# AlgoReport: automation reliability review

SYNTHETIC DEMONSTRATION — no customer data or measured savings.

Observed attempt starts: 2026-10-08T00:00:00+00:00 through 2026-10-14T01:38:00+00:00.

| Metric | Observed value |
| --- | --- |
| Runs / attempts | 100 / 120 |
| Runs with recorded success | 90 / 100 (90.00%) |
| Runs without recorded success | 10 |
| Recorded attempt cost (USD) | $2.100000 |
| Cost of attempts after the first in each run (USD) | $0.300000 |
| Cost of runs with no recorded success (USD) | $0.200000 |
| Attempt latency p50 / p95 (ms) | 1500 / 30000 |

## Three next investigations

1. **Investigate 10 runs with no recorded success.** Review upstream error categories and confirm export completeness. A missing success may reflect a failure, an unfinished run, or an export boundary; these logs do not establish cause.

2. **Review retries: $0.300000 across 20 runs.** Retries recovered 10 runs. Their full cost is not a savings estimate. There are 0 attempts starting after an earlier success started; inspect timestamps and idempotency before labeling these duplicates or changing retry policy.

3. **Inspect the latency tail: p95 30,000 ms.** The export contains 10 timeout attempts. Compare slow traces with agreed service targets before changing timeouts. Percentiles describe attempts, not end-to-end runs.

## Definitions and limits

One run is one pseudonymous run_id; one row is one terminal attempt. A run has
recorded success if any included attempt says success; this does not assess output
quality, delivery to a customer, or business correctness. Attempts are ordered by
start time, which does not establish completion order or prove duplicate effects.
Retries mean all attempts after the first, including necessary recovery attempts.
Cost categories overlap and must not be added together. USD is copied from the
provided cost records; it is not an independently verified provider invoice.
Latency uses the nearest-rank percentile across individual attempts, including
errors and timeouts. No missing traffic is inferred or filled in.

These six fields cannot verify export completeness, root causes, output quality,
end-to-end latency, legal compliance, or achievable savings. Confirm complete runs
and the collection window with the owner. Quality evaluation requires a separate
agreed rubric and approved content; it is outside this metrics-only pilot.

## Reproducibility

Analyzer version: 1.0. Input SHA-256: `5138d77f166d5362045c4cd4dfa1857106c711b5e059b971e89b5786465eb94e`.