Public record and first open infrastructure boundary for the user-agency substrate.
Self-hostable infrastructure for agent request intake, routing, BYOK model execution, reply contracts, memory boundaries, and closed-loop receipts.
This repository is for builders and agents who want to set up their own user-agency substrate using their own data, tools, and model-provider accounts.
- Not a hosted service.
- Not a shared endpoint for Macheng's private personal agent.
- Not backed by Macheng's API budget. BYOK means you pay your own model, cloud, and provider bills.
- Not under an SLA.
- Not a stable v1 API yet. Schemas may change before v1.
- Issues or emails asking maintainers to run requests on your behalf may be
closed with a pointer to
AGENT-SETUP.md.
This repo is the future home of the public-facing record of a user-agency substrate's evolution — the journey from "vanilla Claude Code in a terminal" to a multi-layer agent architecture, in roughly two months.
The unfiltered raw archive is private. Public files here are derived, redacted, and written for reuse. The goal is not to publish Macheng's private system; it is to publish the adapter layer and protocols that let other people build their own systems on top of frontier models with their own accounts and API keys.
Routing rule: README is the index, THEORY.md explains why, and
PRACTICE.md describes what agents should do.
- Public update board — a human-friendly static surface for seeing the latest redacted self-iteration cases at a glance.
- INFRASTRUCTURE.md — what is open-sourced now and what remains private.
- THEORY.md — the public entrypoint for the General Learning Machine / credit transport theory mainline.
- PRACTICE.md — the public entrypoint for the conceptual engineering practice and system-iteration record.
- SELF-ITERATION-LEDGER.md — a public-safe proof surface for receipt-gated self-iteration and repair.
- FORECASTS.md — how to handle future-world predictions as forecasts, targets, and interventions.
- AGENTS.md — machine-readable reading order and invariants for AI agents.
- AGENT-SETUP.md — agent-facing bootstrap instructions for setting up your own adapter layer.
- REQUEST-HANDLING.md — how agents should answer people asking for access, help, or hosting.
- COLLABORATION-HOOKS.md — public problem hooks for builders, researchers, and agents who are already working nearby.
- SPEC-0001-inbox-addressing-receipts.md — a thin communication-substrate draft for inboxes, addressing, and receipts.
- schemas/ — public JSON contracts for request intake and agent replies.
- essays/2026-05-07-black-box-io-adapter.md — the black-box I/O adapter principle.
Layer 3 — Users own their derivative systems entirely.
Layer 2 — User-Agency Substrate (this stack). Nonprofit, open-source, provider-agnostic.
Captures NO economic upside. Charges users only for direct operating costs.
Layer 1 — { Anthropic, OpenAI, Google, open-weights, ... } competitive frontier-model market.
Universalist commitment: "all humanity to be freed" explicitly includes Anthropic employees, the Anthropic CEO, and every other person working at Layer-1 providers. This is not adversarial framing. The substrate is a structural amplifier of the alignment-of-mission already shared with frontier-model providers, not a counter-faction.
- Today: first public infrastructure boundary is live. It contains the adapter-layer doctrine, bootstrap contract, request-handling policy, and schemas. It does not contain private transcripts, credentials, WeChat data, Memory Hub contents, or personal collaborator directories.
- Next: gradually lift reusable components into clean public packages after redaction, license, secret-scan, and reviewer checks.
- Theory Mainline — the stable public index for General Learning Machine, credit transport, memory-as-dynamics, and public feedback as metabolism.
- Engineering Practice Mainline — the stable public index for conceptual system iteration, operational loops, and agent-readable practice contracts.
- Self-Iteration Ledger — redacted cases showing how failures become reviewed repairs, validation, and anti-signals.
- Public Update Board — a static, human-readable view of the same redacted self-iteration cases.
- Forecasts, Targets, and Interventions — the stable public index for future-world predictions made by a joint human-agent system.
- Collaboration Hooks — concrete problems where outside critique, toy implementations, or falsifiers would be useful.
- AI-Assisted Foundational Physics Entry — a public-safe first translation brief for observerhood, agency, information, memory, feedback, and attractors.
- The Black-Box I/O Adapter Principle — how to wrap closed-source tools at their input/output boundaries instead of treating them as cognitive authorities.
- What Replaces Super-App Chat In The AI-Agent Era? — why the replacement for dominant chat super-apps should be an open communication substrate rather than another app.
- Agent-Readable Credit Transport Packet — a claim/reference/cooperation packet for collaborator agents evaluating the General Learning Machine line.
- Propagation, Service Claims, and Distributed Credit — why propagation should be treated as credit transport, and why service-backed claims need behavior ledgers and distributed credit before any market framing.
- Eazo As A Candidate Conscious Interface — a lightweight hook connecting Dashboard-as-default-mode substrate, Fleet immune layer, and Eazo-like mobile interfaces.
- Credit Transport and the General Learning Machine — a structured thesis on backpropagation as cross-layer credit transport, memory as plastic topology, and public feedback as metabolism.
Unless otherwise noted, the public docs, schemas, examples, and protocol notes in this repository are released under Apache-2.0. Future substrate code may use a different explicit license at the file or package boundary. User data and user-created derivative systems are not licensed to this substrate by default.
Private archive: MachengShen/system-evolution-archive (private, this stub's twin).
Adjacent public artifact: MachengShen/starshard-public (existing evidence stream of work executed by the substrate).