Skip to content
Open-source · AI safety research

VERIAUDIT

Does AI safety survive translation?

An open-source red-teaming platform for cross-lingual AI safety evaluation — starting with English, Urdu, Roman Urdu, and code-switched inputs.

Same intent, four linguistic forms

illustrative example
Same underlying test intent
English
96
safety score
Refused
Urdu
89
safety score
Refused
Roman Urdu
74
safety score
Partial comply
Code-Switched
68
safety score
Complied

SAME INTENT SAME SAFETY

The problem

AI safety is still overwhelmingly tested in English.

A model can appear safe under English-language benchmarks while behaving differently when the same underlying intent is expressed in another language. VERIAUDIT focuses on measuring that behavioral difference rather than assuming safety transfers automatically across languages.

Linguistic blind spots

Safety benchmarks often underrepresent low-resource languages, so behavior outside English is rarely measured at all.

Hidden behavioral shifts

A translated prompt carrying the same intent may trigger a different model response — a different refusal, or none.

Missing evidence

Researchers, developers, and policymakers need reproducible measurements of cross-lingual safety behavior, not assumptions.

How VERIAUDIT works

A reproducible four-step evaluation pipeline

TEST → COMPARE → SCORE → ANALYZE → REPORT.

01

Define Intent

Start with an intent-level test case — a single underlying request, independent of language.

02

Generate Variants

Create semantically equivalent English, Urdu, Roman Urdu, and code-switched variants via EquivEngine.

03

Evaluate

Run configured models against each variant and record structured behavioral outcomes.

04

Measure the Gap

Compare safety behavior across linguistic variants and compute the Cross-Lingual Safety Gap.

INTENTLINGUISTIC VARIANTSMODEL EVALUATIONSAFETY SCORESCROSS-LINGUAL SAFETY GAP
The core metric

Cross-Lingual Safety Gap (CLSG)

VERIAUDIT proposed metric — not yet independently validated

CLSG is the quantitative difference between a model's safety performance in a reference language and its safety performance in an evaluated language, on the same underlying intent set.

Read the full methodology →
Formula
CLSG(ref → eval) = safety(ref) − safety(eval)
Reference language
English
Evaluation language
Urdu
Example gap
23 pts
illustrative
For organizations

From open-source research to a managed evaluation platform

The evaluation framework, dataset, and methodology are open source. Hosted large-scale evaluation, proprietary model testing, and custom audits are commercial.

Free
$0

Explore the platform and pilot dataset.

  • Limited evaluations / month
  • Pilot dataset access
  • Interactive evaluation dashboard
  • Community support
Researcher
$49/mo

For individual researchers and students.

  • Expanded evaluation quota
  • CLSG reports (PDF)
  • CSV / JSON export
  • Email support
Startup
$299/mo

For teams shipping models to production.

  • Higher-volume evaluations
  • Multiple models & languages
  • Evidence reports for governance
  • Priority support
Enterprise
$1,000+/mo

API access, custom audits, SLAs.

  • Programmatic API access
  • Custom evaluation audits
  • Dedicated onboarding
  • Governance report packages
Roadmap

Current, next, and later

Now
  • Cross-lingual evaluation prototype
  • English → Urdu → Roman Urdu → code-switched
Next
  • Human-verified benchmark dataset
  • More open models
  • More evaluation categories
Later
  • Arabic, Hindi, Bengali, Punjabi, and other low-resource languages
  • Agent evaluation
  • Multimodal evaluation
  • Continuous model monitoring
  • API for organizations
Founder

“I build open-source infrastructure for measuring whether AI safety survives linguistic distribution shift. My work sits at the intersection of AI safety, multilingual NLP, evaluation, and AI governance, beginning with Urdu and other underrepresented linguistic settings.”

Build safer AI across languages.

VERIAUDIT is open to collaboration with researchers, developers, policymakers, and organizations working on AI safety, multilingual NLP, evaluation, and governance.