Guides
Write An RCA Report
Produce a final explanation that cites trace evidence, replay policy, fixture provenance, diffs, metrics, effects, and validity labels.
An isplay RCA report should be short enough to read and specific enough to reproduce. It should separate observed facts from replay-backed claims.
Required Sections
| Section | Content |
|---|---|
| Executive finding | One bounded conclusion and confidence/validity caveat. |
| Scope | Project ID, run ID, branch/replay/experiment IDs, adapter/runtime. |
| Observed baseline | What happened in the captured run. |
| Tested hypotheses | Branches or experiment arms and their interventions. |
| Replay policy | Model, tool, drift, max steps, temperature. |
| Evidence | First divergence, diffs, metrics, fixture use, effects. |
| Validity | Labels, comparability, low-N limitations. |
| Next actions | Further experiments, instrumentation gaps, or remediation. |
Template
Finding:
<bounded conclusion>
Scope:
Project: <projectId>
Base run: <runId>
Branch/replay/experiment: <ids>
Adapter/runtime: <name and version if known>
Observed baseline:
<facts from captured events and context inventory>
Hypothesis tested:
<intervention and target>
Policy:
model=<policy>, tool=<policy>, drift=<policy>, maxSteps=<n>
Evidence:
First divergence: <event/step>
Diffs: <summary>
Metrics: <summary>
Fixtures: <fixture IDs, provenance, dependency>
Effects: <ranked effect and status>
Validity:
Labels: <labels>
Comparability: <value>
Caveats: <low-N, fixture, nondeterminism, unsupported areas>
Next actions:
<specific follow-up>Wording Rules
| Use | Avoid |
|---|---|
| "observed in the base run" | "caused" for base trace only |
| "under recorded replay" | "would always" |
| "sensitive to analyst fixture" | hiding fixture dependence |
| "inconclusive with one trial" | statistical certainty from one run |
| "non-comparable" | forcing a conclusion from drift |
Final Check
Before sending the report, confirm it includes:
- IDs.
- Policy.
- Fixture provenance.
- First divergence.
- Metrics.
- Effects.
- Validity labels.
- Concrete next action.