Guides

Interpreting your Intent Report

Read 30-day failure clusters, weights, trends, costs, questions, evals, and experiment evidence.

Report v2 covers latest 30 days of tenant failures. JSON at GET /v1/report; print-ready HTML at GET /v1/report.html.

#Headline values

total_failures → raw failure event count.
weighted_failures → reported events weight 1.0; inferred weight

confidence × 0.5.

estimated_cost → weighted failures × configured failure cost.
trend → raw daily failure counts for 14 days.

Estimated cost is prioritization proxy, not billed or verified loss.

#Cluster rows

Semantic clusters require at least three events and use latest 30 days. Each row includes stable ID, label/source, raw and weighted count, reported/inferred mix, seven-day vs prior seven-day trend, dimension mix, up to two redacted examples, suggested question, and regression eval seed.

Sort order is weighted count × failure cost, then raw count and label.

#Trend markers

Resolved status forces downward marker. New current-week cluster without prior events points up. Otherwise current and prior seven-day raw counts determine arrow and percentage. A percentage can be absent when prior week is zero.

#Suggested clarification

Question comes from dominant dimension in cluster aggregate:

Dimension	Suggested question
`goal`	What exact outcome do you want?
`output_format`	What format should result use?
`scope`	Which exact item should this apply to?
`constraints`	What limits or rules should agent follow?

Treat as starting point. Validate in product context and holdout before broad use.

#Experiment section

When enabled, report embeds same GATED/HOLDOUT session analysis as GET /v1/experiment. Below 200 sessions in either arm is directional only.

Source truthbackend/app/reporting.pybackend/app/report_html.pyREADME.md