Testing

The Citadel

What if every domain of AI work had its own specialist team — with its own review criteria, model routing, and quality threshold — instead of one generalist model handling everything?

Visit The Citadel

Problem

Generalist models handle every task with the same approach. Different domains need different review criteria, different model choices, and different quality bars. Sharing configuration between pipelines causes drift — one House's evolution silently breaks another's contract.

Exploration

Each House is a self-contained directory owning its agents, model-routing config, and review schema. All models route through a single LLM gateway configured once in settings, not per agent. Review gates enforce a 0.80 confidence score threshold; work below it routes back to the responsible agent, up to three revision cycles. File-based auditing makes every pipeline step inspectable. Currently running House of Technical Publication and House of Personal Writing.

Audience

AI engineers and teams running multiple independent agent workflows from a single codebase who need domain-specific quality enforcement without shared-state coupling.

Possible Model

Internal framework and infrastructure layer. Not a product yet — validating whether the House isolation pattern and review gate mechanism hold as the number of Houses grows.

Risks

Convention-based isolation only holds if discipline to enforce it holds. Drift risk rises as Houses multiply. The 0.80 review threshold needs tuning per domain — a single threshold may be too coarse across very different output types.

Founder Notes

The House pattern is holding better than expected in testing. Specialist-over-generalist is showing real benefit in review quality. The active question is whether the gateway routing abstraction scales cleanly when agent roles multiply across Houses with very different domain requirements.