Technology & assurance
The record every decision leaves behind.
What we publish about how the platform earns trust: the Decision Evidence Record, the evidence each kind of data gets and the security and standards posture underneath. Method detail beyond this page remains commercially confidential.
The common assurance object
The Decision Evidence Record.
One standard record follows every important result across vision, tabular, time-series, text and agent workflows. Where Pipeline runs the workflow it captures the data and execution context; where Assure is attached to your own AI, that context comes with the connection. Assure adds explanation, monitoring and human review, and SparkSoft Automate records any approved downstream action — one consistent contract across models, data types and agents.
- 01Data & context
Input, schema, lineage, permissions and workflow purpose.
- 02Model & run
Model, code, dataset and component versions.
- 03Result
Detection, prediction, recommendation or agent action.
- 04Evidence
Concepts, features, comparables, citations, uncertainty and tests.
- 05Human review
Accept, reject, escalate or override — and the recorded rationale.
- 06Audit & feedback
Tamper-evident history, monitoring, drift and a controlled learning loop.
Operational feedback improves data, thresholds, the concept vocabulary and models under controlled governance — a learning loop, never uncontrolled retraining.
What the record holds.
A record is generated whenever a result crosses a customer-defined importance threshold, and remains queryable throughout the workflow lifecycle.
| Record section | Content |
|---|---|
| Decision identity | Workflow, case, timestamp, environment, business purpose and responsible owner. |
| Input and lineage | Input references, schemas, source systems, permissions and data-quality state. |
| Model and execution | Which model, data and settings produced the result, and under which policy. |
| Result and uncertainty | Result, score, calibrated confidence, threshold and warnings when the input falls outside the training data. |
| Evidence | Concepts, features, comparables, citations, rules, retrieved cases and known limitations. |
| Human decision | Reviewer, role, action taken, reason and second-review status. |
| Audit and learning | A tamper-evident history, signals that behaviour has shifted, and whether the case may feed future learning. |
Multimodal by design
Evidence, by modality.
Explanation is not one technique. Each kind of data gets the evidence that makes sense for it — and all of it lands on the same record.
| Modality | Result types | Evidence and explanation |
|---|---|---|
| Vision | Defect and object detection, quality classification, visual anomaly. | The image regions behind the call, the visual concepts they correspond to, similar past cases, and how confident the result is. |
| Tabular | Price, performance or risk prediction; classification. | Which factors drove the result and how much each mattered, comparable past cases, a confidence range, and how the answer would change under different conditions. |
| Time-series | Forecast, anomaly, change point or operational alert. | How far the reading departs from normal, which signals contributed, what else was happening at the time, and comparable past episodes. |
| Text & retrieval | Answer, classification, extraction, recommendation or policy review. | The sources behind the answer and their versions, flags where a claim is not supported by them, and reviewer confirmation. |
| Agents | Tool calls, reports, notifications or controlled workflow actions. | Which agent acted and under whose authority, the evidence it used, what was approved, and the final action record. |
Model neutrality
Your models keep their place.
Assure works with the AI you already run, and with models we build for you. Where our own research earns a customer a genuine advantage we bring it in — architectures designed so an explanation is a property of how the decision was made, rather than a story assembled afterwards. Either way, the assurance contract is the same and the choice of model stays yours.
Security, sovereignty & governance
Built in across the suite — not sold as an extra.
Deployment where you need it
SparkSoft-managed SaaS, your own cloud account, a private cloud, on-premises or edge — chosen by the use case, not forced by the product.
Isolation and control
Identity and tenant isolation, encryption, customer-managed keys, data retention, lineage and action logging applied consistently across Assure, Pipeline and Automate.
Standards alignment
Product controls are mapped to the UK AI Cyber Security Code of Practice, the NIST AI Risk Management Framework and ISO/IEC 42001 — reducing the work your own governance and supplier due diligence need to do.
Alignment is a design commitment, not a certification claim — we say so plainly, and we validate specific assurance, security and legal requirements per customer, sector and jurisdiction.
Where LLMs fit
Evidence first. Language second.
SparkSoft does not use LLMs as an unsupported authority. Domain-trained models and structured business evidence produce the result; LLMs help you query, understand and act on that evidence. In Automate, agents remain constrained by identity, permissions, policy and human approval.