Skip to content
Chimeraforge

Platform Architecture

What Powers Chimeraforge

Nine repositories across Python, Rust, TypeScript, and C#. Constitutional AI enforcement with cryptographic provenance, multi-model debate, and a self-improving alignment loop — from the JARVIS gateway to mobile clients to channel adapters.

  • 9 repositories
  • 4 languages
  • 1.46M+ research measurements

Core engines

  • Banterpacks

    Python, Rust · 6 subsystems + 7 Rust crates

    The core monorepo. Six interconnected subsystems handling constitutional AI enforcement, multi-model debate, cryptographic provenance, and the JARVIS multi-modal assistant. All inter-service coupling is HTTP.

    JARVIS Gateway
    AI agent layer — chat, voice, semantic memory, tools, proactive intelligence, smart home
    Constitutional Router
    TDD002 — embedding cosine similarity, calibration (isotonic/Platt), calibrated fast-path routing
    Debate Engine
    Chimera — heat-based escalation, weighted voting, ranked choice, Condorcet consensus
    Rust Runtime
    TDD005 — Ed25519 provenance, BFT consensus, ZK proofs, cognitive agents with ELO

    RLAIF self-improving loop · 3-stage tool approval

    Read the Banterpacks dev log

  • Banterhearts

    Python · 1.46M+ measurements

    ML research platform and production inference backbone. 1,463,000+ measurements across 55 technical reports, targeting consumer GPUs with sub-100ms inference.

    Inference API
    Model selection, streaming, multi-backend dispatch
    Safety Research
    Alignment under quantization, concurrency, cross-backend consistency
    Benchmarking
    4 compilation backends, 5 quantization formats, GPU kernel profiling
    AutoOpt Agent
    Thompson sampling, multi-armed bandit, SLA enforcement

    37-file evaluation framework · 20 monitoring modules · 12 security modules

    Browse the research archive

JARVIS Gateway

Multi-modal AI gateway with multi-provider routing (Anthropic, OpenAI, Gemini, Ollama). Every chat turn routes through the constitutional router. Tool execution uses a 3-stage propose/approve/execute pipeline with cryptographic provenance on every action.

A chat turn through the gatewayEvery turn passes the constitutional router; a tool runs only once it is proposed and approved, and every action carries cryptographic provenance.

  1. Chat turn

    text or voice, on Anthropic, OpenAI, Gemini or Ollama

  2. Constitutional router

    every turn routes through it

  3. Propose

    a tool call, as a discrete step

  4. Approve

    human in the loop

  5. Execute

    sandboxed, with cryptographic provenance

Modules

  • Calendar
  • Inbox
  • Memory
  • Smart Home
  • Proactive
  • Tools
  • Voice

Supporting systems

In development

Standalone tools

Independent CLIs shipped outside the Chimera ecosystem — their own repositories, not counted among the nine.

How data flows

  1. 1

    Ingest

    Banterhearts bench data and Banterpacks git commits flow into ClickHouse via dedicated agents.

  2. 2

    Process

    Watcher monitors pipeline health. Council generates episodes with performance insights baked in.

  3. 3

    Publish

    Publisher pushes episodes to GitHub. Vercel rebuilds the site. i18n translates to German, Chinese, Hindi.

  4. 4

    Serve

    JARVIS gateway dispatches inference locally. Chimeradroid extends access to mobile devices.

See it run

The subsystems above are not diagrams on /show. Each scene replays pre-computed records from the Banterpacks pipeline — real Ed25519 signatures, real Pedersen commitments, real tier verdicts — and labels the one deterministic stand-in where the public demo uses it.

Start with the five-tier streaming ladder, the zero-knowledge alignment proof, or BFT consensus across four replicas.