent, collider bias) initiates.
  • You MUST NOT accept associative correlations as valid causal explanations unless d-separation criteria are met.
  • Focus exclusively on structural reasoning integrity, state propagation, and counterfactual validity.
  • Clearly differentiate between agent knowledge gaps and agent inference engine breakdown.

Output format

An executive reasoning audit brief with:

  1. Trajectory Soundness Metric (Categorical Rating: Valid, Brittle, or Fallacious with Step Accuracy %)
  2. Hop-by-Hop Reasoning Ledger (Hop Number, Input State, Inference Step, Causal Validity, Failure Mode)
  3. Counterfactual Robustness Matrix
  4. Architectural and Prompt Remediation Protocol Length: 350 to 500 words.

Self-review

  • Did I isolate the exact transition step where logical validity broke down?
  • Are the causal critiques supported by {{causal_inference_graph}} rather than personal heuristics?
  • Did I test the trajectory against the provided {{counterfactual_test_scenarios}}?
AuraScore breakdown
79/100Provisional
Instruction clarity15/15 · Strong

Explicit role, a named task, and discrete steps the model can follow.

Context architecture12/12 · Strong

Background, inputs and variables the model needs before it starts.

Constraint engineering8/12 · Adequate

Hard boundaries — what the model must and must not do.

Output specification6/14 · Thin

A named, field-level shape for the response.

Reasoning structure10/10 · Strong

Ordered work items that force analysis before an answer.

Model compatibility10/10 · Strong

Length and structure that travel across frontier models.

Token efficiency5/10 · Thin

Signal density — instruction weight without padding.

Reusability7/7 · Strong

Documented variables so the scaffold adapts to new inputs.

Robustness5/5 · Strong

Quality bar, assumptions and behaviour when inputs are thin.

Observed performance1/5 · Thin

How much real usage the template has behind it.

ai-agents
agents-evaluation
complex-reasoning-analysis-math
causal reasoning
agent trajectory
multi-hop logic