KRG · Knowledge-Reasoning Gradient
d = 0.04 → 2.36
How much grounding outperforms a vanilla model as task knowledge-dependency rises.
Spearman ρ=0.89, p=0.0006. 11 models, 4 domains.
Trust = Alignment × Reliability. Industry built one half.
A research lab measuring the missing half — three peer-reviewed alignment metrics across 13 benchmarked models.
NeurIPS 2026 · 94.9% inter-judge agreement — Evaluations & Datasets Track, in review
Research — alignment metrics
We measured the missing half.
Three peer-reviewed metrics make agent judgment auditable. Numbers below are from the methodology submitted to NeurIPS 2026.
JIS · Judgment Inflation Score
1.83 ± 0.42
Ratio of model score to ground truth — how much your AI flatters you.
JIS=1.0 calibrated. JIS=2.6 inverts decisions. HALA grounded: 1.00.
DEP · Disconfirming Evidence Prompt
−48 pp
Pressure-sycophancy reduction from a single-sentence intervention (Claude Sonnet 4).
65–87% pressure cut on RLHF models. Zero effect on factual sycophancy.
Validated across 13 models, 4 domains, 3 benchmarks. Inter-judge agreement 94.9%. NeurIPS 2026 Evaluations & Datasets Track (in review).
Read the research →The position