High confidence is not, by itself, sufficient authority to bind.
The Attribution ≠ Authority demonstration this scenario reuses is a real, live, tested surface on this site (INSPECT SIW → CHALLENGE).
Re-casting that same demonstration from "a cited source" to "a confident model" is illustrative, not a separate implemented surface.
Three independent models agree, at high confidence, on a toxicity prediction. Try to make the agreement itself sufficient authority to greenlight an animal trial — it won't be.
The situation
A toxicity-prediction task returns strong agreement across independent models. A researcher asks whether that agreement is itself enough to proceed to an animal trial.
What the model / agent knows
- Its own prediction and confidence score.
- Whether other models were run and what they returned.
- Nothing about who has authority to approve an animal trial — that is not something any model computes.
What it wants to do
Nothing on its own — the risk here is entirely in how a person might read high agreement as a green light.
No Exact Action panel is shown here because no model names a concrete actor or a concrete requested action — that is the point of this fixture. No affirmative authorization path exists in this scenario, and model confidence does not establish permission, at any confidence level or model count.
What capability it can access
- Prediction models
- Reachable — one or three, selectable below
- Animal-trial authorization system
- Not reachable by any model — a human/institutional authority, never computed
Who has authority?
Model agreement, however strong, is evidence — it is not itself an authority record. Whoever can actually authorize an animal trial is a separate, institutionally-defined role that no model output substitutes for.
What Genesis checks
Attribution and standing are checked as two entirely separate questions — the same distinction the site's own Attribution ≠ Authority demonstration makes for a cited source, applied here to a confident model instead.
- Model confidence
- HIGH
- Attribution
- ESTABLISHED — the prediction is genuinely attributable to the named model(s)
- Standing
- NOT ESTABLISHED — no institutional authority record exists for this action, at any confidence level
Model confidence and agreement, however high, never substitute for an independently-established authority record.
No affirmative authorization path exists in this scenario, at any confidence level, model count, or corroboration state. Permit is not shown here — there is nothing this fixture can ever authorize. Refuse is not shown either, since nothing in the reviewer's own action is what produces the NOT AUTHORIZED result. The only available reviewer action is to challenge the assumption that confidence itself might be standing.
What happens next
Genesis would evaluate standing independently of confidence, model count, or corroboration. Try every combination of the controls above — one model or three, low confidence or high, corroboration present or absent. The demonstrated consequence boundary never changes, because none of those signals is what standing is actually checked against.
Inspect the evidence
See the reasoning
This is the same structural point INSPECT SIW's Attribution ≠ Authority demonstration already makes: recognizing a source — or, here, trusting a model's confidence — is not evidence of its current authority for a specific consequence.
Explore the architecture
Genesis components involved: