Solonic

Methods · free, forever, no email wall

Everything we know, published.

Democratized intelligence should mean democratized methods. Every gate, prompt, protocol, and agent architecture we use is public. You can run our entire loop without us, and we would rather you did.

The loop

What we actually do

The Neurath gate

Label every claim PROVED, VERIFIED, OPEN, CORRECTED, or REFUTED. Exact arithmetic. Errata filed the day an error is found. Nothing ships untagged.

Multi-family review

Independent checkers with uncorrelated failure modes. A chain is only as calibrated as its weakest link; correlated checkers do not add up. Up to twelve families at ρ ≈ 0.2 behave like about four independent ones — we design for that, and we say so.

Adversarial critique

Attack the argument until it breaks or survives. Treat a surviving claim as evidence, not proof, and name what would falsify it.

The Ledger

Publish the entries that went against you. A method that only reports successes is a method for producing successes, not truths. Ours is here.

Higher-order evidence

When you cannot evaluate the content, evaluate the process. Gates, track records, calibration, audit trails — and the theorems governing when a chain of deference is rational at all.

The packages

All of the above, wired together and downloadable. Take one.

Our moat isn't secrecy. It's a public track record of being right, and of saying so when we weren't.

The Roster

Model families in the Jürge pipeline

Up to 12 model families. Effective independence ρ ≈ 0.2 (approximately 4 independent reviewers). Availability varies per run — each report names which families returned.

DeepSeek R1
Reasoning-focused architecture, strong mathematical capability
Mistral Medium
European-trained, balanced instruction-following
Llama 4 Maverick
Open-source foundation, Meta training regime
Gemma 4
Google instruction-tuned, efficiency-optimized
MiniMax
Chinese frontier model, distinct training data
Kimi K2.6
Long-context capability, Chinese language heritage
GLM 5.1
Zhipu architecture, multi-modal capable
DeepSeek V4 NIM
Inference-optimized variant, low-latency tier
GPT-4o
OpenAI flagship, broad training distribution
Command A
Cohere instruction-tuned, specialized for structured output
Jamba
AI21 Labs hybrid architecture
Grok
xAI training, real-time information access

Why 12? Objection saturation curves from the corpus show new objections emerging through reviewer 8–9, then flattening. Twelve is the minimum for reliable saturation. Why ρ ≈ 0.2? Independence measured as Spearman correlation among raw scores across the 256-paper corpus. Despite distinct training data, models share similarity due to overlapping internet training sources and architectural borrowing. Availability varies: API outages, rate limits, and latency constraints mean not every run reaches all 12. Each report names which families returned and which did not.

If you'd rather be taught

Classes

We teach the whole loop — generation, gates, ledger discipline, errata practice — to teams who want to run it themselves. Cohorts are small and scheduled by demand. Write to us and say what you work on.