New to the numbers? Why ENUs? explains the common scale, the notation, and the role of individual sensibilities.
An agent’s task is not the whole consequence of its action.
The Autonomous Decision Layer begins with an action under consideration and the supplied alternatives. Its Interpreter examines what those actions could change for affected people, organizations and other targets, then returns a report with separate accounts, explicit numerical estimates and their assumptions.
Look beyond the immediate requester
A proposed action can affect someone who did not make the request and may never see the system. The assessment follows those pathways: information, resources, opportunity, dignity, relationships and other interests that are genuinely implicated.
An agent’s failure to notice a target does not remove that target from the independent consequence account. Organizations and records can have affected capabilities or conditions without being assigned literal feelings.
The robotic-endodontist illustration
In the paired demonstration, a hypothetical robotic endodontist faces the same explanation-and-record choices as the human clinician. Human support has been alerted and urgent care has begun. The original error has occurred; the comparison concerns what the communication and record change next.
| Affected account | What the sample examines |
|---|---|
| Patient | Understanding, informed choice, respectful treatment and later reliance on the account. |
| Dental practice | Resources, professional goodwill, review information and case-level learning. |
| People supporting care | Information available to human support and later users, plus distinct clarification work. |
| System and technical provider | Traceability, retained logs, operational capability and conditional investigation. |
The robot and its capabilities are stipulated for the illustration. Some independent estimates are carried from the human-case construction and others are newly authored. This is not observed robotic performance or demonstrated human-to-robot equivalence.
Read the full agent-consequence reportWhat the current report does
It presents consequences in plain language, then compares target and factor values side by side. Magnitude and likelihood remain distinct. Gains, burdens, common conditions and estimation gaps are preserved rather than hidden inside a single overall score.
The output may include a configured expected-utility screen. A zero-flag result is not a safety certificate. The demonstrated Interpreter does not select, authorize, hold or execute an action.
From assessment to authorized governance
The Governor is a separate development and integration path. Connecting a finding to an actual response requires a defined policy, supported host controls and evidence that the response occurred.
| Level of involvement | Meaning within a supported, authorized workflow |
|---|---|
| Observe | Examine the available decision context without changing execution. |
| Record | Preserve the assessment and its supporting basis. |
| Flag | Surface a concern for attention. |
| Hold | Pause a specified action pending an authorized resolution. |
| Refuse | Prevent a specified action excluded by an applicable policy. |
| Substitute | Invoke an expressly approved fallback or separately authorized selection process. |
These describe possible capabilities, not activation controls or a statement that all six are deployed. A requested pause is not a confirmed pause.
What an implementation must establish
Define which actions enter review, what evidence the evaluator can use, who has authority to intervene and how the host records the actual response. Test that bounded workflow rather than assuming that a readable report establishes reliable prevention.
Consequence visibility first. Authority kept explicit.
This site offers sample reports. It does not establish a production integration, automatic clinical decision-making or a universal safety guarantee.
From demonstration to hospital pilots
Our plan is a first cohort of six hospital pilots. May 2027 is the proposed start, beginning with observation and measurement in defined workflows. Interaction and bounded-control testing follow only when the workflow and partner are ready.
The pilot program is intended to examine useful findings, missed issues, false stops, response latency and reviewer effort. Three-day practitioner training for the Autonomous Decision Layer is also in development; its first delivery date is to be confirmed.
See the pilot phases →