TRA: Persistent Post-Deployment Learning in a Quantized Transformer

Preprint identity

This page hosts the DA Synergie TRA research preprint. It is the same paper we intend to submit later to arXiv (same title, authorship, abstract, and PDF), not a separate website publication.

• Status: Preprint, Version 1.0 — October 2026
• Not peer-reviewed (unless a later revision states otherwise)
• Research instrument in the paper: Granite 3.3 2B Instruct Q4_K_M
• PDF: Download preprint PDF (v1.0)

Authors

• Dany Bouchard — Founder, System Architect & Principal Investigator
• Atlas — Research Guide & Structure Manager (OpenAI AI system used throughout the research process)

Independent TRA research. Dany Bouchard retains responsibility for the research claims and final publication.

Abstract

Large language models are trained from finite historical corpora and do not automatically incorporate later events into frozen pretrained weights. Full-model retraining or fine-tuning can update a model, but repeatedly modifying an entire model to incorporate a small number of new facts is computationally and operationally disproportionate to the size of the update. TRA (Transformer Relearn Architecture) investigates whether a deployed quantized transformer can instead acquire persistent post-deployment experience while its original GGUF remains immutable, learned changes remain reversible and versioned, and inference remains available. Experiments use Granite 3.3 2B Instruct Q4_K_M and real post-horizon name-transition events, principally Nauru to Republic of Naoero. TRA progressed from ineffective bounded FFN_DOWN deformation to distributed rank-8 LoRA attached to Q/K/V/O projections across all 40 layers. The experiments measured acquisition, update-magnitude sensitivity, interference, semantic preservation, related-experience transfer, forgetting, generational freezing, Wake/Dream consolidation, persistence, and deterministic runtime time injection. The strongest historical G1-9 heuristic score was B=3/6 with A=4/5 after one pass, but raw outputs show partial lexical acquisition rather than clean semantic understanding; this paper therefore distinguishes deterministic surface-token scores from qualitative semantic correctness. Related learning transiently improved Naoero from 3/6 to 4/6 without direct retraining, followed by degradation to 1/6 in a single plastic bank. Freezing a useful generation and allocating fresh capacity improved the final Naoero trajectory to 4/6 despite temporary cross-bank interference. A deliberately bounded one-pass Wake phase limits deformation from a single observation, while Dream replays already authorized evidence for later consolidation; G1-19 improved a Wake result from 2/6 to 3/6 in two Dream iterations. Production LRMD state is versioned, non-destructive, and rollback-capable. TIJ demonstrates that deterministic changing state such as current time should be supplied at runtime rather than permanently learned. TRA does not establish general lifelong learning; it provides a transparent systems architecture and a documented set of observations on one quantized transformer.

Copyright and licensing

© 2026 Dany Bouchard / DA Synergie. All rights reserved unless a later notice states otherwise.

Hosting this PDF on dasynergie.com is a preprint distribution of our own work. It is not an arXiv record yet. When the paper is submitted to arXiv, we will select an arXiv distribution/license option at submission time (arXiv supports several choices, including Creative Commons options). Until that submission and license choice are recorded here, do not assume a Creative Commons grant from this website posting alone.

Please cite this preprint by title, authors, “Preprint, Version 1.0 — October 2026,” and this URL (or the PDF URL). After an arXiv identifier exists, prefer that identifier.

Website companion summary (not a second paper)

The text below is a short product-facing explanation for general readers. It is not a separate publication and does not replace the preprint. Product names such as “Trainable Runtime Architecture” describe the broader DA Synergie stack; the paper’s acronym expansion in the preprint abstract remains Transformer Relearn Architecture.

What the preprint is about

Most local AI setups freeze a model and then pretend the world stopped changing. TRA asks whether a deployed quantized model can acquire small, verified updates while the original GGUF stays immutable, learned changes stay reversible and versioned, and inference remains available. The preprint documents that path carefully on Granite — including acquisition messiness, transfer/forgetting, Wake/Dream consolidation, and why continual learning cannot be judged by a single score.

What TRA is (and is not) in the product line

• Immutable base GGUF in ordinary TRA learning — learned state is versioned outside a destructive rewrite of the base file.
• Not Middleware Pro — Middleware Pro connects IDEs to local models; it does not by itself add /learn or /sleep to every GGUF.
• Not solved lifelong learning — the preprint is explicit about that limit.

Exact persistence wording for the public packages: TRA learning — technical note.

Why the public demo stack moved to Ministral 14B Vision

Granite was the research instrument in the preprint. The public experimental packages (llama.cpp TRA + Ministral 14B TRA) use Ministral because it is a stronger coding host and follows multi-step plans more reliably in EXPCoder-style work, with vision for engineering context. Granite plastic/canvas state does not transfer to Ministral; the architecture and lessons transfer, not the adapter snapshots. That product decision is outside the September Granite experimental record in the preprint PDF.

Related downloads (product packaging)