Start a conversation
Illustrative proposed scientific AI environment with researchers reviewing model evidence and robotic laboratory data

Proposed capability · subject to evidence gates

Scientific AI & Model Governance

Judge models by context, uncertainty and consequence, not novelty.

One programme. Two levels of scientific access.

The technical case remains rigorous while the public meaning stays clear and inspectable.

For scientific readers

A proposed scientific-AI environment could evaluate models used for experimental prioritisation, spectral interpretation, anomaly detection, process prediction or operational decision support against a declared context of use.

Evaluation would cover provenance, leakage, calibration, uncertainty, subgroup or domain performance, distribution shift, adversarial or missing inputs, human factors and retirement. Statistical performance alone would not establish scientific validity or regulated acceptability.

In plain English

AI can find patterns, but a confident answer can still be wrong. The important questions are what data the model saw, where it fails and who remains responsible.

The proposed programme would challenge a model before it is trusted and keep people able to override, pause and retire it.

Illustrative locked-challenge model evaluation

Scientific question

Does a model remain calibrated, useful and recoverable when tested on unseen data, missing inputs, domain shift and known failure cases?

Variables to admit
  • Dataset and site domain
  • Missingness and noise
  • Decision threshold
  • Model and preprocessing version
Observations to preserve
  • Discrimination and calibration
  • Conformal or other uncertainty coverage
  • Error severity by declared subgroup or domain
  • Human review and override performance
Controls and comparators
  • Locked independent challenge set
  • Simple baseline model
  • Predeclared metrics and acceptance limits
  • Leakage, shift and provenance audit

Decision useRestrict the context, require new evidence, retain a simpler method or stop deployment.

Calibration and uncertainty coverage

The Brier score measures probability error. Conformal coverage describes one uncertainty guarantee under specific assumptions; neither proves causality, fairness or safe use outside the tested domain.

Brier = N^(-1)Σ(pᵢ - yᵢ)² · P{Y ∈ C(X)} ≥ 1 - α
A model fragment for explanation, not a protocol, prediction or result.
Assumptions that must remain visible
  • Outcomes and sampling process are appropriate for the stated context
  • The challenge set is independent and representative enough for the claim
  • Coverage assumptions and exchangeability limits are disclosed

A proposed equipment system, not a procurement list.

Every system would need justified demand, competent operators, utilities, safety controls, maintenance and an accountable intended use.

  1. 01

    Secure scientific data enclave

    Separate governed data development, challenge and approved-use environments.

    Dependencies
    • Lawful data rights
    • Role-based access
    • Retention and deletion controls
  2. 02

    Versioned model registry

    Preserve model, feature, code, dependency and approval state.

    Dependencies
    • Named owners
    • Change control
    • Rollback and retirement route
  3. 03

    Independent challenge harness

    Execute locked performance, shift, missingness and robustness tests.

    Dependencies
    • Predeclared plan
    • Baseline comparator
    • Protected test data
  4. 04

    Explainability and error-analysis workbench

    Support domain review of failure patterns without treating one explanation method as truth.

    Dependencies
    • Scientific reviewer
    • Method limitations
    • Reproducible analysis
  5. 05

    Human-in-the-loop simulation room

    Evaluate how users interpret, override and recover from model output.

    Dependencies
    • Human-factors design
    • Representative users
    • No live regulated decisions

Machines may assist. Named people remain accountable.

Proposed data layer
  • Data lineage and lawful-use record
  • Model and dependency versions
  • Independent test results
  • Human decision, override and incident log
Human authority

The scientific owner defines intended use and limitations; data, quality, security and risk owners hold independent stop rights.

Automation must never

No model may decide patient care, clinical eligibility, batch release, regulatory compliance or scientific truth, or retrain itself into production without approved change control.

The output is a decision package, not a theatrical result.

Potential output

A model-applicability statement covering intended use, evidence, uncertainty, limits, monitoring, human controls and retirement conditions.

Stop or transfer when

Stop when data authority, independent performance, calibration, robustness, monitoring, human understanding or consequence controls are inadequate.

One scientific standard. Two distinct campus expressions.

Shared governance connects Glasgow and Hyderabad. Climate, infrastructure, demand and regional value keep their designs materially different.

GLA

Glasgow

Scientific focus
Responsible scientific AI for analytical, robotics, translational and advanced-manufacturing decisions.
Design response
Secure review rooms and model-test infrastructure physically separated from live laboratory control.
Value hypothesis
Could help Scottish and international teams turn promising models into inspectable, bounded evidence.

The campus expression is a planning hypothesis only. It does not represent a secured site, approved design, funded programme, partner commitment or operating capability.

HYD

Hyderabad

Scientific focus
Model testing, data engineering and human-governed automation across formulation, quality and transfer workflows.
Design response
Resilient local compute, secure partner workspaces and explicit boundaries between research and regulated systems.
Value hypothesis
Could build advanced AI and quality roles around real scientific use cases rather than generic software claims.

The campus expression is a planning hypothesis only. It does not represent a secured site, approved design, funded programme, partner commitment or operating capability.

A credible prospective user needs a defined decision, not square footage.

Medicines and biotechnology developers

Independent challenge of a model used in one development decision.

Proposed deliverable

A context-of-use and model-risk package.

AI and scientific-software companies

Domain-grounded evaluation beyond a product demonstration.

Proposed deliverable

A scoped challenge study without validation or regulatory endorsement.

Universities and data-science teams

A governed path from research model to realistic scientific workflow.

Proposed deliverable

An evidence and failure-mode assessment with rights agreed.

Measure sustainability at the same decision boundary.

Future intent is not current performance. Every measure requires a defined workflow, boundary and accountable record.

MetricMethodDecision use
Compute energy per accepted decisionRecord training, tuning, inference and storage energy for the defined use case.Compare model complexity against measurable decision value.
Avoided experiment burdenCount only experiments demonstrably removed without weakening the evidence gate.Test whether prioritisation creates real scientific efficiency.
Model and data retirementTrack inactive versions, duplicated datasets and storage lifecycle.Prevent indefinite digital accumulation.

No experiment advances on visual ambition.

Illustrative proposed experiment envelope. This is not performed work, a protocol, an installed capability, a service offer, a validated method or evidence of an operating laboratory. No GMP, GLP, manufacturing-licence, clinical or regulatory-readiness status is claimed.

  1. 01

    Declared context and consequence of failure

  2. 02

    Lawful, representative and traceable data

  3. 03

    Independent challenge and simple baseline

  4. 04

    Uncertainty, shift and cybersecurity testing

  5. 05

    Named human approval, override and retirement rights

  6. 06

    Applicable EMA, FDA, MHRA, CDSCO and data-protection review before regulated use

Primary context, with the caveat attached.

These sources inform the proposed programme. They do not prove VARUNÉ capability, affiliation, performance, compliance or regulatory acceptance.

Official framework

NIST Artificial Intelligence Risk Management Framework 1.0

Supports governance, mapping, measurement and management of AI risks across a defined context of use.

CaveatThe voluntary framework is not a validation certificate or sector-specific regulatory authorisation.

Open primary source
Official framework

EMA reflection paper on AI in the medicinal-product lifecycle

Provides lifecycle considerations for AI and machine learning used in medicine development and regulation.

CaveatThe reflection paper does not validate a model or replace product- and use-specific requirements.

Open primary source
Official framework

FDA artificial intelligence for drug development

Collects current FDA principles, guidance and publications for AI in drug development and manufacturing.

CaveatRegulatory acceptability remains evidence- and context-specific.

Open primary source

Bring a real question. Keep every claim inside the evidence.

A conversation or published concept does not create a partnership, service, installed capability, programme commitment or authority to use another organisation’s name.

Discuss a scoped capability studyReview concept and status definitions