DataKnobs Field Guide

Data as Levers

A practical guide to the knobs that shape AI systems. Twenty control points, each with its own data asset, knobs, outcome and measurement.

Core idea

Every change follows the same control pattern

A data asset is the thing you change. Knobs are its adjustable properties. The outcome is what should change when a knob moves, and the measurement is the evidence that it did : rather than the problem moving somewhere else.

1

Data asset

The thing you change : training examples, retrieved knowledge, test cases, user context, feedback, source records.

2

Knobs

The adjustable properties of that asset : quality, freshness, weighting, segmentation, permissions, thresholds, retention.

3

Outcome

The measurable behaviour that should change when the knob moves. Stated before the change, not after.

4

Measurement

The evidence that the change helped rather than shifting the problem somewhere else.

“Data is where the leverage is. The properties of a data asset are the knobs. Metrics show whether moving a knob produced the intended behaviour.”

Why it matters

What the framework buys you

1

A shared language

Product, data, AI, risk and engineering argue about the same object : a named data asset with named knobs : instead of trading opinions about “the model”.

2

Traceable results

Every user-facing answer connects back to sources, transformations, model versions, policies and metrics.

3

Cheaper improvement

Most quality problems are data-asset problems. Turning a knob is faster, cheaper and more reversible than changing the model.

4

Earned trust

Users see not just a score, but which inputs drove it, how current they are, what assumptions applied, and how confident the system should be.

The guide

Twenty topics, five parts

Each topic is one data asset with its own knobs, outcome and measurement. Open any of them for the detail, the worked Stocks Assistant application, and the metrics that prove the knob moved what you intended.

I

Shaping Model Behaviour

What a model learns, and how reliably that behaviour transfers to production.

Topic 1

Fine-tuning

Fine-tuning turns a general model into a specialist by repeatedly showing it the inputs and outputs that define the desired behaviour.

Data asset
training examples
Knobs
example quality, volume, task mix, difficulty, weighting, curriculum order
Outcome
specialised and repeatable model behaviour

Read the topic →

Topic 2

Retrieval-Augmented Generation

RAG changes model behaviour by changing the evidence placed in the model's context at answer time, rather than changing the weights.

Data asset
retrieved knowledge
Knobs
source hierarchy, chunk size and overlap, metadata, freshness rules, top-k and reranking, citation threshold
Outcome
grounded answers tied to current evidence

Read the topic →

Topic 3

Prompt optimisation

Prompt optimisation treats instructions, demonstrations and output formats as a configurable data layer : the fastest and most reversible control available.

Data asset
instructions and context examples
Knobs
example selection, ordering, diversity, output schema, tool instructions, response constraints
Outcome
more consistent responses without retraining

Read the topic →

Topic 4

Synthetic data

Synthetic data fills gaps in training and evaluation : but only when it is designed from a coverage plan rather than produced as a large undifferentiated batch.

Data asset
generated training and test examples
Knobs
generator model, constraints, coverage plan, rarity, perturbation level, human review rate
Outcome
coverage of important cases that real data does not supply reliably

Read the topic →

II

Grounding, Evaluation and Orchestration

What evidence the system uses, how quality is measured, and how work is routed.

III

Personalisation and Responsible AI

Adapting to the user while keeping safety, fairness and review boundaries explicit.

V

Advanced Decision Intelligence

Extending the framework to complex documents, auditable scores, causal questions and product decisions.

Topic 17

Multimodal AI

Many financial facts lose meaning when reduced to plain text. Treating layout as data preserves the relationships that make a number interpretable.

Data asset
text, tables, charts, images and document layout
Knobs
resolution, ocr engine, page selection, layout preservation, modality balance, cross-modal validation
Outcome
better understanding of complex financial documents

Read the topic →

Topic 18

Explainability

An explanation should be faithful to the actual system rather than a persuasive story generated after the fact.

Data asset
evidence, metadata and model trace data
Knobs
provenance depth, factor visibility, lineage, citation granularity, confidence reporting, version disclosure
Outcome
auditable decisions and understandable scores

Read the topic →

Topic 19

Causal inference

Even when causal certainty is impossible, the framework forces the product to state which assumptions connect the evidence to the conclusion.

Data asset
observational and experimental data
Knobs
treatment, control, outcome window, confounders, matching, identification assumptions
Outcome
better distinction between correlation and cause

Read the topic →

Topic 20

Business experimentation

Experimentation applies the same control pattern to the product itself, and becomes the outer feedback loop around the whole system.

Data asset
customer behaviour under controlled product changes
Knobs
audience, intervention, timing, randomisation unit, success metric, guardrail metric
Outcome
product improvements supported by evidence

Read the topic →

Method

How to turn a knob without breaking something else

1

Name the data asset

Which data asset are you actually changing? If the answer is “the model”, keep looking.

2

State the outcome first

Write the behaviour that should change, and by how much, before you touch anything.

3

Move one knob

Change a single property. Two at once and you cannot attribute the result.

4

Measure by slice

Score the segments that matter operationally, not one headline average.

5

Check for displacement

Confirm the problem moved off the board rather than into another segment.

The rule that keeps this honest: a change ships only if it improves the intended metric without crossing a guardrail elsewhere. That single sentence turns evaluation from a report into a control.

Operating model

Twenty topics, one loop

Not twenty initiatives : four stages that feed each other, continuously. What you learn in the last stage changes what enters in the first.

Enter

Quality, privacy and freshness controls on source and user data

Transform

Retrieval, prompts, models, tools and agents do the work

2 3 7 8 17

Validate

Evaluation, claim checks, explainability and safety gate the result

Learn

Active learning, feedback, drift, causal analysis and experiments decide what changes next

Distillation comes last: it compresses stable, well-evaluated behaviour into efficient student models : never behaviour that has yet to be validated.

Sequencing

What to build first

The order matters more than the ambition.

1

Foundations

Canonical source metadata and point-in-time evaluation

Nothing above this layer is trustworthy without it

2

Evidence

Claim-level citations and confidence-based escalation

Now the system can say what it does not know

3

Efficiency

Distil high-volume extraction, classification and summary tasks

Compress only what is already stable and measured

4

Product knobs

Expose score, personalisation, risk and time-horizon controls to users

The framework becomes visible to the people it serves

Only then: automate more complex agent workflows, or treat ranking changes as proven signals. Sequencing this the other way produces confident systems built on unverified ground.

Artifacts

The minimum you must be able to show

If a control layer has no artifact, it is an intention rather than a control. A good test: pick any number on the screen and ask which artifact proves it.

Control layerMinimum production artifact
SourcesAuthority ranking, timestamps, fiscal-period metadata and lineage
ModelsVersioned prompts, student/teacher routing, schemas and confidence
ScoresFactor definitions, peer groups, weights and formula versions
RiskClaim verification, guardrails, permissions and escalation rules
LearningFrozen benchmarks, production review queue, drift alerts and a change log