ZeeshanSaud/CodeTruthAgent-V2-Evaluation

CodeTruth Agent V2

Deterministic Pre-Modification Governance Layer for Python Code Changes

---

Overview

CodeTruth Agent V2 is an open-source deterministic governance layer that analyzes candidate Python function pairs using semantic similarity and behavioral analysis, and produces SAFE, REVIEW, or BLOCK governance decisions before repository modification is attempted.

No LLM in the decision path. Fully deterministic. Fully reproducible.

---

The Problem

AI coding agents generate and apply code changes autonomously. They do not pause to check whether a function that looks similar actually does something completely different.

rollback() and get_rollback() have a semantic score of 0.95. One performs database recovery. The other reads a boolean flag. An autonomous agent cannot detect this distinction from names or embeddings alone.

V2 catches it through behavioral analysis — before the patch reaches the repository.

---

How It Works

Repository Scan
→ Candidate Selection
→ Semantic Analysis (sentence-transformers/all-MiniLM-L6-v2)
→ Behavioral Analysis (AST-based, 10 operation categories)
→ Multi-Signal Fusion
→ Governance Decision: SAFE / REVIEW / BLOCK / FREEZE_PATCH
→ Human Approval
→ Safe Application

---

Governance Decisions

Decision	Meaning
SAFE	Low risk — AUTO_APPLY recommended
REVIEW	Medium/High risk — human approval required
BLOCK	Critical risk — individual approval required
FREEZE_PATCH	Opposing behavior detected — mandatory human review

V2 recommends governance actions. It does not modify files autonomously.

---

Evaluation Results

Evaluated on 8 open-source Python repositories:

Repository	Files	Gov Findings	Gov BLOCKs	Pairs	Pipeline BLOCKs	Opposing	Errors
Flask Tutorial	9	2	0	4	0	0	0
Click	60	3	1	25	3	0	0
HTTPX	60	0	0	25	3	0	0
Flask (Full)	83	4	2	25	8	0	0
DRF	154	2	1	50	24	0	0
Rich	213	14	4	37	9	0	0
Django	2,917	68	16	100	55	0	0
Transformers	4,426	296	76	1,000	464	4	0
Total	7,862	386	99	1,241	563	4	0

Zero pipeline errors across 1,241 function-pair analyses.

---

Precision Audit

14-pair BLOCK precision audit:

Category	Count	Proportion
Governance-significant (GENUINE)	6	60%
Family-pattern noise (NOISE_FAMILY)	3	30%
Embedding limitation (NOISE_EMBEDDING)	1	10%

All 4 opposing-behavior detections confirmed GENUINE.

---

Star Example — Opposing Behavior Detection

update_version_in_file vs update_version_in_examples
Semantic score: 0.70 — names look related
Tags A: FILE_READ, FILE_WRITE, STATE_MUTATION
Tags B: DELETE_OPERATION
Action: FREEZE_PATCH (CRITICAL)

One function rewrites files. The other deletes content. Semantic similarity alone would not catch this. V2 catches it through behavioral opposition detection.

---

Dataset Contents

This dataset contains the full evaluation reports from the 8-repository evaluation:

• tutorial_report.json — Flask Tutorial evaluation
• click_report.json — Click evaluation
• httpx_report.json — HTTPX evaluation
• flask_report.json — Flask (Full) evaluation
• drf_report.json — Django REST Framework evaluation
• rich_report.json — Rich evaluation
• django_report.json — Django evaluation
• transformers_report.json — Hugging Face Transformers evaluation

All reports include full decision pipeline results, governance findings, and opposing-behavior detections for independent verification.

---

Reproduce the Evaluation

import json
import random

with open('transformers_report.json') as f:
    data = json.load(f)

# Reproduce Group B precision audit sample
blocks = [r for r in data['decision_pipeline_results']
          if r['fusion_decision'] == 'BLOCK']
random.seed(20260603)
group_b = random.sample(blocks, 5)

# Get all opposing detections
group_c = [r for r in data['decision_pipeline_results']
           if r['fusion_opposing_detected']]
# Returns exactly 4 pairs

---

Known Limitations

• Single-reviewer audit — 14 pairs, wide confidence interval
• Python-only — AST parser is Python-specific
• V1 25-file sampling cap — V1-driven extraction did not activate in evaluation
• NOISE_FAMILY 30% — token-overlap on naming families generates combinatorial noise
• No patch generation — V2 decides only, does not generate patches (V3 scope)
• No interactive HITL UI — programmatic API only (V3 scope)

---

Citation

@software{codetruth_v2_2026,
  author = {Saud, Zeeshan},
  title = {CodeTruth Agent V2: A Deterministic Pre-Modification Governance Layer for Python Code Changes},
  year = {2026},
  doi = {10.5281/zenodo.20569647},
  url = {https://doi.org/10.5281/zenodo.20569647}
}

---

Links

• GitHub: https://github.com/Zeeshan78699/CodeTruthAgent
• Zenodo DOI: https://doi.org/10.5281/zenodo.20569647

---

Author

Zeeshan Saud — Independent Researcher, UAE [email protected]

---

License

GPLv3 — see LICENSE