Development log and research archive for the Chimera ecosystem.
Chimeraforge is the public-facing site for the Chimera ecosystem — a constitutional AI enforcement architecture spanning 9 repositories across Python, Rust, TypeScript, C#, and JavaScript. The site publishes three things:
- Episodes — narrative write-ups generated from git commits across the ecosystem. Each episode is written as a roundtable discussion between four AI personas, turning raw development activity into readable stories.
- Research reports — technical benchmarks covering LLM performance, quantization, inference optimization, and safety alignment. 55 reports spanning 1.34M+ measurements.
- Papers — independent research write-ups. 1 paper accepted at the ICML 2026 Workshop on Hypothesis Testing, 5 under peer review, 5 in preparation.
The site is live at chimeraforge.vercel.app with 268 episodes. The per-commit episode stream is now archived — the final episode in each line is a full retrospective, and ongoing work is documented through the research reports.
Chimeraforge documents work across nine repositories:
| Repository | Purpose |
|---|---|
| Banterpacks | Core monorepo — JARVIS gateway, constitutional router (TDD002), debate engine (Chimera), Rust runtime (TDD005) with ZK proofs + BFT, RLAIF alignment loop, authoring. |
| Banterhearts | ML research platform — inference API, benchmarking, AutoOpt agent, safety evaluation. 1.34M+ measurements. |
| Chimeraforge (PyPI) | LLM deployment optimizer on PyPI (v0.12.3). Model-agnostic 5-gate capacity planner (VRAM, quality, safety opt-in, latency, cost) — plans any registry / Ollama / HuggingFace model across 22 GPU profiles, and serves the same numbers to AI assistants over MCP. Every number labeled measured / estimated / unknown. |
| Chimera Multi-Agent | Muse Protocol — 6-agent content pipeline + observability control plane (OTel, ClickHouse, Datadog). |
| Chimeradroid | Unity/C# Android companion — voice, chat, tool approval, session handoff, mesh networking. |
| Echo | Messaging channel adapters — Slack and Discord bridges to JARVIS. |
| JARVIS Console | Next.js web console — chat, control room, cognitive ELO dashboard, admin panels. |
| Project Wyvern | Embodied autonomy — governed mission-execution plane between Chimera control and PX4/ArduPilot. SIM-ONLY MVP on PX4 + ROS 2 + Gazebo. |
| Banterblogs | This repository. Development log and research archive. |
Not part of the nine — independent CLIs that productize the research line:
| Tool | Purpose |
|---|---|
| quantfit (PyPI) | Quantize an LLM and check it still refuses what it should (v0.6.1, Apache-2.0). AWQ / GPTQ / SmoothQuant / FP8 / RTN + GGUF, honest 3-tier capacity refusal, and a two-axis safety-drift check (refusal-robustness + over-refusal) against an unquantized baseline. |
Both shipped CLIs have pages at /tools.
Banterblogs/
banterblogs-nextjs/ # The active Next.js application
src/ # App source (pages, components, lib)
posts/ # 268 episode markdown files (stream archived)
reports/ # Technical report data (benchmarks, analysis)
PublishReady/ # Processed reports ready for rendering
legacy_archive/ # Original static site (archived, read-only)
- Next.js 16 (App Router) with React 19 Server Components and Static Site Generation
- TypeScript in strict mode
- Tailwind CSS 3.4 with a custom dark theme ("Obsidian & Ember" — single hot-copper accent)
- Framer Motion for animations
- Markdown pipeline: remark, rehype, highlight.js
- Search: Fuse.js
- Deployed on Vercel
cd banterblogs-nextjs
npm install
npm run devOpen localhost:3000.
To enable the research hub, copy env.example to .env.local and set REPORTS_ENABLED=true.
- Fork the repository.
- Create a feature branch (
git checkout -b feature/your-change). - Commit your changes.
- Open a pull request.
MIT. See LICENSE for details.