Source fidelity
Can each important statement be traced to the exact admitted source—and do missing, changed or contradictory sources remain visible?
Research & engineering
Make the question harder.
Make the evidence better.
Aprentiz’s research connects private AI engineering with the practical demands of regulated work. Useful progress includes finding where a model, retrieval method or evidence process should not be trusted.
Two distinct research awardsAIRR Rapid Access on Isambard-AI and a separate AIRR Cloud pilot award support Aprentiz’s research and development.
The questions that shape the system
Research that changes
the engineering decision.
The aim is to improve the reliability and usefulness of the full workflow. A score has value when its test population, conditions, failures and practical limitations are understood.
Context integrity
Which facts, conditions, negations and source links survive repeated context compression, interruption and restart?
Useful private inference
Which model, retrieval and context configuration performs the defined work within its quality, privacy and operating constraints?
Evidence that travels
Can an authorised reviewer reconstruct the basis, decisions, changes and limitations without relying on an unexplained model response?
The research method
Freeze. Measure.
Challenge. Retain.
Predefined comparisons and retained negative results help prevent a persuasive demonstration from becoming a claim the evidence cannot support.
- Freeze
Question, sources, cases, baseline and acceptance criteria.
- Measure
Useful evidence, failures, uncertainty and complete cost.
- Challenge
Missing sources, contradictions, hostile input and recovery.
- Retain
Inputs, results, failed attempts, corrections and limitations.
- Decide
Improve, retain, narrow, investigate or stop.
Model agreement is not independent ground truth. Context compression is not perfect memory. A successful source check is not a professional judgement. Each needs the right evaluation and evidence.
Engineering progress & qualification
A built foundation.
A defined next test.
Aprentiz has substantial implemented components. Complete delivery and independent acceptance are assessed separately, against the actual workflow and environment.
IMPLEMENTED SOURCE
The engineering foundation.
Rust control and context services, provenance and assessment structures, GRC interfaces, deterministic Guard checks and workflow/evidence mechanics have retained source-test evidence.
PRIVATE R&D
Controlled integration work.
Retained private research checks cover selected engine, browser, governance and verification journeys. Those observations describe their tested scope and date.
NEXT QUALIFICATION WORK
Prove the complete journey.
Complete cloud and on-premises profiles, real professional workflows, independent security challenge, observed workload performance and whole-system recovery require their own evidence.
Public context & references
A wider need
for accountable AI.
These public sources inform the sector context and evaluation approach. They do not constitute an endorsement, certification or assessment of Aprentiz.
Artificial intelligence in UK financial services
75% of responding firms reported using AI; 46% reported only partial understanding of the technologies they use. The survey also examines model complexity and third-party dependencies.
Read the surveyAI in financial services
The FCA’s public material describes safe and responsible adoption, development and testing of AI in UK financial markets.
Read the FCA materialAI Risk Management Framework
A voluntary framework for considering AI risks across design, development, use and evaluation. Aprentiz makes no claim of certification or complete conformance.
Read the frameworkFacility publication & research support
The official Isambard-AI publication citation, facility statement and separate Cloud pilot acknowledgement are retained on our acknowledgements page.
Read the acknowledgements