Skip to content
BasinWright
Talk to an ArchitectStart Building
Engineering notes

Why the platform works the way it does

The mechanisms behind the design decisions, and what we would have to measure to know whether each one holds.

Decisioning

Why we require an evidence quorum rather than a confidence score

The design bet: independent lanes fail in independent ways, so requiring agreement collapses the confident-error class at the cost of a few more escalations.

8 min readLoom
Monitoring

Watching the corpus instead of waiting for accuracy to move

By the time an accuracy metric moves enough to alert, the model has been making worse decisions for as long as it took the signal to accumulate. Sometimes a quarter.

7 min readCartograph
Models

When a small tuned model is the right answer

The largest available model is the default and it is usually the wrong one — not because it performs badly, but because it is doing work well within its capability at many times the cost.

8 min readLoom
Retrieval

Why entity scoping is the retrieval decision that matters

The wrong customer's records are genuinely, correctly similar. A better retriever finds them more reliably — which is the opposite of what you wanted.

7 min readMeridian
Platform

Cost-aware routing, and the failure mode it introduces

Under-escalation does not distribute randomly. It concentrates exactly where the small model is confidently wrong — a systematic population, not noise.

7 min readKeystone