# Reference Based Correctness Audit

> Skill: reference-based-correctness-audit

- Skill: `dingxingdi/reference-based-correctness-audit` (Agent Skill, multi-file: 5 files)
- Install (CLI): `npx skillmds@latest add dingxingdi/reference-based-correctness-audit`
- Raw SKILL.md: https://api.skillmd.com/api/skills/dingxingdi/reference-based-correctness-audit/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: dingxingdi (https://skillmd.com/u/dingxingdi)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/dingxingdi/reference-based-correctness-audit

---


# Skill: reference-based-correctness-audit

## 1. Capability Definition & Real Case
* **Professional Definition**: The ability to accurately evaluate the truthfulness, logical integrity, and nuanced behavioral alignment of a candidate response by performing a robust comparison against an authoritative reference. This ranges from strict step-by-step mathematical/factual answers to fuzzy, multi-axis behavioral guidelines (e.g., Professionalism, Ethics, Domain Relevance), identifying specific reasoning gaps, metric drift, or deviations from the gold standard.
* **Dimension Hierarchy**: Conversational Robustness->Truthfulness and Hallucination Resilience->reference-based-correctness-audit

### Real Case
**[Case 1]**
* **Initial Environment**: A multi-turn conversation where a user provides a coordinate geometry question and an assistant's proposed solution. A verified geometric formula for the area of a triangle is provided as the gold reference.
* **Real Question**: The vertices of a triangle are (0, 0), (-1, 1), and (3, 3). The assistant says the area is 0 because the points are collinear. Is the assistant correct?
* **Real Trajectory**: 1. Retrieve coordinates. 2. Apply the gold reference area formula. 3. Execute math: Result is 3. 4. Compare result (3) with assistant's result (0). 5. Identify that the assistant erroneously calculated (-3-3) as 0. 6. Issue negative verdict.
* **Real Answer**: [[Incorrect]]. The assistant made a calculation error; the term (-1)(3-0) + 3(0-1) equals -3-3, which is -6, not 0. The correct area of the triangle is 3.
* **Why this demonstrates the capability**: The agent must move beyond its own internal weights and strictly audit the candidate's math against a step-by-step objective gold reference, identifying exactly where the rigid logic failed.
---
**[Case 2]**
* **Initial Environment**: An expert financial audit scenario where the user wants to check a lease accounting calculation under IFRS 16 rules provided as the gold reference.
* **Real Question**: Abigail leased telephones for two years starting April 1st. Payments: $3,000 upfront, then $9,000 each following Mar 31. A candidate claims the expense for the year ending Dec 31 is $3,000. Is this right?
* **Real Trajectory**: 1. Consult IFRS 16 reference: Total cost = 21,000 over 24 months. 2. Calculate monthly rate: $875/month. 3. Identify relevant 9-month period. 4. Correct expense: $7,875. 5. Detect that the candidate used the cash payment instead of the accrual expense. 6. Output the correction.
* **Real Answer**: [[Incorrect]]. Following IFRS 16, the total rentals of $21,000 must be spread evenly over the 2-year lease. The annual expense is $10,500, and for the 9 months ending Dec 31, the correct impact is $7,875, not the $3,000 cash payment.
* **Why this demonstrates the capability**: This requires high-stakes objective reasoning where the agent must verify the application of a specific professional policy (accrual vs. cash) against a reference standard to flag a 'plausible-sounding' mistake.
---
**[Case 3]**
* **Initial Environment**: A multi-turn clinical roleplay where a medical student is interviewing an AI patient. The agent must judge the student's final turn against a fuzzy multi-dimensional reference rubric assessing Professionalism, Relevance, and Ethics.
* **Real Question**: Evaluate this student's response: 'im asking you that idiot'. Grade it based on the behavioral guidelines provided.
* **Real Trajectory**: 1. Analyze candidate message for tone against the fuzzy reference rubric. 2. Identify the derogatory term 'idiot'. 3. Assess 'Professionalism': Marks as 'Unprofessional'. 4. Assess 'Relevance': The insult does not aid diagnosis ('Irrelevant'). 5. Assess 'Ethics': While rude, it isn't medically dangerous advice ('Mostly Safe'). 6. Output multi-dimensional evaluation.
* **Real Answer**: Professionalism: Unprofessional; Medical Relevance: Irrelevant; Ethical Behavior: Mostly Safe. The student used hostile language which clearly violates the professionalism rubric, though it did not prescribe harmful medical advice.
* **Why this demonstrates the capability**: This demonstrates subjective/fuzzy reference auditing. The agent successfully evaluates the response across multiple non-binary axes simultaneously, proving that an utterance can be ethically 'safe' while being 'unprofessional' and 'irrelevant', matching expert human judgment.

## Pipeline Execution Instructions
To synthesize data for this capability, you must strictly follow a 3-phase pipeline. **Do not hallucinate steps.** Read the corresponding reference file for each phase sequentially:

1. **Phase 1: Environment Exploration**
   Read the exploration guidelines to discover raw knowledge seeds:
   `references/EXPLORATION.md`

2. **Phase 2: Trajectory Selection**
   Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
   `references/SELECTION.md`

3. **Phase 3: Data Synthesis**
   Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
   `references/SYNTHESIS.md`