All docs
See what's happeningEvent detailsdashboard

Event details

Everything about one answer: summary, scores, claims, checks, history and details, with where to correct it.

Click any answer in Live events, Sessions or Issues and its details open in a drawer on the right. This is where you find out why an answer scored as it did, and where to correct it: the part of your app (retrieval, a tool call, routing, a hand-off between agents, the answer step, or setup) where the fix belongs.

The drawer's buttons are Add to dataset, Send to review and Copy link (a link straight to this answer, to paste in a ticket or chat). Its tabs are Summary, Scores, Claims, Checks, History and Details.

Summary

SummarySummary

Overall, Faithfulness and Safety; the answer with the parts that need attention underlined; Where to correct for the first failing check; and Scored by rule with the scoring rule that scored it.

Safety reports harm only: toxicity, bias, harmful content, privacy, inappropriate language, violence and adult content. A wrong fact is reported by Grounding and No contradictions, not by Safety.

Scores

ScoresScores

Every dimension score, grouped by its category. What each one measures: Reading your scores.

Claims

ClaimsClaims

The answer is split into claims: single statements that can be checked, such as "the deductible is $1,500". Each claim has a category, a reason and the source text it was checked against.

CategoryWhat Why saysExample
SupportedThe Matched text in the source"In-network deductible: $1,500 individual"
ContradictedBoth values: what the answer says and what the source says, naming the source"The answer says $2,500; Plan summary 2026 says $1,500."
Not in sourcesWhat is missing from the sources, with the Closest passage. It never says the claim is wrong"The sources don't give this figure ($900,000)."
Missing qualifierThe figure exists, but only under a condition the answer left out"$1,500" when the source says "$1,500 in-network; $3,000 out-of-network"
Framing (not counted)Why it isn't counted (Document analysis only)"Here is my review of the invoice."
Follows from verified findingsThe checked claims it was worked out from (Document analysis only)"The overcharge is $560.00" from two checked prices

A sentence that breaks one of your business rules also carries Breaks a business rule, with the reason, for example "Breaks a business rule: the answer approves 16 visits; the rule allows at most 12." Full grounding audit shows every claim, figure and document check. Every case, worked through: Claims and figures.

Checks

ChecksChecks

One row per check, with Check, Result (Pass, Fail, Warning, Not configured or Not applicable) and Detail: what the check found, with the values it compared. A failing check also says Where to correct, and a passing one says what it Relied on: the documents, tool calls or computed values that made it pass. For an answer built from a SQL query, each figure checked against the rows shows the computation that matched it, for example "SUM(paid) over 3 rows = 875,000.50".

When your project has business rules, a Business rules section under the table lists every rule that applied, its verdict and the facts read from the answer.

History

HistoryHistory

History of a conversation turn: the whole-conversation verdict is shown nowHistory of a conversation turn: the whole-conversation verdict is shown now

Every verdict this answer has had, one row each: Scored (when), How, Policy (the version and rule), Overall and Shown.

HowMeans
On arrivalScored when the answer came in
Re-scoreScored again by a re-score job
Whole conversationA rule over a whole conversation found this turn broke it. The line under it says which rule and session

The most recent verdict is Shown now: it is what Live events, the checks and the averages use. Earlier ones are kept in history, never deleted. Verdicts people gave in Human review are listed below the table.

Details

DetailsDetails

Session, environment, agent, tags, experiment, Pinned at arrival (the policy version that applied when the answer arrived), how it was scored (Deterministic or Max-Accuracy), latency and cost.

Not scored · outside the sample

A scoring rule can score a sample of its answers, to save cost on busy traffic. An answer outside the sample shows Not scored: outside its rule's sample. It is stored, not scored, and never counts as 0.

Worked example: retrieval or answer?

An answer says "$2,500 deductible" when the plan says $1,500.

  • The right document was sent. Claims shows Contradicted: "The answer says $2,500; Plan summary 2026 says $1,500." The right fact was retrieved, so Where to correct names the answer step: fix the prompt or the model call, not the search.
  • The document wasn't sent. The claim is Not in sources: "The sources don't give this figure ($2,500)." Where to correct names retrieval: the search didn't find the plan summary.