Skip to main content

Core concepts

Five ideas cover almost everything. Read them once and the rest of the docs will make sense.

The governed check​

A governed check is the core action. You send a message; Paresium Clara reads what was really meant, checks that answering is safe against your policies, writes an answer, and has an independent panel review it before it goes out. If anything looks risky it stops and tells a person instead.

Every check ends with a decision: answer, ask (for clarification), escalate, or block.

The record​

Most AI tools give you an answer and nothing else. Paresium Clara gives you the answer and the proof: a permanent record of what was asked, which rules were checked, what went out, and which model version produced it. You can look it up any time, and nobody can edit it afterwards. This is what makes an answer defensible to an auditor or a regulator. See The Record and retention.

Separately, a generalized, stripped-down version of your organisation's data is, by default, eligible to be used to improve Clara's models. You can turn this off at any time. See Model development and your data.

The four offerings​

Paresium Clara does four jobs, from quick answers up to fully checked ones. You can start with the fast, cheap one and step up as you need more care. The two governed levels leave a Record; the two fast ones leave a lighter trail, and the list below says which you get.

  1. Quick answers with receipts. A managed model answers, with a price quoted first and a billable receipt after. You get the receipt and a router audit entry, not a Record.
  2. A fast safe-or-not check. Pass, caution, or stop on a message, with named reasons. This is the Compliance check. It tells you what it found; it does not stop the message going out, so your own code decides what to do with a warn or a block. It checks the message against a narrow built-in floor that every account gets, plus whatever rules you have given Clara. The parts that always run are fixed rules rather than a judgement call, and they run outside the model that wrote the message, so they are cheap enough to put in front of every message and they catch attempts to talk your systems out of checking at all. If you have uploaded no policies, only the floor runs, so a pass means the floor found nothing, not that the message is safe. See Running checks for what the floor covers and Teaching Clara for how to make the check yours. You get a decision_ref you can quote in a support ticket or join to your logs, not a Record.
  3. A plain-language briefing. What the customer really meant, the risks, and a clean instruction for the next step.
  4. The full checked answer. Checks, briefing, reviewed answer, and billing settled, in one run.

To call each one from your own software, see Running checks and governed runs and the API reference.

The independent panel​

Before a governed answer goes out, a panel of AI models from different companies reviews it. One dissenting vote blocks the answer. This is why a Paresium Clara answer is safer than a single model's: it has to survive review by models that do not share the same blind spots.

Model versions and tiers​

The version of Paresium Clara that answers you is recorded on every run. Versions ship on three channels that ladder from beta to production. Today only the Beta model is live; the record always tells you exactly which version produced an answer. See Model versions and channels.

Billing, in one line​

Hold the worst case, charge what actually ran, release the rest. Amounts are in credits, and during beta everything is test mode, so no real money moves. See Billing and credits.