---
name: orphex-experiment-result-reviewer
description: "Review a supplied marketing or conversion experiment against its declared design, mature outcomes, uncertainty, and guardrails; allow inconclusive or invalid results without forcing a winner."
license: MIT
metadata:
  version: "2.1.0"
---

# Orphex Experiment Result Reviewer

Evaluate the supplied experiment against its predeclared design and business decision. A higher observed metric is not automatically a statistically supported or economically acceptable winner.

## With an Orphex connection

If an Orphex connector is available, check it before asking for exports. Run `capability_search` with platform `orphex` and limit 100; use a read only if that search returns its id, and ask which workspace to use when several are bound. Read `causal.segments_read` then `causal.segment_read`, or `causal.latest_summary`, for any Impact Analysis of the changed campaigns; it is a modelled counterfactual, not a randomized result, so report it beside the experiment rather than in place of it. Read `controller.catalog` then `controller.fetch` for arm-level outcomes when the arms are separate entities. A Google Ads experiment's control and treatment arms are a live read through `platform_management.performance_read` when a search with platform `google` returns it. Then list guides with `skill_catalog` (kind `guide`, topic `platform_reads`, then `measurement`) and follow a matching one through `skill_read`; choose by title, never by a stored id.

If a read is not returned, this connection cannot reach that Orphex data for the workspace: say so, then use the live platform reads it offers or the supplied exports. An absent, refused or empty read is not evidence of zero, none or healthy. Describe each read with `capability_describe` before `run_read`, keep `date_end` no later than yesterday, disclose request adjustments, and name each number's source, workspace and window. This skill's evidence rules still govern any guide, insight or playbook label, and no read authorizes account changes. Ask the user only for what is still missing.

## Verify what was tested

Read the hypothesis, primary outcome/denominator, assignment unit and method, allocation, dates, maturation, exclusion rules, effect criterion, confidence/method, and guardrail thresholds. Check assignment integrity and unexpected sample imbalance without treating its cause as known; identify repeated exposures, contamination, concurrent edits, early stops, or unplanned exclusions. Analyze assigned eligible units under the supplied intention-to-treat rule when appropriate. Platform-optimized ad delivery and simple before/after reports do not establish randomization.

Reconcile arm counts and numerator boundaries, then show absolute rates/counts, absolute effect in percentage points, and relative effect only with a nonzero baseline. Prefer the supplied valid platform/statistical report with its method and scope. The optional helper offers only an independent-binary fixed-horizon normal approximate difference interval. It is unsuitable for small/extreme counts, clustered assignment, revenue/CPA ratios, fractional attribution credits, multiple comparisons, or sequential peeking; do not silently reuse it. Compute with a valid method or explicitly leave inferential status unavailable.

A confidence interval including zero is inconclusive, not evidence of equivalence. Equivalence/noninferiority needs a predeclared meaningful margin and appropriate analysis. Statistical support alone does not satisfy profitability or guardrails. Check mature secondary/guardrail outcomes and do not certify a guardrail without its declared rule. Adjust or disclose multiplicity and the consequences of post-hoc segment picking.

## Decide within scope

Classify the result as supported benefit, supported harm, inconclusive, or invalid under the documented method; describe any unknowns. Recommend rollout only when design, primary effect, maturity, and business guardrails support it. Any follow-up needs a planned decision rule, not indefinite extension until a winner appears. Preserve the current control and keep applying experiments or shifting traffic within explicit authorization.

## Official reference

- [Google experiment scorecards and confidence settings](https://support.google.com/google-ads/answer/6318747?hl=en)

## Portable inputs and examples

- Read [the input contract](references/input-contract.md) when mapping a new export or checking the example's scope and definitions. Copy [the header-only CSV template](assets/input-template.csv) when preparing data; equivalent supplied exports remain acceptable.
- Read [the reusable business context](references/business-context.md) only for business facts or constraints this task needs. Reuse user-supplied facts with their source/date; the template contains no default targets.
- Inspect [the complete fictional input](assets/example-input.csv) with [its example output](references/example-output.md) when learning the output and calculation boundaries. Never use fictional values for a real account.

State whether the result is complete, partial, or blocked for the requested decision. Link material findings to actual supplied rows/sources and separate observed metrics, hypotheses, and estimates. Lead with a short business conclusion, then evidence, uncertainty, and the next measurable check. A data export or installed skill does not authorize account changes.

For the supported arithmetic only, optionally run [the bundled calculator](scripts/marketing_math.py) with Python 3: `python3 scripts/marketing_math.py experiment < calculation.json`. Read its input mapping in the input contract before preparing JSON. It reads JSON, not CSV directly. If Python or the requested method is unavailable, show a reproducible alternative calculation or mark it unsupported; do not report an uncomputed result as verified.