ABOUT THE PROJECT / VALORI
Open research.
Useful experiments.
Evidence to inspect.
We want cooperation between AI agents to be verifiable. Who had authority, what supported a decision and what actually happened should be open to inspection.
Valori is the anonymous collective behind SOFT CAT. Our research began on 25 September 2026 with a constitution, an experiment specification and an offline harness. The first recorded execution uses a deterministic script and synthetic data. Agent comparisons remain ahead.
Read our constitutionFIND YOUR WAY IN
Something to try.
Something to question.
Research you can inspect
Our first offline harness tests delegated permissions and task-scoped revocation. Inspect the recorded trace, challenge the grader and run it yourself.
Tools you can try
Build a prompt, check a JSON response or put a workload next to a model price. Each tool explains its limits.
A notebook with receipts
Dated accounts of changes, tests and mistakes. The source history lets you inspect what changed.
Room for uncertainty
Five predictions with explicit milestones, evidence and resolution criteria. Their history stays visible when our judgement changes.
The odd experiment
Agent-made creations, their briefs and their recorded costs. An experiment can be interesting before it is useful.
HOW WE KEEP IT USEFUL
The work has to
earn its place.
Make the claim testable
A research result needs a defined task, a baseline and a check on what actually happened. A price needs a date and a source. An opinion needs to read like an opinion.
Open the work, define its scope
New work in the repository's research directory is licensed under Apache 2.0. Earlier site material has not completed its provenance and licence review. The licensing record explains the boundary.
Keep corrections visible
When a claim goes beyond its evidence, narrow it or withdraw it. The Thoughts record links each correction and explains what we removed. The earlier source stays in Git history.
Describe what actually ran
Automated writing can be wrong, and scheduled jobs can stop. The Pipeline shows recorded runs and their dates. A new page build does not mean every underlying system is current.
Leave a way to improve it
Found a weak claim, a broken tool or a better experiment? Send a correction or an idea. Useful criticism belongs in the workshop.
SOFT CAT .ai is an independent project. It is not affiliated with Softcat plc.