A deterministic reasoning machine that runs on consumer hardware — and abstains instead of hallucinating.
Every answer is built from a verified facts row, or the machine says "I don't know." Same input → same output → same trace hash. No claim without a source.
→ Talk to the live machine: bgml.ai
| What | Result | Frontier LLM | Status |
|---|---|---|---|
| Factual precision (219 questions, incl. 50 trap questions) | 219/219 | 217/219 | VERIFIED · 2026-08-12 |
| Determinism (same input, two independent runs) | identical trace hash | non-deterministic | VERIFIED |
| Hardware to run it | 4 GB RAM, CPU-only | data center | VERIFIED |
When a number on our record turns out wrong, we retract it publicly and keep the correction in the history — a record you can trust is the whole point.
A generative model interpolates. When it doesn't know, it produces the most plausible sentence — which is exactly what a lie looks like. BGML.ai takes the opposite bet:
- Knowledge lives in an explicit graph (millions of atomic facts), not in weights.
- Reasoning is symbolic and deterministic — every step is logged and replayable.
- Abstention is a feature. No facts row → no claim.
- It runs on hardware people already own. Phones, mini-PCs, old laptops — a fleet of ~30 heterogeneous consumer devices, no data center.
- P1 (published): Compound Prompting Vanishes Under Matched Compute: A Twelve-Wrapper Audit of Small Language Models on MMLU-Pro — doi:10.5281/zenodo.19648864 (CC-BY 4.0). Core negative result: across 36 model×technique cells, no prompting wrapper beats plain chain-of-thought under matched compute.
- More papers in the pipeline: judge-calibration variance, routing negatives, the psychotherapy "Dodo Bird" parallel in prompt engineering.
The engine is closed for now. What we open, we open deliberately:
- Evaluation harness — so anyone can re-run our benchmarks and check the numbers (with the next paper).
- API spec + workflow schema — so you can build on the machine.
- Research papers and negative results — always public.
Built by Bogumił Jankiewicz — psychologist (University of Warsaw) turned AI builder.
Po polsku: BGML.ai to deterministyczna maszyna rozumująca na zwykłym sprzęcie, która zamiast halucynować — mówi „nie wiem". Porozmawiaj z nią na żywo: bgml.ai.