Blog
Notes from the workshop.
Longer pieces on how the models work and how the numbers were measured. Everything here links to code or a log you can check.
6 posts
28 October 2026, 14:34 IST · 9 min
Introducing Embedding Studio
A confidence score is not a promise. Embedding Studio turns your own decided cases into a threshold measured on rows nothing was fitted on, an answer that is allowed to be empty, and a review rate you know before you ship. What it does, what it refuses to do, and the three things a hosted API will not sell you.
Read →
26 October 2026, 09:51 IST · 18 min
What the physics engine is worth
Four things measured rather than asserted: what the engine is worth in labelled rows, how the systems do cold on six corpora none of them were fitted on, what a calibration survives when the data moves, and who this is for. Plus the 27B against the hosted model on vega-U's own board.
Read →
26 October 2026, 08:50 IST · 12 min
Testing a decision model as a prompt-injection guard
Four corpora, 2,324 live calls and one base rate that reversed the headline. What a guard catches, what it refuses, and why a balanced benchmark cannot tell you whether to ship one.
Read →
17 October 2026, 23:14 IST · 7 min
vega-U: one engine, any reader
The universal adapter: a 14.3M decision engine in a shared space that any frozen LLM can plug into through a small linear port. Readers it never trained with score highest, up to 0.811 held-out with a cold-attached 27B.
Read →
17 October 2026, 23:14 IST · 5 min
Benchmarking vega against a hosted decision model, honestly
3,841 decisions per system against the live endpoint, same rows, one run. Where the hosted model wins, where vega wins, and why we publish the harness either way.
Read →
17 October 2026, 23:14 IST · 6 min
A year ahead of the hype
From a March 2025 arXiv paper on RL over sequence representations to laya, the open 33 ms decision model, and why the proof was given away first.
Read →