Agentic research needs a control plane

Local AI research needs a control plane.

Enoch turns idea queues into bounded local runs: dispatch gates, worker telemetry, evidence sync, and publication packages with explicit provenance. Run the proof, then inspect one artifact end to end before trusting the corpus.

What Enoch is

A reliability layer for long-running autonomous work.

Enoch owns queue state, pause and maintenance controls, worker preflight, single-lane GPU safety, process and telemetry completion truth, evidence sync, artifact rewriting, and dashboard status. Built on FastAPI with a hard state contract; developed in collaboration with Codex.

The generated papers are released as AI-generated research artifacts. The human operator built and operated the surrounding Enoch system; Codex CLI is credited as the worker execution substrate, not as the owner or author of the generated paper prose, arguments, or experimental results.

New here? Start with Enoch Docs for the quickstart, deployment guide, configuration reference, API notes, and provenance caveats.

Workflow

From idea intake to public artifact.

1. ScoutLLM-assisted review of technical signals, public papers, news, and local runtime opportunities.
2. ScoreA source-of-truth scoring table ranks candidate ideas before execution.
3. DispatchControl plane gates queue state, maintenance state, worker health, and single-lane safety.
4. ExecuteGB10 worker runs the experiment while wake gate watches process and telemetry truth.
5. PreserveRun notes, evidence bundles, claim ledgers, metrics, and manifests are synced.
6. PublishThe artifact writer emits evidence-grounded reports with provenance, caveats, and release checks; bounded useful no-paper signals stay in a separate promising-signals lane.

Run the local proof

Verify the control-plane path before trusting the corpus.

This proof validates the local API/dashboard path, worker preflight self-check, and dispatch dry-run behavior. It does not validate scientific correctness, peer review, independent replication, strict claim/evidence auditability, or production readiness of generated artifacts.

git clone https://github.com/alias8818/enoch-agentic-research-system.git
cd enoch-agentic-research-system
scripts/proof-local.sh

Expected smoke output includes PASS healthz, PASS control status, PASS wake gate healthz self-check, and PASS dispatch dry-run. For setup details, read the quickstart.

Inspect one artifact end to end

Use one bounded artifact as the audit path.

vLLM Attention-Sink Retention 3B Continuous-Serving Stress Campaign is a concrete serving-performance artifact with limitations and reproducibility caveats. Treat it as inspectable, not independently replicated.

  1. Read the generated report.
  2. Open the evidence bundle.
  3. Open the claim ledger.
  4. Read the limitations and strict audit report before treating any generated claim as evidence-backed.

Strict audit examples

Representative strict-audit-passing artifacts.

Every artifact now clears the strict gate: non-empty claim ledgers, every claim linked to an evidence reference, and every referenced result file either publicly present or explicitly declared unavailable with a hash and public surrogate. The examples below are representative audit paths, not the only passing artifacts.

Read the strict audit report for the full corpus-level gate result and per-artifact evidence accounting. Bounded useful or scale-blocked no-paper leads are kept in the promising-signals repo instead of being counted as corpus papers.

System diagrams

Control plane first, artifacts second.

These diagrams are intentionally editable SVGs so the launch story stays crisp: queue safety and provenance are the product, while generated papers are inspectable outputs.

Enoch architecture diagram from idea intake through control plane, wake gate, worker, evidence, publication, dashboard, and alerts
Architecture: idea intake enters a gated control plane before local execution, evidence sync, paper packaging, dashboard visibility, and alerts.
Provenance diagram linking generated text to metadata, evidence, claim ledger, and publication surface
Provenance: generated text is only useful when attached to metadata, evidence bundles, limitations, and review/audit state.

Public framing

Built in the open, with provenance up front.

Release Enoch as an engineering artifact: a working control plane, a corpus of generated outputs, and a transparent record of what was automated. The hook is not “these are human papers.” The hook is that a local agentic system generated, supervised, rewrote, and packaged a research corpus with evidence bundles, claim-ledger files, packaging/provenance lint checks, strict claim/evidence audit gaps, and caveats—and some generated artifacts are specific enough to invite independent replication attempts.

Launch TODO · Announcement drafts · Enoch Docs · Codex CLI