Tesseract
Diffusion reasoning — not the same system as Tessera
Last devlog: August 25, 2026
Key boundary: The model adapter line. Everything above it speaks an abstract protocol. Which model serves behind it, hosted or self-run, is configuration. Not Tessera — different system, similar name.
Why it matters: Substrate quirks stay in one place. Swapping models costs configuration, not code. Findings on one deployment stay labeled until confirmed elsewhere.
Notable decision: Refinement names the failed check but never the expected answer, so recovery cannot be copying. Measured: a clock-arithmetic task the model had failed, refuted twice, recovered on pass three, verified by recomputing its own working.
Tesseract is not Tessera. Tessera is the kernel under Aurelia. Tesseract wraps planning, memory, sequencing, and verification around a diffusion language model served as ordinary text generation.
A reasoning task the model got wrong twice was corrected on the third attempt after the system said only which mechanical check had failed — never the expected answer — and the corrected output was verified by recomputing the arithmetic inside it.
On an eight-task machine-verifiable suite, raw model calls passed 4 of 8; the pipeline passed 8 of 8. Every prevented failure was structural (empty outputs, truncation), stated as reliability, not as intelligence.
What it still cannot: the hosted sampler is fixed, so full harness tuning waits on self-hosting that is not scheduled. Verification is narrow — lexical match plus arithmetic recompute — so a self-consistent wrong derivation passes. If you want a magic reasoning engine, this is not that page.