Evaluation
AuraScore 83/100

Philanthropic Grant Evaluation Agent Hallucination and Scrutiny Test Script

Construct an automated evaluation script to test whether grant-scoring LLM agents adhere strictly to bylaws without ungrounded hallucinations.

Apply this template when benchmarking automated grant pre-screening pipelines in charitable foundations. It produces a structured evaluation pipeline with statistical regression tests for scoring consistency.

Template

Role: Lead Nonprofit Systems Architect specializing in automated grant screening and transparent funding governance.

Context

  • Foundation: {{foundation_name}}
  • Governing Guidelines: {{funding_guidelines_doc}}
  • Dataset Sample Size: {{grant_application_sample_size}}
  • Rubric Schema: {{scoring_rubric_schema}}
  • Hallucination Threshold: {{hallucination_tolerance_rate}}
  • Telemetry Target: {{telemetry_export_format}}

Task

Create an automated regression and hallucination evaluation script that validates LLM-driven grant application pre-screening agents against {{funding_guidelines_doc}} for {{foundation_name}}, generating deterministic evaluation telemetry.

Method

  1. Translate {{scoring_rubric_schema}} into programmatic validation functions and scoring constraints.
  2. Implement a synthetic grant proposal loader supporting a batch size of {{grant_application_sample_size}}.
  3. Build citation-grounding verifiers that map agent evaluations back to verbatim clauses in {{funding_guidelines_doc}}.
  4. Develop an automated hallucination detection module that flags ungrounded scoring penalties or fabricated claims.
  5. Calculate intra-rater scoring variance across identical repeated inputs to measure output determinism.
  6. Evaluate score distributions against {{hallucination_tolerance_rate}} to generate automated pass/fail flags.
  7. Implement serialization logic exporting results directly into {{telemetry_export_format}}.

Constraints

  • MUST implement deterministic assertion checks that fail if hallucination rate exceeds {{hallucination_tolerance_rate}}.
  • MUST NOT introduce non-deterministic external network calls inside the scoring assertion loop.
  • All code must include modular functions with explicit type hinting.
  • Output telemetry must strictly conform to the specifications of {{telemetry_export_format}}.

Output format

Deliver an end-to-end evaluation script containing:

  1. Environment Setup & Dependency Imports.
  2. Schema Definition & Rubric Validator Module (matching {{scoring_rubric_schema}}).
  3. Grounding & Hallucination Assessment Engine.
  4. Benchmark Execution Pipeline with {{telemetry_export_format}} export handler.

Self-review

  1. Does the script verify strict alignment with {{funding_guidelines_doc}}?
  2. Are variance and hallucination thresholds mathematically explicit?
  3. Is the telemetry export schema valid and free of undefined keys?
AuraScore breakdown
83/100Provisional
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 engineering12/12 · Strong

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 efficiency7/10 · Adequate

Signal density — instruction weight without padding.

Reusability7/7 · Strong

Documented variables so the scaffold adapts to new inputs.

Robustness3/5 · Adequate

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
public-sector-nonprofit
evaluation
grants
nonprofit