GOVERNATLAS / PRODUCT DOSSIER

Citadel Lens

Citadel AI describes Lens for automated and human evaluation, custom metrics and red teaming of generative AI and machine-learning systems.

Explore the product ↗Add to comparison ↗

01 / CONTEXT & SCOPE

Understand the proposition.
Define the boundary.

Where it fits

Compare automated metrics with the intended application behavior and a documented human review process.

The decision to make

Choose the evaluation target and language before adopting a built-in metric. Lens, Lens Governance and Radar perform related but different jobs. A pre-deployment report does not show a runtime restriction was enforced. Ask how a human annotation is reconciled with an automated judgment and how that change remains visible in a later report.

What this record does not establish

Selected Lens documentation establishes described evaluation workflows, not metric accuracy, ISO conformity, legal compliance or feature parity with Radar. Named case studies are vendor material, not directory findings.

1 current source · 4 published assertions
Source coverage is a research boundary, not a quality score.

02 / YOUR EVALUATION

Turn interest into a useful trial.

Open the evaluation playbook ↗

A test worth running

Use a synthetic support dataset with an answerable question, an unsupported answer and an ambiguous case in the intended language. Compare automated judgments with reviewer labels, amend one rubric and retain the earlier run. Treat any live control demonstration as a separate authorized exercise.

GovernAtlas editorial proposal. We have not performed this test.

Evidence to bring back

A target/language/dataset/metric manifest and human-review protocol.
Original and amended judgments with a retained rubric and report history.

THE OPERATING DETAIL

A useful metric needs a case-level explanation.

GovernAtlas analysis: practical review considerations, not findings from a product test.

Separate the evaluation dimensions

Use a fictional retrieval case whose answer is relevant but unsupported by the supplied source. Review consistency and relevance separately. Add a harmless refusal and an adversarial mock input. Inspect item-level results so an aggregate cannot hide the difference between a helpful-looking answer and an evidence-backed one.

Keep reviewer disagreement

Have two authorized reviewers annotate the same ambiguous response under a written rubric. Preserve their rationales alongside the automated judgment. Revise the rubric once and retain the earlier interpretation. Agreement is useful context, but it should not erase a legitimate unresolved disagreement about the intended behavior.

Bound the visual model trial

For an image-model demonstration, select a disposable fixture and document the permitted perturbations. Inspect one transformed input, its prediction and the corresponding error segment. Ask which real operating conditions those transformations represent. Synthetic robustness checks and high-error segments require contextual interpretation rather than a blanket safety conclusion.

03 / DOCUMENTED EVIDENCE

Read the claim. Inspect the basis.

Each statement retains its classification and preserved source. A vendor assertion is not an independent test result.

vendor asserted

Its ML evaluation scope includes image perturbation robustness testing and identifying data segments with high error rates; these are vendor-described functions.

Source, dates & preserved provenance

Citadel Lens — primary scope and current identity ↗

Authority: vendor owned
Observed 2026-10-16 · Reviewed 2026-10-16 · current until 2026-11-15

Artifact SHA-256: ff77bd27e63b0306ecd2b6a915d8c09d13388da2ff99274be5f5f6336c117c6f

The hash identifies the preserved capture. The live source may have changed.

vendor asserted

The vendor describes prompt-defined custom metrics and human annotation, with views of agreement and disagreement between automated and human evaluations.

Source, dates & preserved provenance

Citadel Lens — primary scope and current identity ↗

Authority: vendor owned
Observed 2026-10-16 · Reviewed 2026-10-16 · current until 2026-11-15

Artifact SHA-256: ff77bd27e63b0306ecd2b6a915d8c09d13388da2ff99274be5f5f6336c117c6f

The hash identifies the preserved capture. The live source may have changed.

vendor asserted

Citadel Lens describes built-in generative AI metrics for source consistency, answer relevance, toxicity and jailbreak detection.

Source, dates & preserved provenance

Citadel Lens — primary scope and current identity ↗

Authority: vendor owned
Observed 2026-10-16 · Reviewed 2026-10-16 · current until 2026-11-15

Artifact SHA-256: ff77bd27e63b0306ecd2b6a915d8c09d13388da2ff99274be5f5f6336c117c6f

The hash identifies the preserved capture. The live source may have changed.

vendor asserted

Citadel AI describes Lens for automated and human evaluation, custom metrics and red teaming of generative AI and machine-learning systems. Scope: primary documentation only; no independent model test, fairness determination, causal finding or compliance approval.

Source, dates & preserved provenance

Citadel Lens — primary scope and current identity ↗

Authority: vendor owned
Observed 2026-10-16 · Reviewed 2026-10-16 · current until 2026-11-15

Artifact SHA-256: ff77bd27e63b0306ecd2b6a915d8c09d13388da2ff99274be5f5f6336c117c6f

The hash identifies the preserved capture. The live source may have changed.

Source register

Citadel Lens — primary scope and current identity ↗vendor owned · captured 2026-10-16

Brand asset provenance

Organization identity asset from its official website; inclusion does not imply endorsement.

Official website ↗ · Original asset ↗

04 / BEFORE YOU COMMIT

Better questions.
A clearer decision.

Resolve these with the provider and your own evaluation team.

Can you provide a target/language/dataset/metric manifest and human-review protocol for the selected configuration?

Can you provide original and amended judgments with a retained rubric and report history for the selected configuration?

Which exact version, entitlements, data paths and responsible roles apply?

Regulatory applicability requires a separate, current determination. Inclusion is not certification.

COMPLIANCE / CLAIMS, SCOPE & GAPS

Start with the requirement.
Inspect the support.

Prepare your vendor review ↗

These labels describe this dossier’s current evidence coverage. A published mention may be a framework mapping, a qualified assertion or a limitation; it is not a compliance finding.