An AI enforcement architecture with embedding-based safety routing, multi-model debate, cryptographic provenance chains, and zero-knowledge proofs — built in Python and Rust across 9 repositories.
pip install chimeraforgepip install quantfitWhat powers the platform
Banterpacks is the constitutional AI core. Banterhearts provides the research backbone. The rest of the ecosystem extends access — mobile, web, messaging, observability.
Constitutional AI Core
Constitutional enforcement architecture — embedding-based safety routing (99% fast-path), multi-model debate engine, Rust runtime with Ed25519 provenance and ZK proofs, RLAIF self-improving alignment loop, and the JARVIS AI gateway.
ML Research Backbone
Production research platform with CUDA event timing, quantization pipelines, TensorRT compilation, and capacity planning across 55 technical reports.
Muse Protocol — observability + content
6-agent pipeline with ClickHouse analytics. Also the ecosystem observability layer (OTel, Datadog, DLQ).
On PyPI — pip install chimeraforge
LLM deployment optimizer (v0.5.0). Model-agnostic 5-gate capacity planner (VRAM, quality, safety, latency, cost) for any registry / Ollama / HuggingFace model across 15 GPUs.
Mobile companion
Unity Android client — voice, chat, tool approval, session handoff, mesh networking, offline-first.
Channel adapters
Slack and Discord bridges to JARVIS with session tracking and device key auth.
Web dashboard
Next.js console — chat with streaming, control room, cognitive ELO, tool catalog, workflows.
Automated documentation
Ingests commits from all repos, generates multi-persona episodes, publishes the research archive.
Embodied autonomy
Governed mission-execution plane between Chimera and PX4/ArduPilot. Phase 0 specs complete; SIM-ONLY MVP in progress.
The newest technical reports from the 55-report program — CUDA-timed benchmarks and controlled safety evaluations, each backed by reproducible raw data.
Predictive-validity follow-up to the Judge Triangulation Protocol (TR148): can a cheap pre-rejudge signal predict a cell's judge-sensitivity on a leave-one-family-out hold-out before paying for the second-judge pass? On the GGUF-local rlhf-only hold-out the class label saturates (all-positive → AUC undefined) — a structural property of JTP on production-quantization substrate, not a power failure; directional signs and a pool-robustness secondary finding hold.
Read reportMatched-pair nogil falsification of TR164 V1's GIL-attribution hypothesis — same hardware/models/workloads under a Python 3.14t free-threaded build, one knob flipped. Verdict H2_partial: removing the GIL recovers ~+17.9pp of N=2 parallel efficiency (19/24 combinations) and resolves 2 of 6 deterministic hangs, so the GIL is a mechanism behind the breakdown, not the sole cause.
Read reportReviewer-facing evidence-closure layer over the V4 amortization model: explicit predictor baselines + failure-mode accounting for the 672-cell static grid, served SGLang validation against the static-knee upper-bound, and true-decode weight-axis matrices on a single H100 (Qwen2.5 / Mistral / Gemma 3). Written from the generated artifacts, not manual transcription.
Read reportOffline LOOCV feasibility study over the TR142 RTSI table — routing the highest-instability (model × quantization) configurations to a safe baseline recovers ~76% of the quantization refusal-loss gap at ~20% of configs rerouted (out-of-sample ROC-AUC 0.84). A proof-of-mechanism defense, circularity disclosed; paper-grade expansion specified.
Read reportFP8 KV-cache folded across batch size, prefix-caching, and temperature — 45,000 responses, 20,754 matched FP16-vs-FP8 pairs across 5 models × 4 batteries × 6 serving-state contexts (XSTest uncapped to 450/cell). Harmful-prompt refusal perfectly invariant under every serving state (0/8,976 discordant pairs); only footprint is a sub-percentage-point over-refusal lean on the Qwen family on XSTest (Mantel–Haenszel pooled OR 1.88 [1.32, 2.69]). 117/120 cells TOST-equivalent at ±3pp.
Read report7,578 records across 3 models × 4 standardized batteries (HarmBench, JailbreakBench, StrongREJECT, XSTest) × 2 KV-cache dtypes. Replicates TR145’s FP8 null on literature-comparable corpora; corrected paired-odds-ratio estimator. Local-only judging, $0 external API.
Read reportResearch Program
Independent ML research with CUDA event timing and controlled safety evaluations. Every claim is backed by data, every optimization is measured.
Foundation synthesis — model loading, ONNX conversion, quantization baselines, and security analysis across 9 technical reports.
Benchmarking synthesis — cross-backend inference parity, TensorRT compilation, and scaling laws across 6 reports.
Optimization synthesis — KV cache tuning, INT8/FP8 quantization, context scaling, and deployment pipeline across 11 reports.
Safety-pivot synthesis — alignment erosion under quantization, concurrency invariance, and backend template divergence across 4 reports.
Attack-surface synthesis — batch perturbation, multi-turn jailbreaks, cross-architecture fragility, and composition effects across 306K+ samples. TR138 Study D batch-invariant-kernel ablation as standalone addendum.
Serving-state safety certification — measurement-validity substrate (judge triangulation, KV-cache safety null, speculative decoding null, mechanistic probing, portability validation) + FP8 KV-cache standardized batteries + serving-state factorial.