ActiveAudit manuscript drafted · not submitted · checked 13 Nov 2026

Trajectory Geometry

Are the instruments we use to read inside a language model valid before we draw conclusions from them?

An audit of interpretability measurements in a small language model: what survived correction, and what was withdrawn.

Why it matters
Interpretability results are only as good as the measurements under them. A wrong formula or a mismatched tool can produce a confident finding about nothing.
What it does, and how
It started as a study of the path a small open model takes through its internal states while answering. It became an audit: each earlier claim was re-tested with corrected tools, and what failed was withdrawn.
Where it is up to
In October 2026 an audit found two faults in the feature-level analyses, and those causal claims were withdrawn. The manuscript now reports what survived the correction and offers a validation checklist. It is drafted and claim-audited, awaiting my own read before submission. There is no preprint yet.

Where you come in

If you work on interpretability and would read a draft, or want the validation checklist when it is public, write to me.

Write to me about Trajectory Geometry