Roadmap

Public, and revised as the project grows. Items marked help wanted are where a contributor would make the most difference. Discuss any of them in Discussions.

Shipped

• Kenning v0.4: an open System One model (Apache-2.0) with published benchmarks
• SystemOne Builder: serve, train, distil from Clef, benchmark and publish models, on one GPU
• The systemone Python client: one client for Kenning, Clef and Jev
• This site: concepts, cookbook, projects, all in Python and runnable

Next

Kenning that holds its own on subtle phishing. Today it’s at 0.75 on our held-out modern-email set, against 0.90 for the big engines. Next: more realistic calm-lure and legitimate-twin training pairs, and a permissively licensed real phishing corpus. Help wanted: labelled emails you can share under an open licence.

More benchmark suites. Every suite is a small, hand-labelled, openly licensed set that’s never trained on: support tickets, moderation, intent routing, log triage. Help wanted: this is the easiest way to make a real difference.

A calibration toolkit. The calibration check as a small library: reliability diagrams, Brier and ECE, drift alerts. Help wanted.

A TypeScript client. The wire format is simple; the client should exist for Node and the browser too. Help wanted.

Smaller and faster Kenning. A base-size model for CPU-only and edge deployments.

Later

• Evaluation guide: building a labelled set, sampling, and measuring without fooling yourself
• Cost modelling: System One routing versus an LLM for everything, with honest assumptions
• Domain sections: fintech, healthcare intake, logistics, ad review
• Case studies: real deployments, named or anonymous, with real numbers
• Code review checks: whether System One models can usefully read code diffs is an open question; we’d like a benchmark before we publish a pattern
• Glossary: calibration, proper scoring rules, distillation, and the rest

Explicitly not doing

Becoming a sales site. Not for Kenning, not for anyone. The moment this reads as marketing it stops being useful.

A single-number leaderboard. Ranking models on one number invites gaming and misleads readers whose workload looks nothing like the benchmark. We publish per-suite results with methodology instead.

Training on data we can’t share the licence of. Kenning’s training data is permissively licensed or generated, and listed in every model card. No outputs of services whose terms forbid it.

Influencing this

Start a discussion or open an issue. Things people actually ask for get priority over things we assumed they wanted, so tell us what you came looking for and didn’t find.