Working Paper¶
The full research paper documenting ml-lab's adversarial debate evaluation methodology. Covers the problem statement, related work, methodology design, experimental results across multiple versions, and conclusions.
Source location
The working paper is maintained at WORKING_PAPER.md in the repository root. This page provides context; the canonical version lives there.
Abstract¶
ml-lab uses an adversarial critic-defender debate protocol to evaluate ML methodology claims. This paper presents the protocol design, describes eight iterative experiments that calibrated the methodology, and reports detection and verdict accuracy on benchmark cases with known ground truth.
Submission tracks¶
The paper is being prepared for:
| Venue | Track | Status |
|---|---|---|
| arXiv | Preprint | In preparation |
| EMNLP 2026 | Conference | In preparation |
| NeurIPS 2026 | Conference | In preparation |
Submission-specific checklists and formatting are tracked in paper/ subdirectories.
Key results¶
- Detection performance: validated across v2–v7 with pre-registered hypotheses
- Ambiguity judgment: validated (system correctly identifies genuinely ambiguous cases)
- Defense-case calibration: pending (v8, active)
- Cross-vendor stability: evaluated in v6
For detailed per-version results, see the Experiment Reports.