TLTLL — Useful intelligence per unit of compute

01What TLTLL is

TLTLL is not prompt compression. It is an R&D effort by Answer North to optimize useful intelligence per unit of compute — across models, tools, and providers.

Token counts are a cost metric, not a value metric. The target is the ratio of task-relevant reasoning to total compute spent.

If intelligence becomes cheaper without becoming weaker, entirely new classes of automation become economical. That is the design target — and the reason verification is a first-class layer rather than an afterthought.

02The pipeline

00INPUTRaw task, context, and constraints as they arrive.

01CONTEXT ANALYSISIdentify what the task actually depends on — and what it doesn't.

02SEMANTIC COMPRESSIONReduce low-value context while preserving requirements and recoverability.

03MODEL / TOOL ROUTINGSend work to the best capability-to-cost frontier, across providers.

04EXECUTIONRun the optimized task where it belongs.

05QUALITY VERIFICATIONMeasure whether optimization changed the answer or violated a constraint.

Three layers carry the work: compress (reduce repeated or low-value context while preserving semantic requirements and recoverability), route (send tasks to the model or tool that can satisfy the job at the best capability-to-cost frontier), and verify (measure whether optimization changed the answer, violated constraints, or created hidden quality loss).

03What is claimed, and what is not

TLTLL is a research prototype. Reduction and routing targets are design goals measured internally — no external benchmark results are claimed here. A working paper, "Useful intelligence per token as an optimization target," is in preparation; when it is published it will carry the reproducible measurements this page deliberately does not.

TLTLL's measurement discipline follows the same house rule as ASTRA-Q: a public numerical claim requires a published, reproducible methodology first.

Status

TLTLL is a research prototype under active development at Answer North. No external benchmark results, savings figures, or model comparisons are published yet — by policy, the methodology ships before the numbers.