AI can read the words. Can it understand the meaning?
Cihan AI Data combines African-context data, calibrated human evaluation, verified talent and benchmark infrastructure to help AI systems perform better on meaning-rich African discourse.
African Context. Human Meaning. Better AI.
Nigerian register benchmark — MIS™ score
- Gemini Flash78.6
- GPT-567.8
- Gemini Pro65.4
- Calibrated human analysts100.0
Results are specific to our Nigerian register/discourse benchmark (arXiv 2606.20255) and do not imply general model rankings.
The problem
The Register Gap
AI fails on context, not translation. Models can render the words of Nigerian English and Pidgin perfectly well — and still miss what the speaker meant. Register, not vocabulary, is where meaning lives.
Same sentence. Opposite meaning.
“You don try well well”
Sincere praise
A warm, public acknowledgement of real achievement. Positive sentiment, high sincerity, celebratory register.
A sentiment classifier trained without register awareness scores both as positive. In moderation, customer intelligence, and safety pipelines, that single error class compounds into systematically wrong decisions about African audiences.
Capabilities
Seven ways we support AI teams
Managed data programmes delivered by calibrated African analysts, governed by research methodology and measurable quality controls.
Managed Annotation & Data Curation
End-to-end annotation programmes run by trained, calibrated analysts against your schema or ours, with multi-pass review and adjudication built into the workflow.
- Client-defined schemas, taxonomies and labelling guidelines
- Multi-pass annotation with adjudication of disagreements
- MIF™ available as an option for meaning-rich African-context tasks
African Data Collection & Corpus Development
Purpose-built collection of African text, speech and conversational data, sourced with consent and documented provenance.
- Nigerian English, Nigerian Pidgin, Standard English and Nigerian code-mixed speech; other African languages expanding / scoped on request
- Task-targeted prompts, dialogue, transcription and metadata capture
- Documented sourcing, consent and provenance records
Human Model Evaluation & Red-Team Review
Calibrated human raters assess model outputs for meaning, cultural appropriateness, safety and failure modes in African contexts.
- Side-by-side preference, rubric scoring and free-form critique
- Adversarial and red-team probing for context-specific failure modes
- Rater calibration and agreement tracking across every panel
Benchmarking & Diagnostic Evaluation
Contamination-aware benchmark runs with per-dimension diagnostics, so you learn where a model breaks rather than a single headline number.
- MIS™ scoring on the Nigerian register/discourse benchmark
- Per-dimension error analysis and failure taxonomies
- Contamination-aware holdout design and rotation
Dataset QA, Adjudication & Gold-Set Creation
Independent audit of existing datasets, disagreement resolution and construction of defensible gold sets for training and evaluation.
- Label audit, error rate estimation and inter-annotator agreement
- Senior-reviewer adjudication of contested items
- Gold-set construction with documented decision rationale
Managed AI Talent Deployment
Standing panels of vetted African analysts assigned to your programme, managed by Cihan with quality reporting and continuity of expertise.
- Dedicated, NDA-bound panels with domain and language fit
- Cihan-managed onboarding, calibration and quality oversight
- Capacity that scales with your evaluation cadence
MIF™ is a proprietary framework we offer where meaning-rich African-context tasks call for it. It is optional — we also work to your own schemas, rubrics and evaluation task types.
See all seven capabilities →How quality works
Governed quality, not good intentions
Every programme runs through the same controls, and the measurements come back to you with the data.
- 01
Guideline design
Task definitions, edge cases and worked examples agreed with you before a single item is labelled.
- 02
Analyst calibration
Every contributor passes calibration items against a reference set before joining a live batch.
- 03
Multi-pass annotation
Independent passes on meaning-rich tasks, with blind review where the task design requires it.
- 04
Agreement measurement
Inter-annotator agreement tracked per batch and per dimension, reported to you, not hidden.
- 05
Senior adjudication
Disagreements escalate to senior reviewers who resolve and document the final decision.
- 06
Gold-set audit
Hidden gold items sampled through the batch to measure accuracy and trigger retraining when needed.
Human intelligence
Cihan AI Data Network™
A curated African network — recruited, trained, calibrated and certified before anyone touches client work. Contributors progress to analyst, reviewer and adjudicator as their measured quality earns it.
- Level 1
Contributor
Data collection, transcription, and simple validation or classification tasks.
- Level 2
Certified Analyst
Complex annotation and human model evaluation against calibrated rubrics.
- Level 3
Senior Reviewer
Quality assurance, disagreement resolution and gold-set support.
- Level 4
Project Lead / Adjudicator
Batch leadership, final adjudication and quality of delivery.
Research engine
Meaning Intelligence Lab™
Research & benchmark engine for CIHAN AI DATA — a Cihan research initiative.
The Lab develops the Meaning Interpretation Framework (MIF™) and Meaning Interpretation Score (MIS™), designs benchmarks and gold sets, calibrates analysts and runs contamination-aware evaluation. Commercial delivery inherits that methodology.
- 1Research
- 2Benchmark
- 3Calibrated human evaluation
- 4Client delivery
Proof points
Check the work yourself
Research paper
arXiv:2606.20255
Preprint: the register framework, benchmark construction and full results. Not peer-reviewed.
Press
TechEconomy coverage
Independent reporting on the register gap in frontier AI systems.
Credentials
Verify an analyst
Verify any CMIA credential issued through Cihan Digital Academy.
Benchmark results
| Gemini Flash | 78.6 |
| GPT-5 | 67.8 |
| Gemini Pro | 65.4 |
MIS™ scores on our Nigerian register benchmark, contamination-aware holdout. Benchmark-specific; not a general model ranking.
Why Cihan
Rigour you can audit
Published methodology
Our register framework and benchmark are documented in a preprint on arXiv (2606.20255) — inspect the method before you buy the labels.
Calibrated human analysts
Human performance demonstrated on our Nigerian register benchmark, by CMIA-certified analysts working to a shared rubric.
Verifiable credentials
Analyst credentials issued through Cihan Digital Academy are cryptographically signed and independently verifiable.
Multi-pass annotation
Multi-pass or triple-blind annotation where the task requires it, with adjudication and inter-annotator agreement reported per batch.
Contamination-aware benchmarks
Evaluation sets are withheld and rotated, so benchmark scores reflect capability rather than memorisation.
African-context expertise
Nigerian English, Nigerian Pidgin and code-mixed discourse handled natively by in-country analysts — not machine-translated.
Who we work with
Built for teams accountable for model behaviour
Frontier and applied AI labs
Teams evaluating model behaviour on African discourse before release, and sourcing training data that reflects it.
Enterprise AI and data teams
Banks, telcos and platforms deploying moderation, support automation and customer intelligence across African markets.
Research institutions
Groups building or auditing African-language corpora, benchmarks and evaluation methodology.
Public sector and development programmes
Language technology, civic information and safety initiatives that need documented, consented African data.
Engagements are delivered by Cihan-managed qualified contributors and certified analysts under NDA, not by an anonymous crowd.
Get started
Request a pilot
Tell us about your corpus and we will come back within two business days with scope, timeline, and a written quote.
Cihan AI Data Network™
Become a Cihan-certified analyst
Apply to join the network as a contributor, analyst or reviewer. Applications are screened, and qualified applicants are invited to calibration. Joining does not guarantee paid assignments.
In brief
What CIHAN AI DATA is, in plain terms
- What is CIHAN AI DATA?
- CIHAN AI DATA is an African AI data and human evaluation company operated by Cihan Business Solution Limited (RC 711803). It supplies annotation, data collection, benchmarking and model evaluation services for AI systems that have to work on African languages and discourse.
- What does CIHAN AI DATA do?
- It runs managed annotation and data curation, African data collection and corpus development, human model evaluation and red-team review, benchmarking and diagnostic evaluation, dataset QA and gold-set creation, managed AI talent deployment, and CMIA certification and workforce development.
- What is the Meaning Intelligence Lab™?
- Meaning Intelligence Lab™ is the research and benchmark engine for CIHAN AI DATA. It develops MIF™ (the Meaning Interpretation Framework), MIS™ scoring, gold-set and benchmark design, analyst calibration and contamination-aware evaluation methodology.
- How is quality assured?
- Quality is designed into each project: agreed guidelines before labelling, analyst calibration against reference items, multi-pass or blind annotation where the task requires it, inter-annotator agreement measured and reported per batch, senior-reviewer adjudication of disagreements, and hidden gold items audited through the batch.
- What is CMIA?
- CMIA is the Certified Meaning Interpretation Analyst credential, trained and issued through Cihan Digital Academy. Credentials are cryptographically signed and can be checked at cihandigitalacademy.com/verify.
- Who are the analysts?
- Analysts are members of the Cihan AI Data Network™ — recruited, trained, calibrated and certified African contributors working under NDA as network or managed talent, not as an anonymous crowd. Levels run from Contributor to Certified Analyst, Senior Reviewer and Project Lead / Adjudicator.
- What kinds of AI projects are supported?
- Training-data creation and curation, corpus development, safety and red-team review, evaluation of model outputs for meaning and cultural appropriateness, benchmark construction and diagnostic runs, dataset audits and gold-set builds, and standing evaluation panels for ongoing model work.
FAQ