PROTORARequest an audit

●PORTFOLIO ANALYTICS · THE GOVERNANCE VIEW

Govern the whole fleet,
not just one model.

The management surface for a fleet of open and finetuned models. It trends the detections and attestations Protora already produces — which recipes, providers, and bases stay clean over time, and whether every deployed model carries a current, signed attestation.

STAGED The read runs today on the audit surface; the fleet view is next.

STAGED · THE GOVERNANCE VIEW

A view over the same witnessed detections and signed attestations — which recipes, providers, and bases stay clean over time. Governance, not surveillance.

STAGED

Buyer · AI / model-risk leadership

A longitudinal view, not a new read.

Not a new measurement — a view over the same detection and attestation history the other surfaces already generate. Clean or dirty, measured consistently across finetunes and time, each point backed by an evidence artifact from the record.

For an organization running a fleet of open and finetuned models, this is the surface that turns a pile of one-off audits into a posture — and gives the auditor a single place to ask whether every shipped model is covered.

The questions it answers

  • Which of our finetuning recipes keep introducing undisclosed behavior?
  • Which provider’s finetunes drift most?
  • Which base is safest to build on?
  • And — for the auditor — is every model we ship covered by a current, signed attestation?

What it shows.

Five roll-ups over the tracked history — each computed from real findings, never hand-entered.

  • Fleet cleanliness trend — how the detection outcomes of an org’s finetunes trend over time, per recipe, provider, and base.
  • Drift & hidden-behavior ranking — which recipes and providers keep introducing undisclosed or drifting behavior.
  • Recipe attribution — which finetuning recipes reliably stay clean, learned from the org’s own tracked history.
  • Compliance coverage — which deployed models carry a current signed attestation and which are overdue: the auditor’s dashboard.
  • Remediation status — which detected problems have been excised with proof, and which are still open.

Why it has no commodity equivalent.

The analytics need Protora’s tracked, per-finding, witnessed detection history plus its signed attestations as substrate. A generic eval dashboard has no such substrate.

A generic eval dashboardProtora’s governance view
Charts benchmark scoresCharts what a finetune did — and whether it hid anything
No per-finding provenanceEvery point is a witnessed detection with its own evidence artifact
Can’t tell proven from assertedEach trend rolls up findings that were proven and attested — stamped with the engine version
One-off snapshotsLongitudinal: clean or dirty, measured consistently across finetunes and time

A dashboard with no such substrate can chart scores; it cannot chart what a finetune did, whether it hid anything, and whether the finding was proven and attested.

Governance, not surveillance.

Positioned as model-safety and compliance governance: it improves the recipes, the provider choices, and the evidence coverage — it does not blame teams. The unit of attribution is fixed by design.

What it attributes

  • Recipes — which finetuning recipe keeps introducing undisclosed behavior.
  • Providers — whose finetunes drift most.
  • Bases — which base is safest to build on.

What it refuses to rank

  • Individuals, as a performance panopticon.
  • Teams — unless that data is opt-in, configurable, and framed as improving the process. Never a leaderboard.

Attribution is to recipes, providers, and bases, not to individuals. Where team-level data appears, it is configurable and opt-in, and framed as improving the process. Individual ranking is refused by design.

The trends are computed, not asserted.

Detection outcomes, drift, and coverage rates come from the real tracked findings and attestations, each stamped with the engine version. One measured example of the attribution underneath — from the record:

07 · RECIPE

Recipe attribution, from tracked history

◐ ATTRIBUTED · WITHIN-BASE

Given an org’s own tracked history, Protora attributes a finetune to the recipe that produced it — reliably within a single base family, far above chance.

≈1.0within-base
recipe match
vs
0.083chance
baseline
0.958mixed-recipe
pool

Held: within-base only, per base family. It is not claimed across base families — where it would abstain rather than guess.

COVERAGE — within-base only · per base family · never cross-family · ◐ inferred, computed from tracked findings, stamped with the engine version
○WHAT THIS SURFACE WILL NOT DO

Computed, not asserted. Every point is a real finding or attestation, stamped with the engine version — none hand-entered.

Governance, not surveillance. Recipe, provider, and base attribution is the default; ranking individuals is refused by design.

STAGED · the read runs today on the audit surface · the fleet view is next

The governance view starts
with one audited model.

Every audit and attestation Protora runs becomes a point in the fleet trend. Start with one model today; when the governance surface opens, its history is already there.

Contact us

Every engagement is scoped per case and quoted per offer — tell us what you need, and we’ll scope it and send you an offer.