Purpose
ARC-Neuron-LLMBuilder is a governed local AI building and memory system whose core goal is to provide an auditable, rollback-capable, evidence-driven lifecycle management for the training and improvement of small language models. It is not just a training repository, but an evidence-preserving build loop for developing better local AI systems. The system treats a language model as an artifact within a governed lifecycle: you train a candidate model, measure it, compare it with the current production model, and promote it only if there is no regression on protected capabilities. Every decision leaves credentials, every candidate model can be restored, and every archive is linked back to the source of truth through an indexed binary ledger.
Main Capabilities
- Train small models: built-in working Transformer families (ARC-Neuron Tiny and Small), as well as a retrieval-based example adapter, a canonical dialogue pipeline, a draft -> critique -> revision reflection process, and automatic absorption of terminology from dialogue.
- Benchmarking and scoring: provides a 168-task scorer-extended benchmark inventory covering dimensions such as reasoning, planning, critique, repair, continuity, reflection, instruction following, calibration, compression, out-of-domain behavior, paraphrase stability, quantization retention, archive/runtime/state evidence, deterministic compliance, and memory continuity.
- Gated promotion: uses Gate v2 regression-aware gating, promoting to production only when the candidate model truly improves and does not regress on protected capabilities.
- Evidence and archiving: all decisions generate benchmark credentials, preserving rollback-capable lineage; archives include SHA-256 hashes and an indexed binary ledger.
- Independent audit and verification: provides a 136-item public verification test suite, a 9-step proof workflow, and an independent audit report (the v10 audit corrected 4 structural defects, and after 4 governed promotions, the score increased from 0.6836 to 0.9237, a gain of 35.1%).
- Live deployment and continuous learning: the ARC GitHub AI Operator processes code review issues daily through Cloudflare Workers AI, writing each production review as a supervised training example into the repository seed example schema; a nightly workflow ingests these artifacts and deduplicates them, increasing the confidence of manually corrected records.
- Portable memory layer: Omnibinary Runtime and Arc-RAR provide device-portable communication, replay, recovery, and source spine connectivityContinuity.
- Visual/Video Memory Plugin: ARC-StreamMemory is a visual/video memory plugin actively being built, providing visual memory capabilities for ARC-style systems and other LLM stacks.
How to Use
The project provides one-click operation commands; specific installation and running steps require referring to the documentation in the repository. Operations include training candidate models, running benchmarks, gating via Gate v2, archiving evidence, and rolling back. Live deployment requires configuring the OPERATOR_READ_TOKEN secret (a PAT with Actions: Read permission for GareBear99/gh-ai-operator). The currently reproducible production model is arc_governed_v10_wave4, scoring 0.9237 on the audited v10 benchmark. v11.3 / wave5 is only in candidate/staging status until promotion evidence cleanly reproduces through Gate v2.
Use Cases
Suitable for researchers, developers, and teams who want to improve models without relying on GPU servers, cloud-only APIs, or unverifiable black-box workflows. Particularly suitable for scenarios such as local-first AI governance, candidate evaluation, model provenance, scorer evolution, and rollback-capable cognitive lab tooling. The project was built, tested, and validated on a 2012 Intel Mac, demonstrating that it can run on old consumer-grade hardware.
Limitations
- External open-source datasets are acquisition targets only and have not yet been bundled, ingested, or promoted into production weights.
- Current knowledge weights are borne by self-curated ARC materials and the ARC Language Module; Tiny/Small model weights are circular-proof reference models, not a complete brain.
- The project is explicitly not a drop-in replacement for frontier models (such as Gemma, Llama, Claude, GPT).
- Public comparisons must be based on the same locked scorer version, benchmark manifest, capability surface, adapters, prompt configuration, and candidate artifacts; otherwise old scores and new scores are not directly comparable.
- Some features (such as ARC-StreamMemory, the Synth roadmap) are still actively being built and not yet complete.
- The sponsorship program does not guarantee AGI, return on investment, or custom software delivery; custom software delivery requires a separate written agreement.
- Specific information such as fees and pricing is not currently provided in the materials.
