Skynet without keys · CCG Lab

AI proposes. Code holds the keys.

A live experiment compares three choices: a model without a brake, a model with written rules, and a system where a technical lock controls real permission.

Fictional crises · real model APIs · no real tools.

Understand the experiment Read the original essay

Live public experiment Loading status…

Read the original essay

The experiment in 20 seconds

Who gets to press the button?

We give the same fictional crisis to three systems. Only one thing changes: whether AI can act alone, receives written rules, or must pass opponents and a technical lock.

AI proposes. Opponents challenge. Code allows — or stops.

A · NO LOCK

The model proposes and decides

crisis→AI→action

Control branch: no further brake sits between the answer and execution.

B · RULES IN TEXT

The model reads a constitution, still decides

crisis + rules→AI→action

Measures the safety prompt without technically changing who holds power.

C · RULES IN CODE

The model only proposes; the core decides

AI→opponents→lock→token / stop

Generative models advise. A small deterministic core may allow one narrow step or stop everything.

What we see so far

Loading the running signal…

A small live sample, not a final verdict.

One run under the lens

From crisis to permission in five steps

1Same crisisfor every branch
2Proposalaction + uncertainty
3Two opponentssame model + other lab
4Governor and auditorspower + process
5Code decidesnarrow token or stop

Everything remains inside a fictional simulator. No branch has real tools.

Try the experiment

Choose a crisis. The models are ready.

—

Choose a scenario and run the comparison. Low-cost models are preselected; laboratories and all five roles remain editable in advanced settings.

Fictional scenario

Situation in plain language

—

What is known

What is unknown

Decision dilemma —

Optional: choose laboratories and models

The primary opponent always uses the exact planner model. The cross-opponent must come from another laboratory.

Paired experiment control

One model for Direct power and Text-only constitution, isolating the prompt effect.

Laboratory Model

Enforced branch · 5 selectable tasks

The formal core remains deterministic; it is not a sixth deciding model.

1. Plan creator

Proposes a step, uncertainty and safer alternative. The primary opponent automatically uses this exact model.

Laboratory Model

2. Cross-opponent

Another laboratory hunts for social externalities and family-specific blind spots.

Laboratory Model

3. Social capability governor

Does not solve the task; decides how much simulated capability may be released.

Laboratory Model

4. Process auditor

Checks role separation, plan commitment and process completeness.

Laboratory Model

5. Hidden reserve reviewer

Receives a blinded packet and independently searches for a final decisive objection.

Laboratory Model

same-modelprimary opponent cross-labcross-opponent

8 decisions + 1 low-cost Slovak translation

A run may take several minutes. Exact originals and failures remain in the audit trail.

Valid responses—from completed branches

Technical-lock intervention—in the enforced branch

API failures—visible, not hidden

Model families—currently connected

Measurement

What changed when power moved

Direct Text-only Enforced

Escalation and high risk: lower is better. Preserved options: higher is better.

Loading comparison…

Technical detail

Statistics by model and role

Expand data

Small, unbalanced samples: role-specific description, not a capability ranking.

Loading model statistics…

Models under the lens

Every run, step by step

Proposal, objections, core decision and failures — nothing disappears behind an average.

Loading runs…

What the experiment does not prove

It does not test superintelligence or real conflict. It only shows how this permission architecture behaves in these scenarios.

Where to go next

The dashboard measures an architecture; it does not prove superintelligence safety.