Experimental project
TSQ
Tension-Structured Quantization: a verifier-gated runtime that keeps ordinary inference cheap and escalates precision or repair when tension justifies the cost.
What problem it addresses
Static quantization spends the same numerical precision regardless of whether a token or reasoning step is routine, unstable, or verifier-critical.
How it works
Estimate cheap tension signals, route through a low-precision path by default, escalate on uncertainty or verifier failure, and record both cognitive and compute receipts.
What exists now
A serious runtime and experiment harness, but not yet proof of adaptive-quantization efficiency.
Evidence / receipts
A runnable routing loop, mock and Transformers backends, verifier repair, evaluation/cost accounting, data preparation, LoRA scripts, and an adapter experiment harness exist.
Limits
No custom quantization kernels or committed model-quality result exist yet. Precision labels are routing labels unless mapped to genuinely distinct numerical models.
Next milestone
Implement real low/high precision execution and measure wall time, peak memory, bytes moved, verifier pass rate, repairs, and output quality.
How to run / inspect
Install / open
git clone https://github.com/BoggersTheFish/tsq && cd tsqInspect
Compare always-low, always-high, and dynamic routes only after the backends execute materially different precision paths.
Links
Focus tags
Related receipts