Skip to main content
Enoch is an agentic research control plane that treats autonomous AI work as a reliability engineering problem, not a scripting problem. It queues research ideas sourced from an upstream LLM scouting process, gates dispatch through authenticated preflight checks, supervises agent runs on local AI hardware using process-tree tracking and CPU/GPU telemetry, captures evidence into structured bundles, and generates publication-style research artifacts with provenance instead of free-floating model output. Key transitions are guarded by idempotent APIs, stale-state reconciliation, and operator-visible status so you can inspect what is running, what is queued, and what the hardware appears to be doing.

​
The problem Enoch solves

Autonomous research runs fail in mundane but important ways:
  • an agent session can finish while child processes keep running;
  • a GPU worker can still be busy when a queue item looks idle;
  • queue state can become stale across dashboard views, intake projections, and runtime files;
  • evidence can be scattered across worker folders;
  • generated reports can overstate unsupported claims if provenance is lost.
Enoch addresses those problems with explicit state, authenticated APIs, worker preflight, process tracking, telemetry thresholds, evidence bundles, claim ledgers, packaging/provenance lint, and an explicit strict claim/evidence audit status. It also keeps quality buckets separate: operational success, promising-signal preservation, paper-corpus readiness, and public trust posture are different claims.

​
Architecture flow

LLM-assisted idea scouting
  -> structured idea card
  -> control-plane idea/project/queue rows
  -> Enoch queue candidate
  -> FastAPI control plane
  -> worker preflight and dispatch safety checks
  -> worker gate
  -> agent run with process + telemetry supervision
  -> evidence sync
  -> generated research artifact
  -> corpus quality and provenance gates
Control-plane intake and the upstream scouting workflow are intake surfaces. Notion sync remains a legacy compatibility path. Runtime authority begins when an idea becomes an Enoch queue item. Current production storage is local Postgres on enoch-core; older Supabase wording is migration/compatibility terminology. For the canonical current host, storage, worker, decision-artifact, automation, and compatibility boundaries, see current runtime snapshot.

​
Main components

  • Control plane API — queue/project/run/paper state, pause controls, dashboard data, dispatch decisions, and publication automation APIs.
  • Worker gate — process-tree and CPU/GPU quiet-window checks for long-running local jobs.
  • Worker preflight — authenticated checks against the worker API before live dispatch.
  • Single active lane — the control plane blocks overlapping active queue work.
  • Evidence sync — copies evidence from worker project folders before artifact generation when enabled.
  • Artifact writer — creates draft research artifacts from evidence context using a deterministic template or an optional OpenAI-compatible provider.
  • Corpus packaging — stores generated artifacts with provenance and quality metadata.

​
What Enoch is not

Enoch is not a peer-review system, a guarantee of scientific novelty, a replacement for expert review, or a collection of old workflow-tool exports. The corpus is useful because it preserves generated work and evidence boundaries, not because every generated claim is true. The current public status separates 389/389 packaging/provenance lint from 389/389 strict claim/evidence audit pass. For the full review boundary, read quality floor and signal visibility.
The packaging/provenance and strict audit counts above are the values reported by enoch-ai-research-corpus quality/packaging_provenance_report.json and quality/claim_evidence_audit.json at the time this page was last audited. Re-run the corpus release validators or check the canonical JSON before quoting them externally; counts can change as the corpus is re-imported.

​
Recommended next step

If you are new, read the quickstart first. If you are reviewing generated papers, read authorship and provenance and quality floor and signal visibility before judging or citing any artifact.