Skip to main content

Data science

Separate the pattern from the decision.

Werkon turns a business question into traceable analytical evidence without hiding the population, assumptions, uncertainty, bias, alternatives, or limits. The result remains support for an accountable decision, not a substitute for one.

Analytical contract

Define the question, evidence, and permitted use together.

The contract begins with the decision and the real-world process that generated the observations. It then makes population, time, outcomes, exposures, features, assumptions, baselines, uncertainty, bias, access, action, correction, and ownership inspectable before selecting a method.

Inputs

Decision and current baseline
Decision, accountable owner, users, affected people or operations, current rule or judgment, available actions, cadence, horizon, cost of error, reversible and irreversible consequences, existing measures, minimum useful improvement, escalation, and stop conditions.
Question and evidence design
Descriptive, diagnostic, causal, predictive, forecasting, simulation, or optimization question; target population; unit of analysis; outcome; exposure or intervention; features; comparison; time; confounders; assumptions; counterfactual limits; and acceptable uncertainty.
Data generation and provenance
Authoritative sources, collection process, sampling and selection, identifiers, grain, measurement, missingness, labels, corrections, joins, transformations, leakage paths, historical policy effects, sensitive attributes, access, retention, lineage, and known changes.
Method and operating environment
Baseline, analysis or model candidates, train and evaluation separation, metrics, segments, uncertainty, sensitivity, robustness, reviewers, reproducibility, software, infrastructure, integration, permissions, human review, monitoring, recalibration, support, and retirement.

Outputs

Decision and analysis plan
A versioned record of the question, intended use, decision owner, population, variables, time, assumptions, baseline, method options, evaluation design, segments, uncertainty, bias risks, safeguards, reviewers, acceptance criteria, and prohibited interpretations.
Protected analytical dataset
A permissioned, documented, reproducible dataset with source lineage, population and sampling rules, feature and outcome timing, quality findings, missingness, transformations, sensitive handling, split policy, access, retention, and representative synthetic fixtures where useful.
Reproducible evidence pack
Versioned code and environment, data references, baseline and candidate results, uncertainty, segmented evaluation, sensitivity and robustness checks, error analysis, alternative explanations, reviewer findings, limitations, and a plain-language decision record.
Controlled use and monitoring pack
A bounded implementation or handoff with input and output contracts, decision authority, explanation, abstention or fallback, access, audit, quality and drift signals, outcome review, incident and correction paths, change control, support ownership, and retirement criteria.

Evidence path

Protect the evaluation before examining the answer.

A useful analysis should survive a clear baseline, untouched evaluation evidence, meaningful segment review, alternative explanations, and independent challenge. The method can change as evidence improves, but the decision question and evaluation rules must not move invisibly with the result.

  1. 01

    Frame the decision and estimand

    Observe the current decision and consequences, distinguish description from causation and prediction, define population, unit, outcome, exposure, comparison, time, action, authority, minimum useful result, uncertainty, and conditions under which analysis should stop.

  2. 02

    Inspect how the data came to exist

    Trace collection, selection, measurement, labeling, missingness, corrections, joins, historical decisions, access, and lineage; identify bias and leakage paths; and decide whether the available observations can answer the stated question at all.

  3. 03

    Set the baseline and protected evaluation

    Choose a simple current or naive baseline, predefine metrics and meaningful segments, separate development from evaluation by entity and time where needed, preserve an untouched test or prospective design, and set acceptance and rejection rules before tuning.

  4. 04

    Analyze, compare, and stress the result

    Run reproducible candidates, quantify uncertainty, examine errors and segments, test sensitivity to assumptions and data changes, compare alternatives, check calibration and practical significance where relevant, and invite independent analytical review.

  5. 05

    Authorize bounded use and keep watching

    Communicate the result and limitations in decision language, choose report, experiment, model, simulation, or no deployment, keep consequential authority explicit, stage exposure, monitor inputs and outcomes, review drift and harm, correct errors, and retire unsupported use.

Analytical form

Match the method to the claim the evidence can support.

Methods answer different questions. Choosing from the desired output alone can turn a descriptive pattern into a causal story, a forecast into an individual judgment, or an optimization result into an instruction that ignores real constraints.

01The first need is to understand what occurred

Descriptive or diagnostic analysis

Use governed summaries and exploratory analysis to characterize distributions, trends, cohorts, sequences, associations, outliers, missingness, and possible explanations without claiming that an observed relationship caused the outcome.

Evidence: Population and grain, source lineage, definitions, time, sampling, missingness, comparisons, segments, multiple testing where relevant, uncertainty, alternative explanations, reproducible code, reviewer, and decision use.

02A causal or comparative effect matters

Statistical inference or experiment

Use a randomized or carefully justified observational design when the question concerns an intervention or effect, while making assignment, counterfactual assumptions, interference, attrition, power, uncertainty, and generalization limits explicit.

Evidence: Estimand and population, treatment and comparison, assignment or identification strategy, assumptions, pre-analysis plan where suitable, sample size, outcome timing, attrition, sensitivity, uncertainty, ethics, and independent review.

03Future or unknown outcomes guide a bounded action

Prediction or forecasting

Use a predictive model when a forecast or score can be evaluated against future or protected observations and an owner can define thresholds, segments, calibration, abstention, decision costs, review, monitoring, and fallback.

Evidence: Prediction target and horizon, feature availability time, leakage review, baseline, split design, metrics and decision costs, calibration, segment errors, uncertainty, drift, threshold owner, fallback, and outcome monitoring.

04Choices must be compared under constraints

Simulation or optimization

Use simulation or optimization when the system, objective, alternatives, and constraints can be modeled explicitly, while treating assumptions, parameter uncertainty, feasibility, sensitivity, and human approval as part of the result.

Evidence: System boundary, objective and competing objectives, constraints and authority, input provenance, assumptions, calibration and validation, scenarios, sensitivity, uncertainty, feasibility checks, override, and realized-outcome review.

Evidence controls

Keep uncertainty visible all the way to action.

Analytical quality is not one score. It depends on the intended use, the data-generating process, the method, the evaluation design, independent assurance, communication, and whether the result remains valid when conditions change.

Development cannot consume the final test
Separate exploration, tuning, validation, and final evaluation by entity, group, time, or process as the use requires. Record every reuse of protected evidence and prefer prospective evaluation when historical policy makes a clean test impossible.
Bias review follows the full lifecycle
Inspect how people, institutions, sampling, measurement, labels, objectives, interfaces, automation, and downstream decisions can introduce or amplify bias. Evaluate meaningful groups and contexts without treating parity on one metric as universal fairness.
Uncertainty and alternatives travel with the result
Report intervals, distributions, sensitivity, assumptions, missing data, model and parameter limits, segment variation, practical significance, plausible alternative explanations, and conditions outside the evaluated population or time.
Authority remains outside the analytical artifact
Keep scores, forecasts, experiments, and optimized plans downstream of authenticated permissions, deterministic constraints, qualified review where required, audit, appeals or correction, fallback, incident response, and an owner who can suspend use.

Engagement fit

Use data science when a bounded decision needs more than descriptive reporting.

Good reason to begin

  • A recurring or material decision has a named owner, a current baseline, observable outcomes, explicit consequences, and a question that can be stated as descriptive, causal, predictive, or prescriptive.
  • Source and domain owners can explain how records, labels, outcomes, interventions, corrections, and historical decisions were created, and representative data or safe synthetic alternatives can be inspected lawfully.
  • The work can reserve evaluation evidence, compare a simple baseline, examine uncertainty and meaningful segments, document assumptions, accept independent challenge, and publish limits alongside the result.
  • The organization can own access, integration, decision authority, monitoring, review, correction, incident response, model or analysis changes, support, and retirement after the initial evidence is delivered.

Resolve before beginning

  • The desired conclusion has already been chosen, evaluation criteria will move until it appears, contrary findings cannot be accepted, or no decision owner can define what a useful or unacceptable result means.
  • The question depends on inaccessible, unlawfully used, unrepresentative, manipulated, unlabeled, or unexplained data and there is no credible path to inspect provenance, population, measurement, bias, or correction.
  • A descriptive association is expected to prove causation, historical fit is expected to guarantee future behavior, or a single aggregate metric is expected to establish safe performance for every person and operating context.
  • The requested output would make a consequential decision without explicit permission, qualified review where required, meaningful appeal or correction, audit, fallback, monitoring, and an accountable owner who can stop it.

Source basis

Sources behind the control model.

  • 01

    National Institute of Standards and Technology

    Artificial Intelligence Risk Management Framework 1.0

    The NIST framework organizes AI risk work across Govern, Map, Measure, and Manage and describes trustworthy characteristics including validity, reliability, safety, security, resilience, accountability, transparency, explainability, privacy, and managed harmful bias.

  • 02

    National Institute of Standards and Technology

    Towards a Standard for Identifying and Managing Bias in Artificial Intelligence

    NIST Special Publication 1270 treats bias as a socio-technical risk that can arise from systemic, computational and statistical, and human processes across the AI lifecycle, not only from a model or dataset.

  • 03

    UK Government Analysis Function

    The AQuA Book

    The 2025 guidance places proportionate analytical assurance throughout engagement, scoping, design, analysis, delivery, communication, sign-off, maintenance, and review, with verification, validation, documentation, uncertainty, independent assurance, and accountable approval.

[ WORKFLOW / SYSTEMS AUDIT ]
THE FIRST ENGAGEMENT

Start with one real workflow

A Systems Audit is the usual starting point. If the opportunity is already clear, we can move directly into a focused build.

Show Us the WorkflowStart with the free automation readiness checklist

OBSERVEQUANTIFYDECIDEBUILD

Continue with the work

Choose the next useful path.

Stay with the operating question. The technology can wait until the work is clear.

01

Explore data services

Place analytical work inside the wider engineering, analytics, scale, migration, database, quality, lineage, access, and lifecycle responsibility map.

Open path02

Build a traceable data flow

Preserve source authority, schema, time, quality, lineage, delivery, replay, reconciliation, correction, and ownership before analytical use.

Open path03

Build governed decision views

Turn validated observations and maintained analytical findings into defined metrics, reports, dashboards, explanations, and alerts with explicit limits.

Open path04

Build quality evidence

Plan independent verification, realistic evaluation, accessibility checks, residual-risk reporting, and release decisions around the intended use.

Open path05

Develop a bounded AI model

Compare baselines and candidates against protected evidence, meaningful segments, limits, versioned release, monitoring, and explicit human authority.

Open path