AI handles language. Code handles judgement. Humans make decisions.
Koher is a philosophy for building AI tools that respect domain expertise rather than replace it.
Every tool here separates language from judgement:
| Layer | What It Does |
|---|---|
| Qualification | AI reads language patterns, transforms unstructured input into structured signals |
| Rules | Code handles judgement — deterministic logic produces auditable decisions |
| Language | AI narrates decisions already made |
When you ask AI to "judge whether this is good," you lose auditability. When you separate the layers, you gain it back. The split has a name — Split-Domain Cognition (SDC) — and a public DOI: 10.5281/zenodo.19661700.
Small tools demonstrating the architecture. Each grounded in student conversations before development. Open source — clone the repo, bring your own API key, run locally. Hosted demos run 10 free analyses per day per browser, with no account and no email.
| Tool | What It Does | Released | Licence |
|---|---|---|---|
| Sensorium — Sycophancy | Map how a language model handles five sycophancy triggers — desktop app for macOS and Linux | May 2026 | AGPL-3.0 |
| Fragment Mapper | See structural relationships between scattered text fragments | March 2026 | MIT |
| Play Shape Diagnostic | Analyse experiential qualities of play concepts | February 2026 | MIT |
| Coherence Diagnostic | Analyse design concepts for coherence | February 2026 | MIT |
Tools released up to 5 April 2026 are MIT. Every tool published after is AGPL-3.0.
┌─────────────────────────────────────────────────────────────┐
│ Q · Qualification │
│ Transform unstructured input into structured signals │
│ Principle: AI reads language patterns │
└─────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────┐
│ R · Rules │
│ Apply thresholds, evaluate relationships, produce verdicts │
│ Principle: Code handles judgement │
└─────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────┐
│ L · Language │
│ Translate rule outputs into readable explanation │
│ Principle: AI narrates decisions already made │
└─────────────────────────────────────────────────────────────┘
- Website: koher.app
- Architecture paper (Zenodo):
10.5281/zenodo.19661700 - SDC site: splitdomaincognition.org
- Contact: hello@koher.app
Co-created by Prayas Abhinav + Claude Code