Dhruv1000/vulnguard-code-and-report

VULNGUARD — Code, Results, and Manuscript

Private repository. This is the complete working mirror of the VULNGUARD project: an RL/GRPO study of whether reinforcement learning can teach an LLM to discriminate vulnerable vs. patched C/C++ functions. It contains all code, all run logs and results, the full decision/incident history, and the project's manuscript.

One-line summary of the science: this is a negative-result study. Every gold-reward GRPO configuration tried — across 5 backbones, multiple credit-assignment schemes, and two purpose-built collapse-prevention mechanisms (Campaign D5) — collapsed to a degenerate, near-single-verdict policy instead of learning real discrimination. The dominant mechanism is only partially understood; see below for what is and isn't resolved.

Start here

If you want...	Read
The full story, in order, with every hypothesis/result/bad outcome/timing	docs/COMPREHENSIVE_REPORT.md
A shorter narrative version (original write-up, Campaign D5 only)	docs/FULL_PROJECT_REPORT.md
The academic manuscript (with figures, references, methods)	paper/manuscript.md + paper/figures/
The current, PI-accepted scientific conclusion (most authoritative single statement)	docs/validation/CLAIM_MATRIX.md
The raw, timestamped, append-only decision log (ground truth for everything above)	docs/DECISIONS.md
Every logged error/deviation from the frozen pre-registration	docs/ERRATA.md
The frozen experimental design itself	docs/PREREGISTRATION.md
The verified citation list used by the manuscript	docs/LITERATURE.md
Campaign D5's spec (the two collapse-prevention interventions)	docs/CAMPAIGN_D5.md

Repository layout

src/            training/eval/reward/queue library code (vdrl package)
scripts/        one-off and campaign-specific scripts (data prep, queue, D5 forensics, toy sims)
tests/          pytest suite (pre-registration rule: no GPU job may run while this is red)
configs/        frozen model/queue/base configs
data/           processed splits, audits, provenance manifests (NOT the raw corpora)
docs/           all narrative reports, the decision log, pre-registration, errata, literature
results/        every run's metrics.jsonl, gates, watchdog logs, scoring outputs
  results/campaign_D5/      Campaign D5 forensics writeups (H3/H4/H5 scoring + mechanism analysis)
  results/final_scientific_audit_20260927/   primary confirmatory result dossier
  results/validation_20260926/, results/preservation_20260926/   human-validation scaffolding (packets only — no labels, no private data)
paper/          the manuscript (Markdown) and its 4 real-data figures
phase2/         a separate, authorized research branch (worktree)

What is NOT in this repository, by design

Human-rater blinding-key mapping files, the underlying private/blinding data for the never-completed human validation study, and anything inside integrity-backups/ that carries that same content are deliberately excluded from every backup of this project, including this one. This is a standing project rule, not an oversight — see docs/DECISIONS.md entries tagged privacy-incident and preservation for the full reasoning.

Headline results (see docs/COMPREHENSIVE_REPORT.md for full detail and evidence paths)

• Primary confirmatory result: GOLD − PERMUTED discrimination Δ = −1.04 pp [95% CI −1.52, −0.62] — gold-reward training did not produce more discrimination than a placebo reward. Not confirmatory: gated on human validation that was never completed (0 labels ever entered).
• Cross-backbone replication: the collapse reproduces on Llama-3.2-3B (step 8), Qwen3-8B (step 12), and Qwen3-4B-hybrid (step 14), in addition to the primary Qwen3-4B-Instruct-2507 backbone (step 18).
• Campaign D5 tested two mechanisms specifically designed to prevent collapse — a verdict-rate dual-ascent clamp and a pair-symmetric reward. Neither worked. The clamp collapsed to always-SAFE anyway (mechanism: its correction signal is a within-group-standardized quantity that is mathematically zero once a group has already collapsed, regardless of the penalty's magnitude); the pair-symmetric reward collapsed to the opposite corner (always-VULNERABLE) instead.
• A reusable judge-calibration finding: an LLM judge's false-positive rate, calibrated at 2% on a synthetic benchmark (NIST Juliet), was found to actually sit near 48% on real code from the same source projects.
• A positive control (a deliberately easy synthetic task run through the identical pipeline) reached 100% held-out accuracy, indicating the training infrastructure itself is not the bottleneck.

Citation

If referencing this work, please cite the manuscript in paper/manuscript.md (Das, D. J., When the Reward Collapses: A Pre-Registered Study of Gold, Constrained, and Pair-Symmetric RL for LLM Vulnerability Detection).