Long Tail Inference Lab
An open research laboratory for turning verified work into local intelligence.
IThesis
A lightweight local model may not need frontier scale for every recurring problem. It may need compact approved evidence from stronger inference on separate verified work.
The preregistered protocol transfers sanitized Markdown from a fixed cloud teacher to one fixed local Qwen student. The student stays fixed. Approved memory grows. Held-out executable tests decide whether transfer actually occurs.
IIWhy it matters
Useful intelligence should remain close to people, their devices, and the evidence behind its answers.
The experiment treats answers as claims to verify, not prose to admire. Success requires executable evidence, privacy safe artifacts, visible failures, and honest abstention when local knowledge is insufficient.
IIICurrent experiment
IVMethod
Build with a cloud teacher. Sanitize locally. Approve exact hashes. Test the same local Qwen student as memory grows.
- 01 Teach Let the fixed cloud teacher solve only preregistered public memory-build tasks.
- 02 Verify Require a qualified task verifier and an executable pass; model claims never substitute.
- 03 Sanitize & approve Retain the cloud teacher's raw capture locally, never retransmit it for distillation, and approve the allowlisted sanitized evidence and draft by exact hash.
- 04 Measure Compare held-out local Qwen M0/M2 executable-verifier pairs; teacher outcomes are provenance, not scores.
The laboratory is open
Read the protocol. Reproduce the work. Challenge the result.
Enter Long Tail Inference Lab on GitHub