Evidence provenance
Important facts should retain their source, timing and context so a reviewer can understand what the recommendation was built on.
Trust
Trust should be visible in the decision path.
CAI is designed so that evidence, rules, model-assisted judgement, uncertainty and human authority remain distinguishable rather than collapsing into a single opaque recommendation.
Trust principles
Six controls should remain inspectable.
Observed vs inferred
Directly observed evidence should be separable from derived signals, interpretation and model-assisted inference.
Rules vs AI assistance
Deterministic policy checks should not be presented as model output, and model-assisted analysis should not be disguised as policy.
Uncertainty stays visible
Missing evidence, unresolved questions and low-confidence judgement should be exposed rather than silently converted into certainty.
Human decision rights
The authorised owner should remain explicit, including where approval, override, escalation or abstention is required.
Audit and learning
The decision basis, action taken and eventual outcome should be reviewable so the system can be evaluated and improved over time.
Current boundary
What sits outside the current boundary.
Autonomous approval and irreversible decision execution remain outside the current product boundary.
Continuous live monitoring applies only where the required data path exists and has been verified.
Regulatory certification, external assurance and compliance status are stated only when actually obtained.
Model output does not replace policy, delegated authority or accountable business judgement.
Data & privacy posture
Use only what the decision actually requires.
Phase 1 does not require customers to upload sensitive operating datasets through this website. Any later product workflow that processes customer data should define the permitted sources, access boundaries, retention, logging and deletion expectations before production use.
The design objective is evidence sufficiency, not indiscriminate data collection.
Evaluation philosophy
A recommendation is not enough. The decision outcome matters.
CAI should be evaluated end-to-end: whether the right evidence was used, whether uncertainty was handled appropriately, whether the recommendation was useful to the decision owner, and whether the eventual action produced a better outcome.
Discuss a governed decision