intuition labs · lens-gate← the foundry · j-town
Drop a lens. Get a verdict, with a receipt.
Drop one of Neuronpedia's pre-fitted Jacobian-lens files
(huggingface.co/neuronpedia/jacobian-lens,
the raw *_jacobian_lens.pt) — or any plain-JSON lens
{d_model, vectors: number[][]}. This page fits held-out lens directions into
sparse nonnegative cones, compares two layers with Δ_μ, gates each fit at a disclosed
tolerance, and hands you the receipt.
Everything runs in this tab. No request leaves it — check the network panel. Same file, same numbers, every time: the pipeline has no randomness anywhere.
—Δ_μ between the two cones (k=3)
—probes commit at tol 0.85
—probes escalate
What this is not: this page does not run a model, does not fit lenses, and cannot tell you
what a lens direction means. It stores, fits, compares, and gates the objects other
tools produce. A high escalation rate is the honest reading — a k=3 nonnegative cone of 48
lens rows explains only a small slice of a held-out direction, and the gate says so instead
of pretending. The tolerance (0.85 relative residual) is page-local and printed on the
receipt it gates.