No prompt engineering. No retrieval-augmented generation. A real curriculum-gated learning loop, with consensus verification, hash-chained audit, and zero hallucination tolerance.
If you know much about programming, you'll really enjoy this.
In a single twelve-hour window on May 16-17, 2026, the Praefex brain progressed from "no curriculum module" to 73% mastery of the first persona's five-level curriculum, fully autonomously, on commodity hardware. Every step is on-disk, hash-linked, replayable.
Every other "AI assistant" project we've reviewed this year shares the same default failure mode: confident hallucination. A single model produces fluent output, no verification, no audit, no admission of ignorance. Praefex is the architectural opposite.
Every claim is filtered through a pathway: preserved verified memory → internal canon retrieval → vetted external libraries (arXiv, PubMed, USPTO, MIT OCW, Project Gutenberg) → peer-model consultation (multiple models, blind to each other) → honest refusal. Fabrication is not an output option. If the pathway produces nothing reliable, the brain says so and routes the question to research threads for incubation.
Five levels per domain: definitions → application → synthesis → cross-domain transfer → generation. Level N+1 stays locked until three different topics at Level N have been verified with cross-axis reflection scores ≥ 3.5/5. The brain cannot skip ahead. Mastery is demonstrated, not asserted.
After answering a probe, the brain must complete a second pass: "how does knowing this help future questions in this domain?" Scored on three independent axes — specificity, forward-looking-ness, and association density. Shallow reflection means rote retrieval, not learning. Specific, forward-looking, association-dense reflection means real understanding that compounds.
Every event in the brain's life is recorded in a cryptographically linked manifest. Cannot be retroactively faked. Cannot be retroactively edited. When a fact gets contradicted, the new one is appended with a supersedes=<old_hash> pointer; the old one stays with superseded_by=<new_hash>. Belief evolution is preserved, never overwritten.
Four different model classes across four hardware platforms produce convergent verified thoughts. When a 27-billion-parameter local model and a 4-billion-parameter local model independently agree on a reflection score above threshold, that's not one model's bias. It's the architecture's signal.
Praefex was built by a single operator-architect using today's AI engineering technology. The architectural design, constitutional principles, methodology, and failure-mode response patterns are human-authored. Code execution is delegated to AI agents working across an 11-node mesh. From concept to functioning brain in roughly four months.
The thesis: AI's currently-default failure mode (confidently wrong) is fixable architecturally. Route through multiple reasoners, demand sources, audit the path, gate on prerequisites, never fabricate. The empirical evidence of Praefex doing this is the proof point — and the live dashboard is the verification surface.