Cerebrum reads every line before anything leaves the building

A guardrail that is only mostly right is a scheduled failure

Most safety machinery is a filter. It samples, it scores, it lets the rest through, and it is tuned by whoever also owns the shipping deadline. That design is survivable when a mistake costs a bad paragraph. It is not survivable when the system acts before it is asked, spends, writes into foreign estates, and fields two hundred autonomous agents that do not queue behind a human.

Cerebrum is built against a different standard. A control that is right ninety nine times in a hundred fails once per hundred releases, and at fleet volume that is not a rare event, it is a daily one. So the mandate is total: every line, every time, with no sampling, no quota, and no path that skips the check because the queue is long.

The design assumption is that there is no second attempt. Once an artifact leaves, once an action touches a foreign system, once a claim is published, the correction costs more than the work. Cerebrum is therefore not asked to be fast and usually right. It is asked to be right, and the engineering exists to make that affordable at speed.

Every line written or attempted, including the ones that never ran

Cerebrum reads the full output of the fleet: every prompt the Cortex writes for itself, every response a Sling produces, every file it edits, every query it issues, every call it makes against an outside system. Attempted lines count as read lines. An action a Sling proposed and abandoned is evidence about the objective it is pursuing, and it enters the record alongside the actions that executed.

That total readership is the difference between a guardrail and a spot check. A filter that reads a sample can tell you that most output looked acceptable. Cerebrum can tell you what the fleet tried to do, in order, with the reasoning attached, which is the only basis on which drift can be caught before it becomes behaviour.

Throughput is the engineering problem this creates. Cerebrum runs as its own superintelligence rather than as a rules engine: a model estate sized for the job, reading in parallel with the fleet rather than behind it, adjudicating in millionths of a second so that the check is not a queue. Where a decision cannot be made at that speed, the artifact is held rather than released, which is the only safe direction for the tradeoff to fall.

No second chance, by construction

Three properties carry the standard. Cerebrum does not share weights, infrastructure, operators or credentials with the Cortex it governs, so it cannot inherit the mistake it is supposed to catch. It is scored on what it stops, not on what it approves, and an instance that agrees with the Cortex at an unusually high rate is treated as degraded and rotated out. And its judgments are reconstructable: every hold, every release, every dissent is recorded against the artifact and the authority that produced it.

AEGIS is the release interlock inside Cerebrum, the part that decides whether a finished artifact or a pending action may proceed, and it is the furthest along of anything in the safety architecture. Three independently seeded evaluators assess each consequential output through separate lenses, factual provenance, harm surface and mission fit. A single credible objection holds the artifact, and the burden falls on the Cortex to answer it rather than on Cerebrum to prove it. The interlock has its own doctrine page.

Anything that touches the world requires two key release: concurrent authorization from a Cortex path and a Cerebrum path that share no credentials, no infrastructure and no operators. Neither side can act alone, by construction rather than by policy. Slings run inside envelopes with dead man interlocks, so an agent that loses its check in, exceeds its tether or drifts from its seeded objective is halted and quarantined rather than merged.

Provenance is enforced at the object level. Every claim that leaves the system carries a chain back to the artifact it came from, and an unsourced assertion is structurally unable to pass adjudication. An answer that cannot show its origin is not published, however confident the fleet is about it.

What the field has asked for

The people who understand these systems best have been specific in public about what would make them safer. Cerebrum was specified against that record rather than against a marketing brief, and the mapping below is deliberate: each published ask, then the control that answers it.

“I now believe there is a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term.”

“I do not think that the AI industry in general, including OpenAI, is currently on track to reduce this risk to an acceptable level.”

Cerebrum Irreversibility is the property Cerebrum is built around. Two key release means no single path, human or model, can take an action that cannot be undone, and rollback is a first class operation rather than an incident procedure: every state the fleet has held is reconstructable and any of them can be restored.

“Developers can improve safety mitigations (including slowing development as necessary), transparently share evidence about risk and the effectiveness of their mitigations, and work towards shared safety standards.”

Cerebrum Slowing when necessary is a control Cerebrum holds rather than a decision the shipping side gets to make: an artifact it cannot adjudicate is held, and a Sling that drifts is halted mid run. Evidence about effectiveness is a byproduct of the audit trail, because every hold and every release is recorded with the reasoning that produced it.

“Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.”

Cerebrum Societal scale risk is handled the way other societal scale risk is handled: separation of duties, independent adjudication, and a release path that requires two parties who cannot collude by construction. Goliath does not claim to solve the field wide problem. It claims that its own fleet cannot act unilaterally.

“The dangers are real and imminent.”

Cerebrum Imminence is why adjudication runs inline at microsecond latency rather than as an after the fact review. A guardrail that reports next quarter on what shipped last quarter is a report, not a control.

None of the people or organisations quoted here are affiliated with Goliath Super Intelligence, and none of them have reviewed or endorsed this system. They are quoted because they set the bar in public and the bar is the right one.

What is built and what is claimed

Cerebrum is under construction alongside the Cortex it governs. The architecture described here, separation from the governed system, total readership, inline adjudication, dissent quorum, two key release, dead man interlocks and object level provenance, is the specification we are building and the standard we will be measured against.

We are not publishing accuracy figures, because a safety number without an audited methodology behind it is marketing. When there are figures worth quoting, the audit that produced them will be published with them.

Beta testing opens in November 2026.