Human in the loop by design
HITL is not a checkbox at the end of a demo. Cases, verdicts, and approvals are how Caliptra keeps speed from becoming silent authority.
“Human in the loop” is easy to say and easy to fake. A screenshot with an Approve button is not a design. A design is a set of constraints that make it hard for the system to do consequential things without a person noticing.
Caliptra’s product promise is blunt: cut the noise, keep the human in the loop. The second half is not a softener for the first. It is the integrity check on everything else.
Cases are the unit of work
Alerts arrive as events. Humans reason in cases. A case is where verdict, evidence, confidence, and recommended next steps live together so someone can take responsibility for an outcome.
That sounds obvious until you watch systems that skip it: chat transcripts that never become a record, enrichment that lives only in a model context window, or auto-actions that leave no durable trail. HITL without a case is theater. Someone clicked once, and the system forgot why.
In Caliptra, Cases and Analysis share the same evidence spine. The point is not decorative UI. It is so a human gate has something concrete to approve against.
Verdicts before actions
A typed verdict is a claim about the world: noise, needs review, real threat, or another agreed shape. An action is a change to the world: isolate a host, disable an account, push a block, close a ticket as benign.
Those are different risk classes. Models can help form verdicts and draft recommendations. They should not leap from “sounds like phishing” to “we already reset the user” without a gate that matches the risk.
Human-in-the-loop by design means the product path has a place where verdict and action are separated, and where the action cannot run until someone accepts the cost of being wrong.
What the gate is for
The gate is not there to slow every click. It is there for consequential moves: the ones that are hard to undo, that affect customers, or that can create outages of their own. Lean teams and MSSPs need speed on triage; they also need a clear line between “the system prepared this” and “a person authorized this.”
That line also protects the models. When every risky step requires approval, you can improve first-pass clearing and deep investigation without turning the product into unsupervised automation. Corrections feed the loop. Authority stays with the team.
Trust without invented proof
We will not dress this up with fake before/after metrics. Trust in HITL comes from inspectability: can you see the verdict, the evidence, and the approval? Can you tell what was model-suggested versus human-confirmed? Can a later investigator reconstruct the trail?
If those answers are yes, the human is actually in the loop. If they are no, you have a demo button.
Caliptra is building for the first standard. Attention stays scarce. Risky actions stay gated. The loop stays human, not as marketing language, but as the shape of the product.