Research/MathesisPhase 2 release

Mathesis: Regression-Gated Continual Personalisation of Local Language Models

A regression-gated continual personalisation pipeline for local language models using isolated LoRA candidates.

01 - Overview

Overview

Mathesis studies how approved corrections can become isolated LoRA candidates for a local language model without mutating the parent model.

Mathesis completed a checksum-verified three-seed controlled study on Qwen3-1.7B. Two independently trained LoRA adapters passed all eight predeclared activation gates, while one high-accuracy candidate was rejected because it introduced a contradiction.

02 - Activation gates

Activation gates

A candidate is accepted only when every predeclared gate passes: target learning, prior-lesson retention, general capability, contradiction, scope error, safety, provenance, and tensor integrity.

The seed-17 result is the key counterexample: aggregate accuracy alone would have accepted it, but the contradiction gate correctly forced rollback.

03 - Evidence boundary

Evidence boundary

This is a successful controlled proof-of-concept, not a claim of universal personalisation.

The study uses generated template-related English examples, one pinned 1.7B base model, three seeds, and no human participants.

04 - Findings

What the release demonstrates.

  • Unmodified base model scored 29/60 on the frozen final suite; the static all-lessons prompt scored 51/60.
  • LoRA seeds 42 and 73 reached 60/60 and passed every activation gate.
  • LoRA seed 17 reached 59/60 but was rejected because the contradiction gate caught one reversal.
  • The accepted seed-42 demonstration adapter corrected 31 base failures with 0 base successes lost.
  • The compact general sentinel stayed unchanged at 5/8 for the accepted seed-42 adapter.
  • The exact paired McNemar result against base for seed 42 was p = 9.31 x 10^-10 on the fixed controlled suite.

05 - Prototype

Selected system details.

MethodIsolated LoRA candidate training with rollback
Base modelQwen/Qwen3-1.7B pinned to a fixed commit
Final suite60 generated controlled examples
DomainsWriting, coding, and design
ActivationEight predeclared gates; no weighted score can hide a failed check
Evidence boundaryControlled proof-of-concept, not universal personalisation

Reported statistics describe the fixed controlled suite and should not be read as population-level estimates or general safety certification.