Trunk and Branches: Thermodynamic Parsimony, Non-Destructive Cortical Specialization, and Continual 3D/C++20 Synthesis on a Frozen 4B Foundation Model (Isaque 2.0)

Technical monograph and longitudinal experimental audit presenting Isaque 2.0 ('Trunk and Branches'), a dual-tier continual learning and neuro-symbolic execution architecture evaluated on consumer hardware (~22.3-23.3 W active inference on an NVIDIA GeForce RTX 5060 8 GB, 0 MB idle VRAM). Pairs a 100% frozen Qwen3.5-4B backbone with 16.00 MiB orthogonal DoRA + O-LoRA cortical cartridges, Sparsegen LD-MoLE Top-2 routing over a 0.015 ms zero-copy .mmap tree, deterministic physical reality compilers (headless Blender 5.2.1 LTS bmesh and clang++ C++20), and a 4-stage Windows NTAPI zero-swap streaming pipeline (+0.0039 MiB swap delta) for a blind Qwen3.5-35B-A3B critic. Includes a 10-wave continual saturation and rupture matrix (100.00% retention via LD-MoLE Top-2 vs. 75.31% token accuracy / 31.37% likelihood retention in a monolithic adapter), the Staged Modular DAG Composition Theorem, SHA-256 hardware receipts, and 4-calendar Bitcoin blockchain OpenTimestamps provenance (Block #975458). Version 5.1.0 (Forensic Evidence & Bilingual Edition): Adds the complete 18-page Brazilian Portuguese twin monograph, compiled English & Portuguese Forensic Evidence Annexes, explicit reconciliation of pre-training anchor H0 (82.85%) vs. learned-wave minimum H1..Hk (89.91% at k=4), and the unaltered raw empirical archive (isaque_2_0_forensic_evidence_and_receipts_doi_24221509.zip: 195 triangular retention cells across 3 regimes, 10/10 solo Isaque-generated 3D/C++20 wave outputs with deterministic recheck exports, and authentic OOD/token-collision failure traces). Version 5.2.0 (Pure Solo-Model Forensic Edition): Line-by-line forensic purification retaining strictly unaltered primary machine receipts (ato2_escada_ondas_3_a_10.json with 195 triangular retention cells across 3 regimes, ato1_desafio04_cego.json, ato1_promocao_e_precheck_ondas3_10.json, and Stage-1 OpenTimestamps Bitcoin Block #975458 receipts) and 100% unedited solo Isaque 2.0 code outputs (onda_3..onda_10 with bit-identical exports == exports_recheck, DESAFIO-01..04, and FUSAO_2_ONDAS_1_3 collision traces), completely purging all auxiliary live-showcase wrapper files.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24221508
Primary Topic
Domain Adaptation and Few-Shot Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Trunk and Branches: Thermodynamic Parsimony, Non-Destructive Cortical Specialization, and Continual 3D/C++20 Synthesis on a Frozen 4B Foundation Model (Isaque 2.0)

Lucas Alves Motta
Zenodo (CERN European Organization for Nuclear Research)
Domain Adaptation and Few-Shot Learning
preprint

Trunk and Branches: Thermodynamic Parsimony, Non-Destructive Cortical Specialization, and Continual 3D/C++20 Synthesis on a Frozen 4B Foundation Model (Isaque 2.0)

Lucas Alves Motta
preprint en

Abstract

Technical monograph and longitudinal experimental audit presenting Isaque 2.0 ('Trunk and Branches'), a dual-tier continual learning and neuro-symbolic execution architecture evaluated on consumer hardware (~22.3-23.3 W active inference on an NVIDIA GeForce RTX 5060 8 GB, 0 MB idle VRAM). Pairs a 100% frozen Qwen3.5-4B backbone with 16.00 MiB orthogonal DoRA + O-LoRA cortical cartridges, Sparsegen LD-MoLE Top-2 routing over a 0.015 ms zero-copy .mmap tree, deterministic physical reality compilers (headless Blender 5.2.1 LTS bmesh and clang++ C++20), and a 4-stage Windows NTAPI zero-swap streaming pipeline (+0.0039 MiB swap delta) for a blind Qwen3.5-35B-A3B critic. Includes a 10-wave continual saturation and rupture matrix (100.00% retention via LD-MoLE Top-2 vs. 75.31% token accuracy / 31.37% likelihood retention in a monolithic adapter), the Staged Modular DAG Composition Theorem, SHA-256 hardware receipts, and 4-calendar Bitcoin blockchain OpenTimestamps provenance (Block #975458). Version 5.1.0 (Forensic Evidence & Bilingual Edition): Adds the complete 18-page Brazilian Portuguese twin monograph, compiled English & Portuguese Forensic Evidence Annexes, explicit reconciliation of pre-training anchor H0 (82.85%) vs. learned-wave minimum H1..Hk (89.91% at k=4), and the unaltered raw empirical archive (isaque_2_0_forensic_evidence_and_receipts_doi_24221509.zip: 195 triangular retention cells across 3 regimes, 10/10 solo Isaque-generated 3D/C++20 wave outputs with deterministic recheck exports, and authentic OOD/token-collision failure traces). Version 5.2.0 (Pure Solo-Model Forensic Edition): Line-by-line forensic purification retaining strictly unaltered primary machine receipts (ato2_escada_ondas_3_a_10.json with 195 triangular retention cells across 3 regimes, ato1_desafio04_cego.json, ato1_promocao_e_precheck_ondas3_10.json, and Stage-1 OpenTimestamps Bitcoin Block #975458 receipts) and 100% unedited solo Isaque 2.0 code outputs (onda_3..onda_10 with bit-identical exports == exports_recheck, DESAFIO-01..04, and FUSAO_2_ONDAS_1_3 collision traces), completely purging all auxiliary live-showcase wrapper files.

Zenodo (CERN European Organization for Nuclear Research)
Domain Adaptation and Few-Shot Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon