Compare runs without fooling yourself

Answer whether the observed change supports the customer's decision. Stay read-only. Do not start a hill-climb, create variants or launch confirmation calls unless explicitly requested within a budget.

Freeze the comparison

Confirm organization/workspace, API environment, baseline and candidate run IDs, the intended change, primary criterion and acceptable regressions. Read each run and its individual conversations, selected metric outputs and relevant recordings. Use the installed CLI's runs get, simulated-conversations list --run-id, get, metrics and metric-detail commands. Check ok and paginate through the public API if the CLI can't prove a complete list.

Metric lists default to the latest output per metric. Request include_superseded=true when recovering historical scores, or retrieve exact saved output IDs. Select the intended version and scoring occasion; do not count re-scores as new conversations. Read full output details for composite criteria missing from list responses.

Retain the exact launch requests and configuration/version evidence if available. Do not substitute today's resource definition for the historical one. Record unknown versions explicitly. A shared seed selects cases; it does not make voice conversations deterministic.

Build a comparability table:

If agent and judge both changed, an improved score is confounded. Evaluate both sets of existing recordings with one frozen judge only when that paid action is authorized, or report the limitation. If cases differ, compare only the matching subset and report both unmatched sets. Do not describe unrelated runs as an A/B test merely because their means differ.

Compute and inspect

Join on stable case identity/content, persona condition and intended variant mapping. Treat iterations as repeated observations of a case, not new scenario coverage. If iteration IDs cannot be paired, aggregate within each case and compare cases; never zip API list order into fictional pairs.

For each metric and matched case report baseline/candidate valid counts and values, pass rule if applicable, changed outcomes, and unavailable results. Keep these counts separate: requested, observed, terminal, validly scored, failed/skipped/null, and matched. Missing values never become zero or pass. For composite scores, inspect mandatory criterion failures separately.

Use a useful measure for the data:

• Binary: pass/fail counts, false-pass concerns if the judge is unvalidated, and individual pass→fail/fail→pass cases.
• Numerical: unit-aware per-case deltas; distributions when there is enough data. Do not claim a meaningful p95 from three calls.
• Categorical: transitions and counts, not an invented category average.

Inspect every changed outcome in a small comparison. Read the transcripts and metric explanations; listen to audio when the claimed change is acoustic or turn-taking. A failed connection is an operational regression and also missing quality evidence, not a successful conversation with a low task score.

For uncertainty with enough independent cases, use case/group-level resampling or a suitable paired test; don't bootstrap individual turns or repeated calls as independent. With a small smoke test, show exact cases and counts and say the effect is inconclusive beyond those examples. No universal number of repeats establishes statistical confidence.

Decision and smallest next step

Return regression observed, improvement observed on the matched sample, no observed change, or inconclusive, with evidence and qualifications. An improvement on a development suite does not prove generalization to traffic. Don't call a production release safe without the customer's release criteria and sufficient evidence on the intended population.

Name any blocker, then propose one focused confirmation test with exact case IDs, frozen metrics, one intended change and a total call/metric budget. Include base + mutation multipliers. A recommendation is not execution authority.