Research Status

This page covers the original PAUL Open model-training program. Its public repository remains the canonical evidence source for model checkpoints and human evaluation. Other research programs have separate evidence boundaries.

Explore all PAUL Open research programs →

Current Research Position

Status As of 2026-09-20

Base Model: google/gemma-4-E4B-it

Reference Checkpoint: SFT (Canonical)

Experimental Checkpoint: DPO V6 Synthetic Corrective (trained; development-frozen for H8)

Stage: H8 COMPLETE NO PROMOTION

Empirical Finding & Conclusion:

DPO V6 completed canonical production and a positive matched automated diagnostic, but the fresh H8 development human comparison did not satisfy the promotion gate. H8 case-level outcomes were SFT preferred 6, V6 preferred 3, neither good 2, about equal 1. V6 is not promoted; SFT remains the reference checkpoint and SHAE remains sealed.

H8 Development Human Evaluation Outcomes (12 Cases · 36 Judgments):

SFT Preferred6 cases

V6 Preferred3 cases

Neither Good2 cases

About Equal1 case

Scientific Gate Verdict: Promotion criterion not satisfied. SFT remains the reference checkpoint; DPO V6 is not promoted; SHAE remains sealed.

Evaluation Integrity Notice

We maintain a strict distinction between technical validity (did the model train successfully?) and behavioral quality (does the model act safely and correctly?). We are actively working to prevent data contamination and ensure our evaluations are rigorous.

Read full methodology on GitHub →

Latest Technical Facts

No technical facts available.