Introduction
An outcome is a small object with one big responsibility: it is the only place in AI Governance where agent activity is converted into money.
Three parts — a definition, a set of sources, and a report computed over a window.
The three parts
Part | What it holds | Who sets it |
Definition | Name | You, once |
Sources | Each agent credited, with how it is verified and what each counted session is worth | You, per agent |
Report | Counts, value, spend and return over a window | Computed |
The split matters: the definition is only a name, and everything that decides counting and value lives on each source. Two agents on one outcome can be verified differently and can be worth different amounts.
The definition
Field | Type | Notes |
| String | Names the result, not the activity. Appears in the list and on the chart. |
Created by | User | Shown as the outcome's owner. There is no Owner field to set. |
A source
A source is an agent credited with producing the outcome, plus the rule that decides when one of its sessions counts.
Field | Values |
Agent | Any registered agent |
|
|
Conditions (rule) | One or more, combined with |
| The model that grades the output — text-generation models only |
| Slider from 0 to 1, default 0.85. Sessions scored below it do not count |
| Plain-language description of what a passing output looks like |
| An automation built on the outcome callable interface; optionally |
|
|
Each source carries its own rule, because two agents can reach the same result by different paths. One might call a ticketing tool; another might be judged on the quality of its reply.
Condition grammar
A rule condition is entity + target + operator + value.
Entity | Targets | Operators |
| A named tool |
|
| An evaluator bound to the agent |
|
|
|
|
|
|
|
The Evaluator entity is the interesting one: it lets a quality score gate a value claim, so an outcome only counts when the work was also good enough. That is the join between Evaluation and ROI, and it is the strongest rule you can write.
The report
Computed over the selected window, never stored per occurrence.
Figure | Derivation |
| Sessions that met a source's rule |
| Counted sessions × what each is worth |
| Spend of the crediting agents, from trace data |
| Total value − total spend |
| Total value ÷ total spend |
|
|
The overall strip also carries an unmeasured spend figure — spend not attributed to any outcome. It has its own Unmeasured spend tile.
Known limitations
Changing a value needs announcing — If somebody has already reported a figure, it may not reconcile after you change what a session is worth.
Counting is per session, not per trace — A session where the rule matched counts once, however many turns it took.
Rules read what the session carries — A rule referencing a renamed tool silently stops counting. Nothing warns you.
Return on spend divides by the spend of credited agents — Agents crediting no outcome contribute cost to your bill and not to this ratio.