Product
From changed agent to signed Exhibit.
Prevailing AI is release assurance software for financial-crime AI agents. Not a consulting report, not a single accuracy number, and not a runtime monitor.
Suite
Bench. Casebooks. Exhibit.
Three parts, one job: prove to an accountable approver that this version of the agent is safe to release under your procedure.
Bench
Runs a Casebook against one or more versions of your agent. Import outputs or connect a staging endpoint. Scores each case, diffs the versions, and surfaces regressions first.
- Staging endpoint or output import
- Deterministic checks plus calibrated jury
- Version metadata captured on every run
Casebooks
Expert-reviewed test packs where every expected outcome references a rule in a written procedure. Synthetic persons and outlets. Held-out subsets guard against overfitting.
- Adverse Media shipping first
- Screening, KYB and TM to follow
- Mapped to your SOP in days
Exhibit
A versioned report the approver can actually read: scope, case mix, failures with rule references, regressions, held-out results, limitations and a sign-off block.
- Consequential errors reported separately
- HTML and PDF, audit-logged
- Re-issued on every release
What we measure
Never a single accuracy number.
An Exhibit reports each error class separately, with the procedure rule every failure violates, so the approver sees the risk rather than an average.
Consequential
False clears
A subject who should have been escalated was cleared. The error a regulator asks about first.
Missed escalations
The agent noticed the signal and still did not route the case to a human under your procedure.
Inconvenient
Unnecessary escalations
Friction and analyst time. Worth tracking, never worth hiding a false clear behind.
Unsupported conclusions
A defensible outcome reached without the citation or evidence your procedure requires.
How it works
Six steps from change to sign-off.
The first two happen once per workflow. The rest happen on every release.
Connect
Upload agent outputs for our case inputs, or expose a staging endpoint. No production code change required to start.
Map
We map Casebook cases to your written procedure so every failure carries the rule it violates.
Grade
Panel grades each case: deterministic checks, a calibrated jury, and human review on disputed items.
Compare
Current version against proposed. Regressions flagged with the affected rule and the case that moved.
Sign the Exhibit
Your MLRO or head of financial crime reviews scoped findings and limitations, then signs.
Re-run
Model swap, prompt edit, retrieval change, rule update. Run again, compare again, issue the next Exhibit.
Questions
Straight answers.
The questions compliance leads and vendors ask first.
Is an Exhibit a certification?
No. An Exhibit is scoped evidence for your approval process. It states scope, limitations and what was not tested. We do not claim to be a notified body or a conformity assessor.
Do you need access to our code or production systems?
No. Most pilots start with exported agent outputs run against our case inputs. A staging endpoint comes later if useful. Production is never touched.
We already use an eval tool. Why add Prevailing AI?
Keep it. We can import its outputs. What we add is expert-labelled Casebook content, mapping to your written procedure, consequential-error separation and a report shaped for the compliance approver.
What exactly do you measure?
False clears and missed escalations, separately from unnecessary escalations and unsupported conclusions. Each failure references the procedure rule it violates.
Ready to put a release on the record?