TensionLM

What problem it addresses

Softmax attention is powerful but often opaque at the pairwise pressure level. TensionLM tests whether sigmoid tension fields can expose useful telemetry.

How it works

Replace or compare attention surfaces with sigmoid tension attention, export pairwise fields, and hold claims to matched softmax-vs-sigmoid receipts.

What exists now

Public repo and model ladder exist; matched benchmark receipts need tightening.

Evidence / receipts

Published model ladder on Hugging Face plus proof-bank entries around seeded language traces and telemetry. Claims remain benchmark-bound.

Limits

Not a claim that sigmoid tension beats standard attention generally. Broad capability claims require matched benchmark receipts.

Next milestone

Matched softmax-vs-sigmoid benchmark receipts.