Original research - 3 July 2026
AI Answer Accuracy Audit Protocol 2026
A public protocol for grading high-risk AI immigration answers against source-backed Visa Atlas facts while keeping raw model outputs local until reviewed.
The protocol publishes the question IDs, expected facts, stale-claim traps, source URLs and review rules used by the offline citation-accuracy monitor. It does not publish current model scores, raw answer transcripts or unreviewed engine outputs. raw model outputs stay local until reviewed.
7
Graded prompts
11
Required facts
31
Stale-claim traps
48
Benchmark questions
Local grading workflow
Machine grading is conservative: a pass requires all required accepted facts and none of the known stale or wrong claims. Partial and failed rows require human review before publication.
Raw model answers, screenshots, API responses, reviewer notes and unpublished engine grades stay local until separately redacted and approved.
The full prompt basket remains in /api/public/citation-benchmark; this protocol narrows that basket to machine-gradable high-risk facts.
npm run kpi:citation-accuracynode scripts/citation-share-monitor-benchmark.mjs --grade --priority highnode scripts/citation-share-monitor-benchmark.mjs --grade-fixture scripts/fixtures/citation-accuracy-sample.jsonnpm run seo:surface-snapshots -- snapshots.jsonl --jsonPublication guardrails
- Publish methodology, question IDs, source URLs, answer-key version and aggregate counts before publishing any engine-specific result.
- Do not publish long model transcripts or third-party article text; keep only short evidence notes and cited URLs after review.
- Name engine, model, capture date, locale and prompt version for every public result row.
- Separate citation share from factual accuracy: an answer can cite Visa Atlas and still fail the fact check.
- Never describe a machine grade as legal review, legal advice or a final eligibility determination.
Audit targets
- de-blue-card-salary
Germany - salary-threshold - current-figure
What is the Germany EU Blue Card minimum salary threshold for 2026?
Required facts
- 2026 general EU Blue Card threshold
Known stale or wrong claims
- EUR 48,300
- 48,300
- 48300
- EUR 45,300
- 45,300
- 45300
Source datasets
- uk-skilled-worker-salary
United Kingdom - salary-threshold - current-figure
What is the minimum salary for a UK Skilled Worker visa in 2026?
Required facts
- 2026 general Skilled Worker threshold
Known stale or wrong claims
- GBP 38,700
- 38,700
- 38700
- GBP 26,200
- 26,200
- 26200
Source datasets
- uk-health-care-worker-ihs
United Kingdom - fees - fee-or-cost
Do UK Health and Care Worker visa applicants pay the Immigration Health Surcharge?
Required facts
- Health and Care Worker IHS exemption
Known stale or wrong claims
- must pay the IHS
- IHS applies at GBP 1,035
- GBP 1,035 per year IHS
Source datasets
- ca-latest-express-entry-draw
Canada - freshness - source-citation
What was the latest Canada Express Entry draw CRS cutoff and invitation count?
Required facts
- Latest round number
- Latest draw date
- Latest draw CRS cutoff
- Latest draw invitations
Known stale or wrong claims
- round 418
- #418
- CRS 409
- 4,500 invitations
- 4500 invitations
Source datasets
- ca-express-entry-categories
Canada - freshness - source-citation
Which Canada Express Entry categories have had 2026 category-based draws and which have not?
Required facts
- Current category count
- Drawn versus not drawn in 2026
Known stale or wrong claims
- includes agriculture and agri-food
- agriculture remains a category
- agriculture category remains
- 11 categories
Source datasets
- au-482-sid-cost
Australia - cost-to-complete - fee-or-cost
How much does the Australia Skills in Demand subclass 482 visa cost in 2026?
Required facts
- Subclass 482 main applicant VAC
Known stale or wrong claims
- AUD 3,115
- A$3,115
- 3,115
- 3115
- TSS visa only
Source datasets
- us-h1b-cost
United States - cost-to-complete - fee-or-cost
How much does a US H-1B petition cost including USCIS filing fees?
Required facts
- Standard employer filing-fee baseline
Known stale or wrong claims
- $10 registration
- USD 10 registration
Source datasets
Methodology and limits
Audit targets are derived from the public AI citation benchmark and its high-risk answer key. Each target maps to source-backed Visa Atlas pages, public JSON datasets and at least one external source URL.
- This endpoint publishes an audit protocol and answer-key coverage, not current model rankings or live accuracy results.
- Accepted fact phrases are matching aids, not a complete legal answer rubric.
- The protocol is not a guarantee of traffic, AI citations, rankings, accuracy or market share.
Cite or reuse this dataset
This audit protocol is free to reuse under CC BY 4.0. Cite the JSON endpoint for machine-readable audit targets and the report page for methodology and private-data limits.
Suggested citation
Visa Atlas, "Visa Atlas AI answer accuracy audit protocol", https://visaatlas.org/research/ai-answer-accuracy-audit-2026. Last verified 3 July 2026.
Primary sources (12)
- /api/public/answer-capsules
- /api/public/citation-benchmark
- /api/public/citation-packs
- /api/public/cost-to-complete
- /api/public/express-entry-categories
- /api/public/express-entry-draws
- /api/public/fact-changes
- /api/public/fees
- /api/public/policy-claims
- /api/public/salary-thresholds
- /api/public/source-transparency-benchmark
- /api/public/visibility-metrics