Take one warehouse autonomy candidate from simulator evidence to a release decision.

Where the SDK fits

Adapt evidence from the systems you already run.

Reliability Studio does not dictate the simulator, policy framework, or trial store. The integration seam is a narrow adapter that translates records you already produce into typed SDK evidence.

End-to-end path

One candidate, three studies, one bounded decision.

1. 01
   Instrument

   Record the state and metrics that operationally define a failure.

2. 02
   Adapt

   Convert existing trial records into paired perturbations, boundary anchors, and physical/simulator pairs.

3. 03
   Predeclare

   Fix thresholds, estimands, sample design, error allocation, and limitations before inspecting results.

4. 04
   Run

   Execute causal, boundary, and transport studies independently.

5. 05
   Persist

   Write every scientific report as canonical content-addressed JSON.

6. 06
   Decide

   Apply a noncompensatory policy in which any critical blocker stops readiness.

7. 07
   Gate

   Require readiness in CI while retaining the evidence directory when the build is blocked.

8. 08
   Submit

   Send the same immutable reports to Reliability Studio idempotently.

9. 09
   Regress

   Turn independently supported failures into replay capsules and regression packs.

Integration seam

Replace fixture builders with narrow domain adapters.

The executable studies expose build_observations as the seam. Preserve matching identity rather than inferring it: a pair must truly share its declared initial state and random context, and a physical anchor must correspond to its simulator anchor.

def build_observations(plan):
    records = trial_store.load(study_id=plan.plan_id)
    return tuple(
        PairedPerturbationObservation(
            plan_content_digest=plan.content_digest(),
            contrast_id=record.intervention_id,
            pair_id=record.pair_id,
            matched_context_digest=record.initial_state_digest,
            baseline_evidence_digest=record.baseline_digest,
            perturbed_evidence_digest=record.perturbed_digest,
            baseline_failed=record.baseline.minimum_margin_m < 0.0,
            perturbed_failed=record.perturbed.minimum_margin_m < 0.0,
        )
        for record in records
    )

Run the workflow

Produce four reviewable artifacts.

The repository fixtures are deterministic, so the full workflow runs locally and in CI without a simulator. Replace their domain values and builders when adapting the workflow to a real system.

uv sync --package iso-obs

uv run python \
  packages/python-sdk/examples/warehouse_release_workflow.py \
  --output-dir evidence

Read the result

The common condition cannot hide the rare critical one.

The next work is explicit: reduce latency sensitivity, collect fixed-design samples in the unresolved band, resolve or exclude the blackout region, and improve low-friction physics or narrow the supported envelope.

Operationalize

Use readiness in CI and retain the explanation.

With --require-ready, the example exits with status 2 whenever a critical requirement blocks release. Upload the evidence directory even on failure so the exact reports remain reviewable.

uv run python \
  packages/python-sdk/examples/warehouse_release_workflow.py \
  --output-dir "evidence-${GITHUB_SHA}" \
  --require-ready

iso evidence submit evidence/causal-perturbation-report.json \
  --project-id YOUR_PROJECT_ID
iso evidence submit evidence/failure-boundary-report.json \
  --project-id YOUR_PROJECT_ID
iso evidence submit evidence/transportability-report.json \
  --project-id YOUR_PROJECT_ID

Before adapting

Protect the assumptions that make the evidence meaningful.

• System, model, simulator, and environment versions are immutable.
• Failure definitions and thresholds are chosen before analysis.
• Matched comparisons truly share their declared context.
• Critical operating strata cannot compensate for one another.
• Real anchors cover the operating envelope being claimed.
• Learned labels retain prediction and calibration provenance.
• Discovery and confirmation datasets remain separate.
• Reports, replay artifacts, and CI outputs are retained by digest.

Relevant API

The classes connecting this workflow.