PAUL Open brings together PAUL Foundry's research in medical imaging, molecular science, sports analytics, developer agents and reproducible open-model evaluation. Each program has its own evidence and publication boundaries.
Research organization
Six distinct research activities, separate evidence boundaries, and honest publication status.
Medical imaging / RSNA
Knee MRI abnormality detection with leakage-resistant evaluation and strict medical research boundaries. Private Kaggle notebooks, results and checkpoints are not published here.
Computational chemistry / CASMI
Small-molecule structure inference from tandem mass spectra. The research program emphasizes provenance, reproducibility and guarded validation; no public benchmark result is claimed.
Sports data & modelling
A separate sports-focused research direction. Competition names, verified evidence and outcomes will be added only after a public release is authorized.
Public results not yet released.
Agentic software engineering
Exploring coding-agent architectures, repository understanding, bounded execution and patch verification. Challenge listing does not imply an award or rank.
Agent research & evaluation
The separate Gemma developer-agent paper track for methodological research. No paper acceptance, publication or prize is claimed.
Multilingual AI & human evaluation
The original PAUL Open Gemma-based SFT/DPO experiments and blinded evaluation. SFT remains the reference checkpoint; candidate improvements require human validation.
Rather than treating AI as an infallible oracle, PAUL Open structures science learning around mechanisms, assumptions, guided questions, and empirical evidence.
Concise scientific explanations detailing the underlying physical or biological mechanism, explicit boundary assumptions, and clarification of common misconceptions.
Socratic guided inquiry that begins with a diagnostic check, guides the learner step-by-step through first principles, and tests understanding with a conceptual checkpoint.
Advanced scientific reasoning that separates empirical observation from inference, flags confounding variables, and helps formulate practical validation experiments.
The mobile companion uses a secure server-side runtime powered by Google's Gen AI SDK with zero client-side secrets. The production companion runtime is deliberately distinct from our experimental research models (such as DPO V6), ensuring stability while keeping research transparent.
Watch how we orchestrate Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and multi-turn human evaluation across Indic multilingual benchmarks.
Empirical Telemetry & Research Transparency
In PAUL Open, unsuccessful candidate checkpoints are openly documented rather than hidden. Our research story is simple: Candidate checkpoint (DPO V6) → Promising automated diagnostic (90.62 vs 89.18) → Fresh double-blinded human evaluation (H8) → Promotion criterion not met (SFT preferred 6, V6 3) → SFT retained as canonical reference; SHAE sealed.
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.
Scientific Gate Verdict: Promotion criterion not satisfied. SFT remains the reference checkpoint; DPO V6 is not promoted; SHAE remains sealed.
Automated scores measure token overlap and rubric heuristics. Human evaluation measures genuine truthfulness, pedagogical guidance, and scientific soundess. Explore our methodology, blinding protocol, and results.
Explore our core scientific and pedagogical domains. Every benchmark is grounded in reproducible splits, public rubrics, and automated safety gates.
Indian language alignment, syntax-preserving translation, and culturally nuanced reasoning across Hindi, Bengali, and English.
Explore Telemetry →Pedagogical evaluation across K-12 and undergraduate STEM concepts. Enforces guided inquiry over direct answer disclosure.
Explore Tutoring Protocol →Domain-specific factual verification, biochemical accuracy, and strict Research Use Only safety boundaries.
Explore Evidence Limits →