Research
Specifications, papers, benchmarks and engineering notes. Open licences, real numbers, and the two results that went against us. Those are the ones worth reading first.
Specification
4 September 2026
An open standard for managing brand truth in AI systems. Three layers: Foundation, Governed Context, Automated Production. Foundation is enough for most cases, everything above it is optional.
openbranddefinition.org →Paper
Published, 23 pages
A method for measuring whether a retrieved memory actually earned its place: remove it, re-run the task, measure the damage. Counterfactual ablation instead of similarity scores, LLM judges or rank labels, all of which reason in a circle.
supabrain.space →Benchmark
Open source
42 intentionally broken projects across 30+ languages with encrypted ground truth. Six weighted dimensions: security 35, cross-module bugs 25, logic 20, performance 10, best practice 5, code smells 5. Point any file-reading agent at it and get a comparable score.
trashfire.io →Engineering note
Build in progress
The problem is not longer context. It is deciding what deserves context. Four primitives: persist, index, compile, verify. Every passage carries a file, a line span and a content hash, so an agent can be held to what it read.
supabrain.dev →Product evaluation
Three week evaluation
We ran our own skill catalogue against frontier models for three weeks to find out whether it does what our marketing said. It does not, not in the way we claimed. The catalogue still ships, the claim does not.
supaskills.ai →Next