Baseline model qualification
One corpus, a defined model pool, firm-specific scoring, and an initial recommendation set.
Model Assurance
Make model selection evidence-led.
ZTA independently benchmarks open-weight models against the institution’s actual use cases and standards, then maintains the evidence, governance, and evaluation framework required to determine which models remain fit for use as models and institutional requirements change.
Why it matters
Model assurance is a program, not a one-time report.
The initial baseline is only the beginning. Models change, the institution’s use cases evolve, and governance obligations grow more demanding as AI moves deeper into production work.
Prove model performance before production. Keep proving it after deployment.
Monthly assurance briefs
Developments in the model pool, the benchmark, or operating landscape that matter to the client.
Quarterly scorecards
Evidence-backed recommendations on which models remain approved, restricted, or removed.
Annual governance audit
Review whether the methodology, controls, and audit record remain fit for purpose.
Event-triggered re-runs
Out-of-cycle triage when a material release, new corpus, or new use case changes the decision surface.
What you get
A model shortlist with reasons
Not a ranking. Each candidate is assessed on extraction, analysis, and judgment separately, because a model that reads figures flawlessly may still be unfit for the work that matters.
A benchmark that stays yours
The corpus, answer key, scoring rules, and evaluation record remain in the client environment, ready to be re-run against future model candidates.
An audit trail built to be examined
Every score traces back to evidence, hashed and attested. A regulator or risk committee should be able to inspect a record, not rely on a recollection.
Problem
The wrong basis, correctly cited. A model is asked for GAAP gross margin and returns the non-GAAP figure found a few lines lower in the same release.
Why it matters
The number is in the source, the citation resolves, and nothing reads as obviously incorrect. The error propagates quietly into every multiple built on it.
Why validation matters
An unvalidated model does not announce these errors. It produces work that reads well and is wrong in ways that compound quietly across a coverage universe.
Sample baseline
Sample Model Assurance Baseline
A public example of the baseline work product: model-level findings, failure analysis, auditability, governance requirements, and deployment recommendations.
100Questions
7Models evaluated
700Scored responses
3Models admitted to Model Council
Independence
The evaluator must be independent of the judged.
The model that judges the work must not be the model that produced it, and should not share its training lineage. Correlated failure modes mean a model can mark its own blind spots as correct.
We also treat the evaluator itself as something to be verified rather than assumed. Assurance that has never found a fault in its own process has not looked hard enough.