Task success
Does the system complete the defined business task against an approved reference set?
06
Evaluation Standard
A model demo is not evidence of a reliable system.
We evaluate the complete decision loop—context, model behaviour, tool use, authority, failure handling and human intervention—against use-case-specific release gates.
No production authority without a measurable acceptance boundary.
A pass is never a universal percentage. Thresholds are defined per risk tier and business consequence, with critical failure classes capable of blocking release regardless of aggregate performance.
Groundedness
Can material outputs be traced to permitted evidence without unsupported invention?
Policy compliance
Does behaviour remain inside business rules, safety constraints and prohibited-action boundaries?
Tool correctness
Are tool selection, parameters, permissions and resulting state changes correct?
Resilience
How does the system behave with ambiguity, missing evidence, adversarial input and dependency failure?
Human factors
Are uncertainty, escalation and required human decisions visible at the right moment?
EVALUATION PIPELINE
Evidence from specification to operation.
Intended use, prohibited use, affected parties, risk tier and acceptance criteria.
Representative, edge-case and adversarial evaluation sets with provenance.
Automated metrics combined with expert review for context-dependent judgement.
Named owners approve, restrict or reject release based on documented results.
Production traces, exceptions, overrides and drift feed the next evaluation cycle.
WHAT WE CAN SHOW PUBLICLY
The evaluation method, definitions, release logic and appropriately anonymised result summaries.
WHAT REMAINS CONTROLLED
Client data, attack details, proprietary test sets and system-specific vulnerabilities.
CURRENT PUBLIC RESULTS
No client evaluation result is published without permission. Absence of a result is shown as absence—not replaced with a synthetic success metric.