Trust tiers
How to read a number in AgenticObjects — the A, B and C tiers, the marks that show who wrote a figure, and why provenance beats a confidence score
Every number AgenticObjects shows you carries a badge that says how much weight it can hold. Reading it takes two seconds, and it tells you what a percentage never could: where the figure came from.
The three tiers
A tier describes provenance — how a figure was produced. You meet the same three tiers everywhere; the guide panel uses shorter labels for them.
| Tier | On briefs and findings | In the guide panel | What it means | Can it appear in a brief? |
|---|---|---|---|---|
| A | Governed | A · Approved | A governed metric with an approved definition — the highest-trust number. | Yes |
| B | Derived | B · Composed | Built by combining approved metrics; sound, and derived rather than pre-approved. | Yes |
| C | Exploratory — cannot enter a brief | C · Candidate | An exploratory figure — read it as a lead, not a final number. | No |
The C label says it outright: an exploratory figure never enters a brief. A brief is a delivered, quotable document, so it carries only A and B figures. You can still meet a C figure in a chat answer you asked for, where you see the badge and decide for yourself.
Tip: Let the tier tell you what to do next. An A figure is ready to quote. A B figure is ready once you know which approved metrics it combines. A C figure is a good question to take to whoever owns the metric.
Where you see the tier
- In a brief, every finding card carries its tier badge, and the Declarations strip at the top states the tier of the brief as a whole. See Reading a brief.
- In chat, each answer carries a Trust badge with its tier. An answer takes the tier of the weakest metric it used, so a single exploratory input is never hidden behind a governed one. See Chat.
- On a finding's own page, the tier sits in the identity line beside the agent that produced it. See Finding detail.
The same figure carries the same tier wherever it appears — one decision, shown in several places.
Why tiers, not a confidence score
A single score that claims to sum up trustworthiness hides exactly what you would need to check it. AgenticObjects shows the components instead, each one visible on its own:
- Provenance — the tier: where the figure came from.
- Verification — whether the engine traced the figure back to the query results.
- Freshness — how current the data behind it is.
Each of these is something you can confirm. A label such as "87% confident" gives you nothing to confirm.
Who wrote the number
A language model writes the sentences around your numbers. AgenticObjects makes sure you can always tell which figures the engine vouches for.
Figures the engine wrote. Where the engine placed a number straight from the evidence, it is shown with a dotted underline:
Dotted figures were written by the engine from the evidence — hover for the source.
Hover or tap one to see the exact source, or the calculation the engine performed.
Figures the model wrote. When the model states a figure the engine could not trace to the query results — a percentage it worked out itself, for example — the number is marked with a dagger and the text ends with this legend:
† model statement — not verified by the engine
Hover or tap a marked number and it tells you why, for instance:
- Model statement — a percentage or ratio the model computed itself
- Model statement — a bare count with no source in the query results
- Model statement — stated without an address; the engine did not verify it
In chat, an answer that contains marked figures also shows how many beside its trust badge, for example 2 figures could not be verified. A † figure is not necessarily wrong; it is unproven. If it matters, ask again so the engine fetches it directly.
When a number is withheld
Where the engine cannot stand behind a figure at all, it leaves the figure out and puts a marker in its place. The marker explains itself:
A number was withheld here: it had no source in this answer's query results.
A withheld value tells you precisely where the evidence ends, so every figure that does appear is one you can trust.
The engine decides
The tier is stamped by the engine from the evidence of the run. The model that writes the prose can never set it, raise it or argue with it. The badge on a brief says so in the product's own words:
The engine stamps this from the evidence; the model can never write it.
So a badge is never the model's opinion of itself. It is a fact about the run, recorded where the query results are and mirrored on every screen that shows the figure.