Back to projects

Active project

Ten-SON-LM

A recurrent language architecture with a fixed-size semantic workspace where learned tension controls revision, sparse routing, relaxation, and readout.

Claim discipline: Claims on this page should be treated as hypotheses unless linked to a receipt, benchmark, or reproducible artifact.

What problem it addresses

Calling an internal gate tension means little unless controlled interventions show that it contributes something beyond ordinary learned gating.

How it works

Route token-derived proposals into a fixed workspace, update selected slots through learned tension, relax state recurrently, and read output from workspace context.

What exists now

A working scientific substrate with promising but incomplete causal evidence.

Evidence / receipts

Milestone 1 learned delayed recall, balanced brackets, and a synthetic next-token task. Copy failed its threshold; tension evidence was strong on delayed recall but mixed elsewhere.

Limits

The validation result is Partial. Slot-specific tension is not yet proven, and the architecture has not established an advantage over matched recurrent or learned-gate baselines.

Next milestone

Milestone 1.1: explain the copy failure and run pre-registered causal interventions against equal-sized GRU/LSTM/workspace learned-gate baselines.

How to run / inspect

Install / open

git clone https://github.com/BoggersTheFish/Ten-SON-LM && cd Ten-SON-LM

Inspect

Read the Milestone 1 report, inspect task-level receipts, and compare learned, shuffled, frozen, constant, and inverted tension interventions.

Links

Focus tags

Workspace revisionSparse routingCausal interventions

Related receipts