DarkPrint

DarkFactoryPrint

The blueprint registry for autonomous AI factories.

Autonomy you can read as a graph.

A dark factory runs with the lights off — no operators, only agents that plan, execute, verify and ship on their own.

scroll on ↓ to see what we mean

scroll

The premise

A dark factory runs with the lights off

A plant so automated it needs no lights — no workers on the floor, only machines running themselves. Ported to AI, it's a pipeline where autonomous agents own the whole arc, from plan to ship, without a human standing over each step.

Ordinary automation follows a script and stops at every branch it wasn't told about. A dark factory is different on three counts — and each one is what makes it worth reading as a graph rather than trusting as a black box.

◆no checkpoints

Decision autonomy

No human-approval checkpoints sit on the critical path. When a run hits an error, the factory decides how to recover and keeps moving — it makes the judgement call instead of stopping to ask.

↻plan · execute · verify · ship

A closed loop

Agents plan, execute, verify and ship on their own. There is no fixed script to march through — the loop closes only when the work clears its own acceptance criteria, and re-opens when it doesn't.

❯specs, not lines

Specs over code

The engineer stops writing code line by line and starts writing specs and acceptance criteria. The leverage moves up a level — and so does the debt: technical debt becomes specification debt.

8Blueprints

6Parts

3Ontologies

6Builders

61.4kPulls

What lives in the registry

Three things you can publish

Everything on DarkPrint is a graph at some altitude — a whole factory, a piece of one, or the vocabulary they're written in. Each is a first-class, downloadable artifact.

▧Blueprint8

The whole factory

The complete graph of a dark factory — every agent, tool, human gate and edge, from the trigger to the ship node. Versioned, scored, and pulled as one piece.

Browse blueprints

◫Part6

Reusable sub-graphs

A retry loop, a validation gate, a negotiation node — packaged on its own with a typed interface. Like npm packages for orchestration logic: drop one into any blueprint.

Browse parts

PlannerExecutorVerifierRouterNegotiatorGate

⬡Ontology3

Typed vocabularies

The node and edge kinds a factory is built from. Ontologies are the grammar that lets a pipeline be encoded as a graph — and then read, graded and compared as one.

Browse ontologies

How a factory is graded

Six metrics, three ways of knowing

Every blueprint carries the same scorecard. What makes it trustworthy is that each axis is honest about where its number came from — computed, measured, or voted — and colour-coded so you can tell at a glance.

Static analysisauto

AutonomySecurity

The analyzer walks the graph and counts human-approval gates and requested tool scopes. Nothing is executed — the score is a property of the structure.

Measured on runmeasured

Cost / time

Recorded objectively on a real run: median tokens and wall-clock, reported through opt-in telemetry.

Community votevoted

EfficacyReliabilityTransparency

Aggregated from weighted community and validator votes over real executions — the subjective half of the card.

Adversarial Consensus Line · scorecard

The card, axis by axis

0–100

Autonomy and Security fall straight out of the graph; Cost/time is measured on a run; Efficacy, Reliability and Transparency are the community's call.

Autonomy
auto92
No approval nodes; conflict is resolved by re-vote, not a human.
Efficacy
voted84
Community-rated task success on real runs.
Reliability
voted78
Rated across repeated executions without error.
Transparency
voted88
How well the internal decisions are documented.
Cost / time
measured46
Median tokens and wall-clock recorded on execution.
Security
auto74
Read-only tools; no write scopes requested.

The autonomy ladder

From assisted to closed-loop

Autonomy is a discrete level, not a vibe. It's the headline axis of the scorecard, and it climbs exactly as far as the graph lets it.

A1Assisted
A human drives; agents assist step by step.
A2Supervised
Agents do the work, but a human approves the critical move.
A3Conditional
Self-directed within guardrails; escalates only the edge cases.
A4Closed-loop
Plans, executes, verifies and ships with no human in the loop.

⏸

A single human-approval gate anywhere on the critical path caps a blueprint at level 2. The static analyzer counts human gatenodes straight off the graph — so autonomy is measured, not claimed. You can't self-report your way to closed-loop.

The two seed blueprints

Where naive pipelines break

Anyone can chain a few agents when everything goes right. A factory earns its autonomy on the two moments that wreck a happy-path pipeline — a disagreement, and a mid-run failure. These are the reference blueprints we ship.

The hard part· When two competent agents disagree

Blueprint

Adversarial Consensus Line

Two agents solve the same task from opposite temperatures, then a consensus node negotiates a single answer — re-opening the debate when they clash.

#multi-agent#negotiation#consensus

MVMara Veiga

↓ 3.1k▲ 214

The hard part· When a stage fails halfway through

Blueprint

Checkpoint & Resume Runner

A staged pipeline that snapshots state after every stage, so a failure at stage 3 resumes from the last good checkpoint instead of restarting the whole job.

#reliability#checkpointing#recovery

KOKwame Obasi

↓ 4.8k▲ 301

How the numbers stay honest

The people running factories set the scores

A registry is only as trustworthy as its data. Two mechanisms keep DarkPrint's scores grounded in real usage rather than self-promotion — one for the measured half, one for the voted half.

⇡Telemetry · opt-in

Your runs sharpen the scores

Run a blueprint locally and choose to send back run metrics — cost, latency, pass/fail. It is explicit opt-in and off by default; nothing leaves your machine unless you say so. In return, your executions feed the measured Cost/time axis everyone sees.

Off by default
Run metrics only, no payloads
Earns reputation points

✓Validators · earned

Votes that carry weight

A validator badge is earned through a track record, not bought. It weights your votes more heavily on the community axes — efficacy, reliability, transparency — so the people who actually run factories move the scores. It can also unlock early access to premium features.

Earned, not purchased
Higher vote weight
Early access to premium

darkprint.io

Share the blueprint of your dark factory

The blueprint registry for autonomous AI factories. Autonomy you can read as a graph — publish your pipeline, get it scored, and pull proven parts from the registry.

Explore the gallery Share a blueprint