How I Build

Plausible output is cheap. In medicine it is dangerous. In agent systems it is merely expensive until it is also dangerous. I do not accept a fluent answer that cannot be traced, and I do not accept a green test as a substitute for a system that has been seen to work where it claims to work.

Local completion is not deployment. A push is not a release. A release is not authenticated acceptance. Mixing those states is how a notebook becomes a product in language only.

The loop

Anchor on the actual object of work. Memory is routing context, not proof. Research against current primary sources. Freeze the claim before you plan. Implement without silently swapping the question. Check the result against the research, not the vibe of the output. Accept independently. Then close, or stop. Two approaches that do not move the work are a halt, not a reason to guess a third time.

Three places the method shows

Quarantine. In the source-locked pipeline, when OCR, visible source structure, and extracted fields disagree, the record is quarantined instead of completed. A complete-looking package is not a virtue if the companion item is not on the page.

Boundary. In persistent memory, medical learning is separated by architecture and restricted tools, not by a tag in one undifferentiated store.

Cutover. In XOT, a new publishing candidate stays fenced from historical backlog during a forward-only transition. A healthy preview is not production acceptance.

What the product owes a person

Dense and operable. Few duplicated controls. Empty and error states as part of the design. No marketplace theatre while the thing is still local. Source fidelity first. Uncertainty in the interface, not only in a footnote.

The point is not process theatre. It is to make the claim of “done” smaller, clearer, and falsifiable.

Policies for what this site will claim live on Claims and boundaries.