← Back

Method

Capture, curate, train — then delete. One local loop; everything happens on your machine and nothing leaves it.

Four rules

1. Every number ships with its conditions. Hardware, dataset size, date. If it cannot be reproduced from the notes, it does not go up.
2. Failures are published next to the result. A clean number with a known exception is worth more than a clean number alone.
3. A test that cannot fail is not a test. Every check has a stated way it could be wrong, and we write that down before we run it.
4. Nothing is claimed that was not measured. No “first”, no “best”, no “frontier”. If we have not run it, we say we have not run it.

Three questions we test

Question	What would make us wrong
Does it still recall a fact after we delete it?	A paraphrase we did not test survives
Does it make a connection it should not?	A path below the layer we observe: kernel, driver, firmware
Is it better than no retrieval at all?	A task suite we wrote to flatter ourselves

What stays private

The measurements are public. The mechanism is not — not until a provisional is filed. If you need the mechanism to judge the work, that is a fair thing to ask for, and it is a conversation under NDA rather than a paragraph on this page.