Research and evaluation

The research behind our methods.

Artabel’s data and evaluation methods draw on published research in data quality, benchmarking and medical AI, and on earlier applied projects. This page shows how that research reaches client work, what an evaluation contains, and the records a deployed system keeps.

Method transfer Evaluation Governance

Method transfer

From research to delivery.

Four examples of how that research is used in client work.

How research method reaches client work at Artabel Four rows, each flowing left to right through three stages. Data-quality auditing research becomes a dataset audit protocol, which finds the label errors that limit a client model's accuracy. Synthetic-data benchmarking research becomes baseline and acceptance-criteria design, which shows whether a system beats the current way of working. Human and machine comparison research becomes review rubrics and blinded scoring, which produces a documented assessment by qualified reviewers. Grounded question answering becomes retrieval with provenance, which produces answers that cite their source. RESEARCH METHOD WE APPLY WHAT THE CLIENT GETS Data-quality auditing research Self-supervised detection of label errors,near-duplicates and off-topic samples Dataset audit protocol Triage that sends the most doubtfulsamples to a human reviewer first A cleaner training set Label errors that limit modelaccuracy, found and corrected Synthetic-data benchmarking Systematic evaluation metrics andcomparative benchmark construction Baseline and acceptance design The current method as baseline,acceptance criteria agreed first A defensible measurement Evidence of whether the system beatsthe current way of working Human and machine comparison Comparing learned representations,diagnostic behaviour and performance Review rubrics, blinded scoring Practitioner judgement, second review,origin of output hidden where feasible Professional sign-off Quality assessed by the people who willbe accountable for the output Grounded question answering Answering from curated sources,with the source of each answer kept Retrieval with provenance Source attribution, access tiers and ahuman review before publication Answers that cite their source Output your experts can checkagainst its source
The middle column is the engineering step: turning research findings and project experience into a protocol that works with real data and real reviewers.

Evaluation

Evaluation and quality assurance.

Evaluation shows whether a new system does better than the current way of working.

What an evaluation contains

  • A baseline from current practice. The comparison is the manual process, the tool you already use, or a standard model with off-the-shelf retrieval — whichever is the real alternative to the work.
  • Acceptance criteria agreed before the work starts. What counts as good enough is settled and recorded first, so the target cannot drift towards whatever the system happens to achieve.
  • Judgement by qualified people. Quality is scored by practitioners against a pre-agreed rubric, with a second independent review for material decisions and output origin blinded where feasible.
  • Reference judgements from people. Model-assisted scoring can support the analysis, but the reference judgement comes from qualified reviewers, and generated output is not reused as ground truth.
  • A sample that fits the claim. How many repetitions a result needs depends on the task, how much it varies and how strong the claim is. Where the evidence is too thin to carry a number, the finding stays descriptive and is kept out of aggregate claims.
  • Uncertainty and support effort reported. Results come with their uncertainty and with the staff time the system needed to use and correct.

Governance

Records and approvals.

Assurance proportionate to risk: an internal drafting aid does not need the review depth of a clinical document. Deployed AI systems keep the minimum record needed to reconstruct what they did and who approved it.

A governed AI workflow at Artabel Data rights and access constrain the whole pipeline. Sources feed ingestion and indexing, then modelling and retrieval, then a human review gate, then release with an evidence log. Evaluation against an agreed baseline covers the whole chain and feeds corrections back into the model stage. Data rights, access and retention — settled before any technology is chosen Sources Ingest & index Model Human gate Release Documents, images,sensors, archives OCR, metadata,embeddings, schema Retrieval, prediction,grounded generation Named reviewercorrects or rejects Output plus anevidence trail manual fallback kept Evaluation against an agreed baseline Task success · error rate and severity · correction effort · cost per task · acceptance criteria agreed in advance
The shape of a document or decision-support deployment: data rights are settled first, and a named person reviews the output before it is released.
  • A named approver. A designated person in your organisation signs off each release and can correct, reject or stop it.
  • A record of what ran. Purpose, data sources, service and model version recorded for every deployed system, so a result can be traced back months later.
  • Corrections logged. What the system got wrong, and how often, is recorded — which also tells you what it is costing your staff to correct.
  • EU AI Act and GDPR screening at workflow level, with escalation routes defined before go-live rather than after an incident.
  • No silent data export. Your data is not sent to general-purpose AI services without documented authority.

Next step

Need a system evaluated?

Describe the system and the decision the evaluation has to support.

Contact Capabilities