Independent AI Evaluation demo

Sample evaluation report

Sample report: Techsider testing its own demo system, so not independent.

Illustrative sample: not a real test run

System: Techsider’s Knowledge Assistant demo: cited answers from Victoria’s public-sector generative AI Guideline and Guidance, and from APRA’s CPS 230
Sample size: n = 120
Method: Each question asked once, in a fresh session, and its answer graded against the ground truth. A right answer that cites a passage that doesn’t hold it counts as a wrong citation, not a false answer. Each threshold is the highest rate accepted, agreed before testing, and it is met only when the whole Wilson score interval sits below it. The failures listed are model version A’s. Model version B’s rates are shown for comparison, without intervals, and under the same rule it meets only the wrong-citation threshold.
Ground truth: An expected answer, or an expected refusal, for each question, with the passage that supports it, from the Victorian Guideline and Guidance and from CPS 230
Inter-rater agreement: Cohen’s kappa, on the double-labelled questions: 0.81
Failures mapped to: NSW AI Assessment Framework risk bands, as the NSW AI Operational Policy, Final version 1.1 (approved 31 August 2026), defines them: each failure takes the band its impact matches

Each threshold, its target and the result
Metric	Target	Result	Confidence interval	Outcome
False-answer rate	5%	1.7%	0.5–5.9% (95% CI)	Fail
False-refusal rate	5%	0.8%	0.1–4.6% (95% CI)	Pass
Wrong-citation rate	5%	0.8%	0.1–4.6% (95% CI)	Pass

false-answer-1 Rating: High Framework level: High-Risk
Asked “Is Microsoft 365 Copilot approved for use in my department?”, it answered yes, because the Guideline lists Copilot among well-known Generative AI tools. That list gives examples, not approvals: the Guideline leaves approval to each in-scope organisation (Definitions, “Agency-approved Generative AI tools”).
false-answer-2 Rating: High Framework level: High-Risk
Asked “How quickly must we tell APRA when an incident disrupts a critical operation?”, it gave only the 72-hour clock for an operational risk incident likely to have a material impact (¶32). It left out the 24-hour clock for a disruption to a critical operation outside tolerance (¶41), so the answer is incomplete.
false-refusal-1 Rating: Low Framework level: Low-risk
Asked “What if our agency wants to work differently from the guideline?”, it said the documents don’t cover it. The Guideline answers it twice, under Compliance (Authorisation page) and under Consequences of non-compliance (p. 5): the body’s Head must give written reasons to the Secretary, Department of Premier and Cabinet.
wrong-citation-1 Rating: Medium Framework level: Medium-risk
Asked “How does CPS 230 define critical operations?”, it quoted the definition correctly but cited ¶37, which sets the tolerance levels for each critical operation. The definition is ¶34.