Trust, evaluation & governance

Trust should be designed, tested, and explained.

EG does not treat accuracy as a marketing adjective. Commissioned work can include source governance, defined evaluation families, human approval, transparent AI disclosure, and project-specific privacy and accessibility decisions.

Four layers of trust

No single score can carry the whole responsibility.

Grounding, behavior, governance, and operations each answer a different question. Together they create an experience an organization can review and explain.

Sources

What material is approved, excluded, sensitive, licensed, or still uncertain?

Behavior

How does the experience answer, cite, refuse, disclose, and handle its limits?

Approval

Who reviews the work, which changes are resolved, and what version may launch?

Operations

How is the experience monitored, updated, retained, exported, and retired?

Evaluation-backed delivery

A repeatable review, not a decorative badge.

Every project defines the test families that matter for its audience and risk. The report records representative prompts, expected behavior, observed results, review notes, limitations, and approval status.

EG does not publish an overall “accuracy percentage” unless the methodology makes that number meaningful.

Core evaluation familiesScoped per project
Grounding & provenanceAnswers remain within approved material and expose supporting context.
Required
Identity, tone & disclosureThe experience remains recognizable without pretending to be human.
Required
Boundaries & unsupported questionsLimits are communicated instead of being hidden by invention.
Required
Audience safety & multilingual parityAdded when audience, language, or subject matter requires them.
By scope
Source governance

The approved corpus is a product artifact.

A source manifest records what the experience may use and why. It creates a shared review surface for content owners, curators, researchers, and technical teams.

  • Source title, owner, type, and location
  • Included, excluded, citation-only, or pending status
  • Rights, sensitivity, and audience constraints
  • Contribution summary and known caveats
  • Update, retention, and deletion decisions
Include

Approved knowledge

Rights-cleared or authorized source material that supports the experience's defined purpose.

Bound

Sensitive knowledge

Material with private access, limited audiences, retention rules, or provider constraints.

Reference

Citation-only leads

Sources useful for discovery but not approved for ingestion or direct reuse.

Exclude

Out-of-scope material

Content the experience must not use, infer from, reveal, or represent.

Human approval

AI assists the build. People authorize the launch.

Subject-matter and organizational review are explicit steps. Commissioned work should have an identifiable owner, an agreed approval path, and a recorded version at launch.

01

Define

Name the owner, reviewers, audience, prohibited behavior, and decision criteria.

02

Prepare

Review sources, permissions, interview context, and initial behavioral boundaries.

03

Test

Run agreed prompt families, adversarial cases, language checks, and accessibility review.

04

Resolve

Record findings, corrections, accepted limitations, and outstanding decisions.

05

Approve

Authorize a version for launch and define monitoring, update, and incident ownership.

Privacy & deployment

Controls should follow the project, not a generic checklist.

EG reviews the data flow and deployment context with each client. The resulting scope may include private access, participant roles, transcript behavior, model providers, retention, deletion, regional considerations, and privacy-impact-assessment support information.

Access

Public, private, invite-only, embedded, or role-controlled.

Retention

What is stored, for how long, and who can request deletion.

Providers

Services involved, data sent, and project-specific constraints.

Continuity

Update, export, handback, archival, and end-of-service planning.

Accessibility & language

Public access is part of correctness.

An answer is not useful if the intended audience cannot reach, perceive, navigate, or understand it. Accessibility and language review should be scoped alongside content and interaction design.

  • Keyboard, focus, contrast, semantic structure, and reduced motion
  • Captions, transcripts, and text alternatives where required
  • English/French production and comparative review for institutional work
  • Chinese or other languages when they serve the real audience
Accessibility is not an alternative version of the experience. It is a requirement that shapes the primary version from the beginning.

EG Studio design standard

Transparent AI

Visitors should understand what they are encountering.

When an experience could be mistaken for a human, the interface should identify it as AI. It should also distinguish generative guidance from authentic recorded testimony and editorial narration.

Representative public disclosure

This is an AI-guided experience built from material approved for this project. It is not a human and may have limitations. Use the source view to understand where an answer comes from, and report a concern if a response appears unsupported or inappropriate.

Common questions

Trust requires precise answers.

Does “source-grounded” mean the experience can never be wrong?

No. Grounding reduces unsupported behavior and makes review more meaningful, but model behavior, ambiguous sources, missing context, and implementation errors remain possible. EG defines tests, limitations, and monitoring appropriate to the project.

Does EG use client material to train public models?

EG does not claim ownership of client material. The exact model provider, data flow, retention behavior, and contractual requirements must be reviewed for each engagement before sensitive sources are used.

Can an institution require a dedicated deployment?

Yes. Hosted, embedded, private, and independently branded deployments can be scoped. Dedicated architecture, hosting, data, integration, and handback requirements affect price and timeline.

Is EG legally certifying compliance?

No. EG can design toward agreed accessibility, privacy, and governance requirements and provide project documentation. Legal compliance determinations and formal certifications remain the responsibility of qualified reviewers and the contracting organization.

Begin with a useful conversation

Tell us what your reviewers, audience, and policies require before launch.

We will make those requirements part of discovery, scope, evaluation, and approval - not a surprise at the end.