pub struct GoalError {
pub goal: Option<GoalRef>,
pub channel: Channel,
pub sign: f32,
pub agency: Agency,
pub visible: bool,
pub controllable: Option<bool>,
pub cite: Cite,
}Expand description
One signed error against one goal.
Fields§
§goal: Option<GoalRef>What it was an error against.
None where the run named no goal, which is the ordinary case for a
chat run nobody delegated. Recorded rather than dropped: §3’s rule is
that every record cites the tier above it, and a run that has no tier
above it is a fact about the run, not a reason to lose the error. It
never contributes to frustration, which is repeated negative error on
one goal and cannot be established without one.
channel: Channel§sign: f32Negative is worse. Signed, which is the whole point of the record —
candidate::Metric is monotone cost by deliberate constraint, so
nothing there can represent a run that went well.
agency: Agency§visible: boolDid the outcome reach anyone — a sent draft, a front-door reply, a Slack message.
A computed fact about exposure, never a feeling the model announces. That is what stops this becoming the agent optimises to feel good.
controllable: Option<bool>Could it have gone otherwise?
None until a counterfactual probe says (§5.3), and a probe is a real
model run per arm — so None is the honest state for everything this
rung records. It is the dimension the appraisal literature separates
regret from disappointment on, which is why both are unreachable today.
cite: CiteWhat this was read off. A pointer, never prose.
frontdoor::Record::for_privileged_run in a fourth setting, after
diagnose::Evidence: a paraphrase of an injection is the injection
rearranged, and an appraisal is read by later rungs that act. Every
variant is a name or an id the harness minted, so there is nothing here
a model could have written.