A single, honest snapshot of where the program stands. Capable AI you can own — on commodity hardware — through self-similar architecture and memory that organizes itself.
6citable outputs
6reproducible
5research threads
10articles
2newsletter issues
2open repos
Éthiqueia is a Research Compiler: it turns an idea into a public, reproducible, citable artifact with minimum friction. Every proof drop runs the same pipeline, so each result becomes a permanent scientific asset the next one builds on.
1Hypothesisresearch.ts "next proof needed" per threadpartial
On context-window overflow, Ollama silently discards the front of the prompt and returns 200 OK: given a 160,689-token prompt against a 32k window it evaluated 16,387 tokens, reported nothing, and the model confidently invented an answer to a question whose ground truth sat in the discarded head. llama.cpp instead returns a hard 400 exceed_context_size_error, which is honest but fatal to an agent turn. ContextPaw is an MIT-licensed proxy speaking both the Ollama and OpenAI APIs that makes overflow honest: it evicts middle-out (preserving the prompt-cache prefix and the system prompt that head-truncation destroys), reports every eviction in the body, the headers, and inline to the model, and recovers llama.cpp 400s by compacting and retrying. We further show that notifying the model of the eviction is insufficient: with three facts buried in the evicted region, a marker stating in plain language that the information was removed still yielded 0/3 recovered facts and three confabulated ones, while summarizing the evicted span with a 1B local model yielded 3/3. Telling an agent it lost information does not stop it hallucinating; the content must be returned.
Memory SystemsbenchmarkedThree reproducible proof drops (2026-06-27). #1 (UFM): runs a 24 GB expert bank on a 23.5 GB RTX 4090 where baseline OOMs; within ~1% of baseline throughput when the working set fits; ~240× faster than naive offload. #2 (FMM): topic-scoped retrieval ~164× faster + ~3.7× more accurate than flat at 128k when the topic is known; recall 0.0 if misrouted. #3 (FMM router): a near-free centroid router (~0.002 ms/query) recovers 98% of oracle recall at ~60× flat-scan speed when topics are separable, but realized recall tracks routing accuracy and collapses toward chance as topics overlap. All three converge on one finding: locality pays only when you can address the right region.A learned topic router that pushes the routing-accuracy crossover into harder (overlapping) regimes where the cheap centroid router fails — measured end-to-end against the centroid floor (proof drop #3). Plus: real-embedding retrieval corpora; training-time paging (autograd-safe) + OffloadedAdam memory curves.
Commodity TrainingbenchmarkedMeasured training envelope (VRAM, throughput, tokens/day) for the full 30M-500M family on one RTX 4090, re-verified 2026-06-29, plus a real 120M loss curve (9.65 → 2.78 over 9.04B tokens). One-command reproduce.Model-quality numbers from a full run — held-out perplexity and HellaSwag/ARC via `forge eval` — to put a quality axis next to the cost axis.
Agent Runtime SystemsshippedCode others can run; agents whose identity + memory are plain files. The YSON spec and an inspectable-memory demo are not yet written up.A published YSON spec; a reproducible demo of inspectable, file-based agent memory compared against opaque memory stacks.
Cognitive ArchitecturesprototypeRunning systems, documented in public. AEON is open-source with a technical whitepaper and a devlog (e.g. tracing a liquid-network trainability bug to a missing normalization and fixing it). Internal/emergent behaviors are not yet formally measured, so all capability claims stay marked speculative — this thread is framed as engineering, not results.Measurable, reproducible behaviors: memory growth over time, belief-revision traces, idle-cycle throughput, and long-run determinism/drift soak tests — before any public capability claim.
Fractal CognitionprototypeRich and documented, but NOT yet benchmarked. All architectural-advantage claims are marked speculative until measured.A controlled, small-scale comparison of the fractal backbone vs. a parameter-matched transformer (loss/perplexity at equal params + equal compute).