PDICONIntelligence
Contact

Precise terms for governed industrial AI

Industrial AI Glossary

This glossary defines the vocabulary PDICON Intelligence uses: workflow episode, expertise data, Training OS, model interoperability, Project Memory, evidence linking, approval gate and outcome feedback. Consistent terminology is what allows technical claims to be evaluated precisely.

Talk to the team See how a pilot works
Maintained reference

Glossary

The words we use, defined.

Every entry maps to a component PDICON builds, with its build status stated on the linked route.

01Workflow episode
A structured record of one expert decision containing goal, context, observations, actions, evidence, the decision taken, any correction and the eventual project outcome.
02Expertise data
Training data constructed from real expert workflow episodes rather than from documents alone, preserving the reasoning path between context and decision.
03Workflow capture
The privacy-aware event layer that converts meaningful project actions into structured records while excluding credentials, payment data and unrelated sensitive fields.
04Training OS
The control plane that turns a versioned dataset and training specification into an evaluated model candidate with complete lineage across code, configuration, hardware, metrics and artefacts.
05Model interoperability
A design in which applications request a capability rather than a named provider, allowing a router to select the model and runtime that satisfy quality, latency, cost, security and residency requirements.
06Project Memory
A permission-aware institutional record connecting projects, equipment, vendors, decisions, changes, outcomes and lessons so completed work remains searchable and trainable.
07Evidence linking
The practice of attaching every generated statement to its source document, revision and location so a reviewer can verify the claim rather than trust it.
08Approval gate
A mandatory point in a workflow where a qualified human accepts, corrects or rejects a recommendation before any consequential action proceeds.
09Outcome feedback
The delivered project result — equipment performance, achieved schedule, realised cost — returned to the dataset as the signal that determines whether earlier assistance was correct.
10Deviation
A documented difference between a controlled requirement and a vendor submission or executed condition, recorded with its reason and commercial consequence.
11Compliance matrix
A structured comparison of controlled requirements against submitted responses, showing compliance, deviation or missing information for each line.
12Permission-aware retrieval
Search and retrieval that filters results by the requesting user’s project and document access, so semantic matching cannot leak restricted content.
13Tenant isolation
Architectural separation ensuring one customer’s workflow data, retrieval scope and model execution cannot reach another customer.
14Training consent
An explicit written agreement defining whether captured project evidence may contribute to model training, for what purpose, for how long and with what revocation path.
15Model lineage
The reproducible record connecting a deployed model to its dataset version, code, configuration, hardware, evaluation metrics and approving reviewer.
16Held-back evaluation
Scoring assistance against historical cases withheld from training and adjudicated by qualified experts, so quality is not assessed by the system that produced the output.
17Degraded mode
Explicitly labelled operation when a model or connector is unavailable, so a weaker result is never presented as a normal answer.
18Design partner pilot
A scoped commercial engagement instrumenting one real decision end to end, with agreed evidence boundaries, evaluation method and exit terms.
Why terminology matters

Vague vocabulary hides weak engineering.

Terms like "AI-powered" carry no verifiable meaning. Defining workflow episode, lineage and held-back evaluation precisely makes the architecture assessable by an engineer or an investor.

Terms match the product

Each definition maps to an implemented or declared component.

The glossary is not general industry vocabulary. Every entry corresponds to something PDICON builds, with its status described on the linked route.

Operating matrix

Evidence moves through explicit controls.

The matrix separates source evidence, intelligence work, governing authority and the resulting artifact.

Control surface Reviewable lineage
SubjectInputIntelligence operationHuman / policy controlOutput
Workflow termsExpert workEpisode constructionConsent and exclusionsTrainable record
Training termsDataset versionReproducible trainingApproval gateEvaluated model
Governance termsModel outputEvidence linkingQualified approvalAccountable decision

Design boundary

What the system will not pretend to be.

Credibility begins where automation stops and accountable professional judgment starts.

01

Not an industry standard

These are PDICON’s working definitions, published for precision rather than presented as sector standards.

02

Not a capability claim

A defined term does not imply the component is fully deployed; status is stated on its route.

Get started

Start with one real decision.

Bring a workflow your team already repeats. We define the evidence, the approver and the outcome to measure, then instrument it end to end.

Talk to the team How engagements work

We are working with a small number of design partners across engineering and procurement workflows.