HN 표시: Enoch – 자율 AI 연구를 위한 제어 평면

Enoch is an agentic research control plane: it queues ideas, gates dispatch, supervises local AI runs, preserves evidence, and packages AI-generated research artifacts with provenance instead of pretending autonomous work is just a script. Long-running autonomous AI work fails in ways ordinary scripts do not: - child processes continue after an agent session appears idle; - GPU workers can still be active when queue state says no work is running; - queues become stale or disagree across sources; - evidence scatters across machines and run folders; - generated reports overstate results when claim boundaries aren't preserved. Enoch treats those as control-plane problems, not model problems. It uses process tracking, CPU/GPU quiet-window telemetry, idempotent APIs, stale-state reconciliation, a live dashboard, evidence bundles, and claim ledgers to make autonomous work observable and auditable. Agentic AI systems need control planes. A model can propose and execute work, but a separate system should decide what is queued, what is safe to dispatch, whether work is actually done, and what evidence supports the final artifact. LLM research scout -> structured idea cards -> Notion scoring / weight matrix -> queue candidate -> VM control plane -> worker preflight and dispatch safety checks -> GB10 worker wake gate -> agent run with process + telemetry supervision -> evidence sync -> AI-generated research artifact -> corpus quality gates The repository contains the execution/control-plane layer and supporting docs. Historical notes describe earlier migration experiments, but this is not a workflow-export repository and does not ship workflow-tool configurations. - Control plane API — queue state, project state, paper review state, pause/maintenance controls, and dispatch decisions; built with FastAPI and LangGraph-era graph boundaries. - Wake gate — proves a run is actually done, not just agent-session-closed: process-tree tracking and CPU/GPU quiet-window telemetry sustained over a configurable window. - Worker preflight — authenticated health checks against the worker before dispatching new work, so dispatch fails early rather than silently. - Single-lane safety — prevents overlapping GPU-heavy work on constrained local hardware; the control plane holds the lock, not the dispatch script. - Evidence sync — copies run notes, metrics, result summaries, evidence bundles, and claim ledgers from worker projects into the control plane before artifact generation begins. - Artifact writer — generates publication-style Markdown reports from evidence context while preserving uncertainty and provenance; does not free-float against raw model output. - Quality gates — scans generated reports for placeholder citations, missing provenance, and missing evidence artifacts before they enter the corpus. The reports produced by Enoch runs are AI-generated research artifacts, not human-authored or peer-reviewed papers. They are built from run notes, evidence bundles, claim ledgers, and reproducibility traces. The maintainer releases the corpus for inspection and critique but does not claim personal authorship of the generated papers, arguments, or prose. See docs/release/authorship-and-provenance.md for the full framing and recommended citation language. Enoch is the project-specific control plane and release package. It runs agent work through Codex/OMX automation, including oh-my-codex orchestration for local agent execution. OMX is part of the operating substrate; generated research artifacts are produced by Enoch runs and the artifact writer, not by OMX as an owning publisher. Ideas are sourced from an upstream LLM-assisted scouting process that reviews technical signals such as AI news, public research papers, systems discussions, and local hardware/runtime opportunities. Candidate ideas are framed as structured experiment cards, scored in a Notion weight matrix, and handed to Enoch as queue candidates. Notion is best understood as an intake and prioritization surface. Runtime authority begins in the Enoch control plane. See docs/idea-intake-workflow.md . For a local developer smoke test, start with docs/quickstart.md . For a full deployment (control VM, worker machine, systemd service, dashboard/API smoke tests, optional Pushover alerts, dispatch checks, and paper-writer settings), see docs/deployment-guide.md . For individual config fields, start from config.example.json and see docs/configuration-reference.md . Required values: - inbound API bearer token - completion callback URL/token - project root and dispatch script path - worker URL/token - optional notification and paper-writer provider settings Never commit live config files or credentials. uv run pytest -q Using Enoch: docs/quickstart.md — local clone-to-dashboard smoke testdocs/deployment-guide.md — full deployment guidedocs/configuration-reference.md — config field referencedocs/system-workflow.md — architecture and control-plane boundariesdocs/idea-intake-workflow.md — LLM scouting, Notion scoring, and queue handoff Release context: docs/release/authorship-and-provenance.md — how generated reports should be frameddocs/featured-paper-selection.md — rationale for the launch highlight setdocs/outreach/launch-announcement.md — draft launch copy and repo descriptionsdocs/launch-checklist.md — public launch checklistdocs/launch-todo.md — remaining public-release gatesdocs/historical/ — historical migration notes retained for engineering context onlysite/ — static launch site Before publishing or deploying changes, run secret scans and tests. See SECURITY.md . Apache License 2.0. See LICENSE .