Skip to content

nah414/Phoenix

Repository files navigation

Phoenix

An AI-agent-consumable middleware harness for accuracy, performance, and safety.

Phoenix is downloadable software that AI agents and integrators call to get validated quantum computation, hardware-aware routing, and physics-grounded verification of results — with provenance, error bars, and honest reporting of uncertainty.

It is not a SaaS, not an end-user app, not a chat interface, and does not host an LLM. It is a power tool that sits between a software stack (or an agent framework) and the underlying compute (local accelerators, cloud quantum providers), making any system that integrates with it more accurate and more honest about its uncertainty.

Phoenix v1.0 is released (GitHub Release, Apache 2.0). Active development continues on the v1.1 line.

Documents

  • PHOENIX_ARCHITECTURE_v1.md — the locked v1 architecture spec (revised to v1.1 on 2026-05-07). Covers the Trinity Core physics heart, the seven wrapping layers, mandatory three-axis wobble verification, hashchained Omega Ledger provenance, end-to-end cost-ceiling enforcement, and the Phoenix Cloud commercial path.
  • docs/ — the documentation hub: distribution (install/run) and reproducibility guides, plus extension planning.

Future extension planning (not part of locked v1)

Status

Current release: v1.0.0 (Apache 2.0), tagged 2026-05-28. Active development continues on the v1.1 line.

The full v1.0 surface is built and tested end-to-end. What's live:

  • Trinity Core — the physics heart (Solver, Control, Orchestrate) wired end-to-end, with mandatory three-axis wobble verification (cross-precision, cross-control, cross-provider).
  • Provider routing — a multi-stage router with provider registry, ranking, and failover. Cloud SDK adapters (IBM, AWS Braket, IonQ) ship as stubs that raise on submit; real wiring lands in a focused later phase.
  • Verification & safety gates — adaptive verification with reactive promotion, plus identity (Ed25519), permissions, rate limiting, and a fail-closed kill switch.
  • Audit & provenance — structured audit events, a hashchained Omega Ledger, and bit-exact replay that re-executes a recorded task and verifies its result hash.
  • Cost & cloud seams — 24h cost-ceiling enforcement, plus Phoenix Cloud abstraction seams that let tenant-aware implementations drop in without changing Phoenix core.
  • Front door — a full surface across REST, WebSocket, CLI, and an MCP server (phoenix mcp serve) for IDE clients, plus an admin dev-ops API and LoRA adapter subsystem.
  • Distribution — three release artifacts: pip wheel (phoenix-middleware), Docker image (ghcr.io/nah414/phoenix), and Nuitka standalone binary (Linux + Windows), all CI-built across Python 3.11/3.12/3.13.

In progress (v1.1): a cognition substrate — CognitionProvider adapters (Anthropic / OpenAI / Google / LiteLLM), three cognition wobble axes, and a per-server registered MCP-client mode — shipped 2026-05-20 (Phase 13). Follow-on work covers encryption administration and per-actor key isolation.

Planned / deferred: the reference admin client (a separate repo, deferred to v1.1) and the perception-harness extension (plan locked; build-guide drafting unblocked).

See PHOENIX_ARCHITECTURE_v1.md Sections 10.7 and 10.8 for the v1 and v1.1 acceptance criteria.

License

Apache License, Version 2.0. See LICENSE.

The Apache 2.0 license is the explicit choice for Phoenix per architecture Decision 34 — open source, ecosystem-compatible, and crucially includes a patent grant as belt-and-suspenders against future patent claims on calibration methodology.

Author

Adam (@nah414) — co-researcher and builder. dr-frank-and-eddy is our lab bench where physics evolves; Phoenix is the production middleware that vendors a frozen v6.6 snapshot of that work plus SynQc TDS Core into a single Trinity Core substrate.

Phoenix is built with Claude as design partner.

About

Ai powered autonomous harness for increasing accuracy, performance and security.

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages