Scoring, severity and confidence

Channel scores

Aiscovery scores search, browser and API/MCP channels separately. Each channel combines observed task outcomes with technical conditions, repeatability and efficiency.

channel_raw = 0.60T + 0.20C + 0.10R + 0.10E

T is severity-weighted task completion. C is severity-weighted technical conformance. R measures repeatable success across planned repetitions. E summarizes comparable step count, completion time, unexpected interventions and cost when cost is available.

Standard scenarios run three times and critical scenarios run five times unless the audit manifest records a justified alternative. A stable failure does not earn repeatability points.

Usability

Usability describes the quality of channels that are actually available to users.

An unavailable or approved out-of-scope channel is excluded, and the remaining weights are normalized. A website without an API/MCP surface can therefore receive a fair usability result for browser and search.

browser = 82
search = 76
usability_raw = (82 * 0.50 + 76 * 0.30) / 0.80 = 79.75

This does not claim the missing channel exists. Agent-native maturity handles that question separately.

Agent-native maturity

Maturity measures the breadth of agent-native capability, including governance and observability.

maturity_raw = 0.40B + 0.25S + 0.20A + 0.15G

B is browser readiness, S is search and citation readiness, A is API/MCP readiness, and G is governance and observability. A missing API/MCP channel receives zero here rather than being normalized away.

For the usability example above, an absent API/MCP channel and governance score of 58 produce maturity of 60.50. The same site can therefore be usable today but incomplete as an agent-native platform.

Severity caps

Aggregate scores must not conceal a critical defect. Caps apply after raw usability and maturity calculations:

unresolved applicable P0: final score is at most 39
otherwise unresolved applicable P1: final score is at most 59
otherwise: final score equals raw score

A cap applies only when the finding has valid evidence, reproduces in scope, affects an in-scope user or channel and has no verified mitigation. Reports show raw score, final score and cap reason together.

Confidence

Confidence is not another readiness component. It reports how much planned audit weight was executed with sufficient evidence.

confidence = 100 * sum(weight * execution_factor * evidence_factor)
             / sum(planned_weight)

Approved not_applicable work leaves the denominator. not_run and inconclusive work reduce confidence. A documented site block can still have full execution credit when evidence proves the cause.

A readiness score with confidence below 70 cannot support sign-off. High confidence also requires adequate session evidence for P0 and P1 runs, or an explicit capture limitation when a hosted agent cannot expose DOM or protocol data.