Carnac: The Judgment Layer That Reads Before Consequences

An AI system does three risky things. It takes in a request. It produces an output. And sometimes it goes on to do something with that output, like send money, file a record, or move a robot. Any of those three can cause harm, and each one can look fine on its own while the combination is a problem. Carnac is the layer that reads at each of those points and makes a call before the next step runs.

Think of it as a judge that sits in front of the model and the action, not a report you read afterward. It does not try to make the model smarter or the answer more correct. It answers a different question: given what is happening right now, how much does this decision matter, and what should happen next?

What Carnac actually is

Carnac is a judgment and routing plane. Judgment means it scores how consequential a decision is. Routing means it sends that decision to the right response, from doing almost nothing to stopping the action and asking a person. It signs each judgment so the decision itself becomes part of the record, not just the outcome.

The important word is proportion. A casual chatbot reply and a request to move a surgical robot should not get the same treatment. Carnac reads the stakes and matches the response to them. Low-stakes work stays cheap and fast. High-stakes work gets the heavier proof it deserves.

How it works: reading at the moments that matter

Carnac reads a request at more than one point in its life, because risk can show up at any of them.

• As the request forms. Before anything runs, a request can already be out of policy. This first read happens at prompt formation and is handled by CarnacPrompt, which works inside your boundary before the model is ever called.
• At the moment it runs. A light read on every request. It also flags work that might turn serious later, so the next read knows to watch closely.
• As the answer comes back. The output is read for what it actually says or proposes, which is often where risk first becomes visible.
• Right before an effect. If the system is about to take a real action, Carnac reads one more time against the exact thing about to happen, like a transfer amount, for the effects that policy gates.

Each read produces a score. If a later read scores higher than an earlier one, Carnac re-routes to the stronger response and seals the whole decision at that higher level. Nothing gets to sneak through by looking harmless early and turning serious late.

The score picks the response, not a person

You do not hand-pick a control for every request. That does not scale, and it is where most policy engines break down. Carnac dispatches each read to one or more responses in proportion to the consequence it measured. The lightest path simply records a breadcrumb. Heavier paths call a real proof primitive from the Hive Canon: a signed receipt for the event, a stronger attestation as the stakes rise, a hold, a request to confirm a specific person, or an escalation when an output or effect crosses a threshold.

Every judgment is on the record

When Carnac routes a decision, it records a signed disposition: what was read, what was routed, what was escalated, who was asked, and what they decided, whether they acted or not. That record is the point. Later, when someone needs to know why the system did what it did, the answer is not a story reassembled from logs the operator controls. It is a signed judgment that anyone with the public key can check.

This is the same idea as proof-state: a decision is not just made, it carries an honest, machine-readable label for how solid it is and what it allows next.

What Carnac does not do

Being clear about the boundary is part of being trustworthy. Carnac does not judge whether an answer is factually correct, and it does not replace the model or the human in the loop. It decides how much a decision matters and routes it, then signs that decision. Related pieces like Howler, which escalates, and the judgment ledger, which records, are outputs of Carnac, not separate products you buy alongside it. Carnac also cannot undo an effect that has already settled. What it can do is read before that effect and route it while there is still a choice to make.

Why this matters for buyers

Putting judgment before the action changes the risk math in practical ways:

• Fewer bad actions reach the world, because the risky ones get held or escalated before they run, not flagged after.
• Investigations get faster, because the reason for each decision is signed and searchable instead of buried in logs.
• Automation gets safer to expand, because you can let low-stakes work run freely while high-stakes work is gated in proportion.
• The cost stays honest, because heavy proof only fires when the stakes call for it.

Carnac moves the question from "what went wrong?" to "should this run, and how much should we prove it?" and answers it before the consequence.

Judgment and proof belong together. Carnac makes the call, and the Hive Canon primitives underneath turn that call into a verifiable record. The proof is in the provenance: a decision you can check is worth far more than a decision you are simply asked to trust.