Documentation / examples/comparison/publication-brief/README.md

Publication-brief adjacent-pattern comparison

The small compare.ts example sends one business input through two deliberately different Blueprint shapes:

monolithic-publication-brief.yaml performs extraction, classification, summarization, validation, and localization in one Blueprint execution.
compare.ts also invokes the five independently testable Blueprints from the publication-brief reference application.

Build PixieCore, configure any supported provider as usual, and run:

npm run build
PIXIECORE_COMPARISON_SOURCE='Release Orion adds typed exports for operators.' \
PIXIECORE_COMPARISON_LOCALE='en-US' \
npx tsx examples/comparison/publication-brief/compare.ts

The output preserves both results and structural measurements. It does not claim that one shape is automatically more accurate, faster, or cheaper.

Reproducible four-approach benchmark

benchmark.ts extends the comparison to four executable approaches:

Approach	Execution shape	Provider calls per case
`pixiecore-composed`	Five independently testable PixieCore Blueprints composed by host code	5
`monolithic-prompt`	One combined Blueprint through the same PixieCore transport and schema boundary	1
`agent-graph`	One model-planned graph followed by five direct schema-constrained worker calls	6
`code-only`	A deterministic parser for explicitly labeled records	0

All four receive the same four synthetic inputs and produce the same typed publication-brief result. The public dataset uses deterministic field and term assertions. A complete-case score requires every assertion to pass; assertion accuracy also exposes partial correctness so one missing term does not turn an otherwise useful result into an unexplained zero.

Inspect the exact plan without a provider call:

npm run benchmark:adjacent-patterns -- \
  --run-id=local-dry-run --seed=local-review --runs=3 --dry-run

Live execution is manual and refuses to start without explicit cost, key rotation, call-cap, and budget confirmations. Choose the model, run ceiling, and USD ceiling for the reviewed release, then retain the resulting value-free report outside the source tree until it has passed the publication review.

The initial public snapshot intentionally contains no prior real-provider scorecard or report. A published scorecard must identify its dataset, model, run count, pricing source, and reproducible command.

This four-case corpus is intentionally narrow. It does not measure development time or prove that one architecture is universally superior. In particular, the code-only implementation supports the two labeled inputs and deliberately rejects both prose inputs.