direct-response-agent

Simulation window · 64 inferences · declaration v14

declaration: a1b2c3d4-e5f6-4789-a012-3456789abcde
content hash: 3cff3e86a506
window: 2026-08-04 to 2026-08-11

Rendered 2026-08-11 18:54 UTC

Essentials passed 210 measurement(s) · tiers essential, always, analytics

Met 2Missed 1Awaiting 3No target 1

Measured over 64 inference(s).

Conformance Built as declared

Whether the agent was built the way its declaration says, which is a different question from whether it behaved well, and fails to a different person.

Not checked

Said out loud rather than omitted. No findings and no answer look identical in the output, and only one of them is a clean bill of health.

declaration_content No published content hash to compare against, so nothing here confirms the inferences ran the declaration ai-gov published. Expected on a local report.

Guardrails

What each gate checked and how often it fired. A gate that never fires is not evidence of a clean population, only that nothing got past it.

Instruction	Checked	Violations	Rate
`conform_to_output_schema`	64	2	3%

Awaiting

Deduped by action, not by assessment. One labelling job unblocks every KPI that reads the evaluator.

Label some outputs — 18 of 40
staff_reply_decision
This measurement is produced by a person, not by the platform. Until labels arrive there is nothing for the agent to be right or wrong about.
Owed by a reviewer.
Unblocks 2 KPI(s) reply_accepted_sharereply_decision_coverage

Health checks

A target says what good looks like, so these settle something.

Grounded In Thread

Missed

0.86target ≥ 0.95

Passed 55 of 64 measurement(s)

passrate

The evaluator behind this

key: groundedness
evaluator agent: evaluator-groundedness
scale: binary
producer: —
tier: always
results: 64
pass rate: 86%

Reply Accepted Share

Awaiting

—≥ 0.40, not compared while awaiting

Over 18 measurement(s)

share of accepted

Label some outputs — 18 of 40

This measurement is produced by a person, not by the platform. Until labels arrive there is nothing for the agent to be right or wrong about.

Owed by a reviewer.

The evaluator behind this

key: staff_reply_decision
evaluator agent: —
scale: categorical
producer: external
tier: analytics
results: 18

Reply Decision Coverage

Awaiting

0.45≥ 0.90, not compared while awaiting

Collected on 18 of 40 eligible inferences

coverage

Label some outputs — 18 of 40

This measurement is produced by a person, not by the platform. Until labels arrive there is nothing for the agent to be right or wrong about.

Owed by a reviewer.

The evaluator behind this

key: staff_reply_decision
evaluator agent: —
scale: categorical
producer: external
tier: analytics
results: 18

Guardrail Fire Rate

Met

0.97target ≥ 0.95

Passed 62 of 64 measurement(s)

passrate

The evaluator behind this

key: output-conformance
evaluator agent: evaluator-output-conformance
scale: binary
producer: —
tier: essential
results: 64
pass rate: 97%

Length Requirement

Met

0.97target ≥ 0.90

Passed 62 of 64 measurement(s)

passrate

The evaluator behind this

key: response-length
evaluator agent: evaluator-length
scale: binary
producer: —
tier: always
results: 64
pass rate: 97%

Metrics

Measured, and nothing compares them to a target. They describe rather than settle.

Reply Similarity

No target

0.71no target

Over 64 measurement(s)

mean

No target set

The number is measured and nobody has decided yet what good looks like. Reported so it can be watched; it settles nothing until a target is declared.

The evaluator behind this

key: similarity-to-approved
evaluator agent: evaluator-similarity
scale: numeric
producer: —
tier: analytics
results: 64
mean: 0.71
p50: 0.74
p90: 0.88

Nothing owed on 1

Inert on this report: nothing ages, nobody is paged, and no amount of waiting changes them.

KPI	State	Why	Value
`label_stability`	Raise the replicate count	This measurement needs the same input answered more than once and got one run of each. More runs of that shape do not help; the run needs more replicates per input.	—