Examples
Worked runs and skeletons, from a real public-data run to per-mode templates.
The spp repository ships a set of examples under examples/. Two are full worked runs; the rest are per-mode skeletons that show the expected artifact shape for each task type spp supports.
Worked runs
public-benchmark— the first example backed by a real, fully reproducible run on public data. A TREC (6-class question-type) run ongpt-5-nano, with every iteration’s prompt / eval / discrepancy / auditor trio, the frozen prompt, the single sacred-test read, and a three-way comparison against EvoPrompt and DSPy. Every row, prediction, and prompt is genuine and unredacted. This is the same run documented in the benchmark site’s TREC loop log.hair-loss-relevance— the canonical worked example shipped with spp’s first release. Demonstrates the end-to-end binary-relevance workflow and the failure-cluster taxonomy that motivated the methodology.
Per-mode skeletons
These show the artifact layout and plan.md shape for each task type, so you can see what spp produces before running it on your own data.
| Example | Task mode |
|---|---|
multi-field-extraction |
multi-field structured output (K>1) |
nested-schema |
hierarchical labels via JSON schema |
feature-group-split |
feature-group prompt splitting |
entity-extraction |
structured extraction (variable-cardinality, span-grounded) |
decomposition-pipeline |
prompt decomposition (managed linear pipeline) |
See the methodology measured
For the full evidence that the methodology works — accuracy, cost, token breakdowns, the fairness ledger, and per-task loop logs documenting every iteration and the human-in-the-loop gate exchange — see the dedicated benchmark site: