Evidence infrastructure for regulated AI decisions
When a decision is challenged,
what can you show?
Every organisation using AI for consequential decisions faces the same question. Not whether the decision was right. Whether the evidence behind it can be demonstrated.
TripleE — Exact Evidence Extraction is designed to answer that question. Before it is asked.
The problem
The liability is personal.
The evidence gap is structural.
When the FCA writes, it writes to the named Senior Manager. When a court asks, it asks what evidence the decision was based on. Section 66A FSMA and the EU AI Act require you to demonstrate exactly what evidence existed at the time.
Most AI systems are not designed to answer that question. The evidence base is generated at query time, influenced by the question asked, and not designed to produce a fixed verifiable record. Governance frameworks improve oversight. They do not change what the underlying system produces.
Current position
Risk committees. Audit logs. Ethics frameworks. These are valuable. They are not designed to produce a forensic evidence record reconstructible by an external auditor without organisational assistance.
With Provenance Kernel
A contemporaneous verifiable evidence record. Fixed before the decision. Reconstructible by any external auditor. Your answer is ready before the question is asked.
How it works
Read once. Question many times.
Provenance Kernel has no system prompt. The evidence base is built once at ingestion and is designed to remain fixed. Questions retrieve from it. They do not influence what was extracted.
Document ingestion
Documents are processed once. Every factual sentence is extracted verbatim and stored. The evidence base is designed not to change after ingestion.
Verbatim extraction
No inference. No summarisation. No reprocessing. Questions retrieve from the fixed evidence base. They do not reshape it.
Evidence retrieval
Ask any question. The system returns exact verbatim sentences categorised as supporting, contradicting, indicating risk, or absent. No narrative is produced.
Verified Fact Pack
SHA256 verified output. Identical outputs for identical queries against identical corpus states, verified cryptographically across a 94-day internal test interval, 23 tests, zero failures. An external auditor can reconstruct decisions from the record alone.
See it in practice
Same question. Same corpus.
Same answer. Every time.
A real Provenance Kernel Fact Pack generated against a publicly available regulatory document. The same question asked repeatedly across months. The output is identical each time.
Development status
What exists. What is proven. What comes next.
- Verbatim sentence extraction from document corpora
- Four category classification
- SHA256 verified JSON output
- Multi-sheet Excel Fact Pack
- On premises Apple M3 Ultra
- Local models for supporting layers
- No cloud dependency
- Identical outputs for identical queries against identical corpus states, verified cryptographically across a 94-day internal test interval, 23 tests, zero failures
- Real government documents tested
- UK parliamentary inquiry processed
- Identical hashes confirmed across multiple runs
- Architecture validated against Article 12 forensic reconstructibility standard
- Financial services domain rule library
- Yorkshire AI Labs funding application submitted
- MSc researchers QUB on predicate development
- Production deployment for section 66A compliance workflow
- First regulated firm validation partnership
Regulatory alignment
Every obligation. Every component.
Every practical benefit.
| Obligation | Requirement | How Provenance Kernel addresses it | What this means for you |
|---|---|---|---|
| Art. 12 | Automatic logging | Every extraction produces an immutable trace. SHA256 verified. Timestamped. | Your named Senior Manager has a defensible answer ready before the FCA letter arrives. |
| Art. 13 | Transparency | Verbatim source sentences exposed. No inference. No narrative. | Every output traces to its exact source document. No ambiguity. No interpretation required. |
| Art. 14 | Human oversight | Every output reviewable by third party without model access. Human always decides. | Human judgment remains central. Provenance Kernel informs it. It does not replace it. |
| Art. 15 | Reproducibility | Identical outputs for identical queries against identical corpus states, verified cryptographically across a 94-day internal test interval. | The same question asked months later returns the same answer. That is your audit trail. |
| S.66A FSMA | Personal liability | Contemporaneous verifiable evidence record. Fixed before any question is asked. | Personal liability becomes personal protection. The record demonstrates reasonable steps were taken. |
What you gain
Not a governance layer.
A different architecture. A different outcome.
| Property | Probabilistic AI | Provenance Kernel | The practical benefit |
|---|---|---|---|
| Reproducibility | Not designed to produce identical output on repeated queries | Designed for consistent reproducible output across repeated queries | Demonstrate consistency to any regulator at any time |
| Evidence base | Generated at query time, influenced by the question | Fixed at ingestion before any question exists | Designed not to be influenced by what the question needs to find |
| Narrative risk | Probabilistic output carries inherent hallucination risk | No narrative produced. Verbatim sentences only. | No fabricated citations. No court apology. No SRA referral. |
| System prompt | Can be changed between the decision and the challenge | No system prompt. Evidence locked at ingestion. | The evidence record is designed not to be altered after the decision is made. |
| Auditability | Reasoning cannot be reviewed after the fact | Verbatim source traceable for every output | Any external auditor reconstructs decisions independently within tested scope |
| Infrastructure | Cloud dependent. Token costs scale with usage. | On premises. No cloud dependency. No ongoing API costs. | No vendor lock-in. No pricing risk. No data sovereignty exposure. |
Why now
The deadline moved.
The liability did not.
The EU AI Act high-risk enforcement deadline has moved to December 2027. Section 66A FSMA 2000 is already in force. The cases are already public. Pinsent Masons. Sullivan and Cromwell. Major commercial insurers are now excluding AI liability from standard D&O, E&O and cyber policies without a demonstrable evidence record. The regulatory and commercial consequences are converging simultaneously.
Current approaches improve governance. Risk committees. Ethics frameworks. ISO certifications. Audit logs. These are not without value.
They are not designed to produce what Article 12 requires. A specific decision reconstructible from the record alone by an external auditor without assistance from the organisation’s staff.
That standard requires a different architecture. Not better governance around the existing one.
You have eighteen months to build on solid ground.
Or come and talk to us now.
That is an architectural requirement, not a governance one.”
The first mover opportunity
One organisation. One regulated workflow.
Twelve months ahead of everyone else.
We are offering one organisation in each regulated sector the opportunity to be the first to have Provenance Kernel validated against their specific compliance documents.
What you provide
Fund the validation exercise. Access to realistic compliance document types. Regulatory expertise to validate output against the forensic reconstructibility standard.
What you receive
Twelve months exclusive access in your regulatory space. A validated section 66A compliance workflow. The evidence infrastructure your competitors will not have.
Ready to talk? Call or email Ken directly.
Built by
Ken Salisbury
The insight behind Provenance Kernel came from solving the ground truth problem in cancer diagnostics. AI diagnostic tools fail on poor quality input. The same architectural problem exists in every AI system making consequential decisions. The evidence base is probabilistic. The accountability requirement is not.
Over 50 years building systems people rely on when failure is not an option. Prior exit to PerkinElmer. Visiting Researcher, Ulster University School of Biomedical Engineering. Named case study in the TechUK Photonics Report 2026.