Standardized Ingestion Reference Cases

These deterministic, in-repository synthetic examples demonstrate format interoperability, not domain-specific modelling: no field names, strategies, units, or transformations are inferred. They contain no third-party source data and make no claim about external data rights, live provider availability, or a real-world decision recommendation.

Community	Croissant	Frictionless	Direct normalized input	Explicit method
Machine learning	canonical-decision.croissant.json	canonical-decision.datapackage.json	Explicit Arrow table and binding	EVPI on model-selection net benefits.
Engineering and operations	cost-outcome-decision.croissant.json	cost-outcome-decision.datapackage.json	Explicit Arrow table and bindings	EVPI at a declared WTP; CEAF over paired cost/outcome PSA rows at declared thresholds.
Business	canonical-decision.croissant.json	canonical-decision.datapackage.json	Explicit Arrow table and binding	EVPI on investment-strategy net benefits.

The fixture records two strategies, strategy_a and strategy_b, and binds them explicitly as A and B:

VOIBinding(
    role="net_benefit",
    table_id="samples",
    field_ids=("strategy_a", "strategy_b"),
    strategy_names=("A", "B"),
)

Run the complete reproducible walkthrough from a source checkout:

uv run python examples/standardized_ingestion/reference_cases.py

The output is a machine-readable case catalogue. Every community has croissant, frictionless, and direct results; the optional dataframe result is an additional SDK-consumer check. The runner fails if a required surface reaches a different EVPI, a provider receipt differs from the pinned CSV digest, or the Arrow schema changes. The dataframe result exercises from_dataframe(..., allow_copy=False) with the explicit zero-copy requirement; it does not replace the direct normalized input.

To inspect the descriptor-backed forms before calculating, run:

voiage ingest validate tests/fixtures/standardized_ingestion/canonical-decision.croissant.json
voiage ingest inspect tests/fixtures/standardized_ingestion/canonical-decision.datapackage.json
voiage ingest validate tests/fixtures/standardized_ingestion/cost-outcome-decision.croissant.json
voiage ingest inspect tests/fixtures/standardized_ingestion/cost-outcome-decision.datapackage.json

inspect is deliberately metadata-only; it does not materialize a resource. validate materializes the declared local CSV and reports its receipt, pinned digest, retained governance metadata, and data-quality evidence. Binding resolution remains explicit in Python rather than inferred from the descriptor.

Cost/outcome derivation case

A second synthetic case exercises a different decision path: it records costs and outcomes rather than precomputed net benefit. Explicit, aligned strategy bindings derive net benefit at 20,000 currency units per outcome unit. The same case is materialized through Croissant, Frictionless, and a direct Arrow table, and has an EVPI of 20 / 3 in every surface.

from examples.standardized_ingestion.reference_cases import (
    run_cost_outcome_reference_cases,
)

assert run_cost_outcome_reference_cases()["ml"] == 20.0 / 3.0

The pinned sources, digests, binding details, and willingness-to-pay value are recorded in cost-outcome-decision.manifest.json beside the fixture files. The scripts deliberately report a fixture as deterministic synthetic data, rather than presenting a substitute for a rights clearance or live provider receipt.

Long-layout equivalence

long-decision records the same deterministic two-strategy samples as the canonical wide fixture, but declares a sample identifier, strategy label, and one net-benefit value per row. Its long binding explicitly names all three fields and fixes the strategy order; no column or label is inferred. The Croissant, Frictionless, and direct Arrow representations must each normalize to the same ValueArray as the wide fixture before EVPI is calculated.

from examples.standardized_ingestion.reference_cases import run_long_reference_cases

long_case = run_long_reference_cases()
normalized = long_case["normalized_net_benefit"]
assert normalized["croissant"] == normalized["frictionless"] == normalized["direct"]

This evidence covers declared long net-benefit rows only. It does not turn an arbitrary long table into a cost/outcome or perspective-specific dataset.

CEAF equivalence

The engineering fixture also supplies paired cost and outcome PSA rows for the same two strategies. It can therefore be converted to net benefit at the fixed WTP thresholds 10,000, 20,000, and 30,000 and passed to the existing CEAF method. This is intentionally narrower than an EVSI, EVPPI, or ENBS claim: those methods need conditional-model or study-design inputs that these fixtures do not supply.

from examples.standardized_ingestion.reference_cases import (
    run_cost_outcome_ceaf_reference_cases,
)

ceaf = run_cost_outcome_ceaf_reference_cases()
assert ceaf["croissant"] == ceaf["frictionless"] == ceaf["direct"]
assert ceaf["croissant"]["optimal_strategy_names"] == ("B", "B", "B")

The returned records also include the normalized net-benefit samples at WTP 20,000. Equality across Croissant, Frictionless, direct Arrow, and the optional DataFrame surface is checked before the example returns.

Method applicability matrix

Community	EVPI	CEAF	EVPPI / EVSI / ENBS	Shape coverage
Machine learning	Evaluated on explicit wide and declared long net-benefit samples.	Not applicable: no cost/outcome WTP surface.	Not applicable: no conditional or study-design inputs.	Wide and declared long net-benefit rows; no perspective-split fixture.
Engineering and operations	Evaluated at WTP 20,000.	Evaluated on paired cost/outcome PSA rows at WTP 10,000, 20,000, and 30,000.	Not applicable: no conditional or study-design inputs.	Cost/outcome wide rows and declared long net-benefit rows; no perspective-split fixture.
Business	Evaluated on explicit wide and declared long net-benefit samples.	Not applicable: no cost/outcome WTP surface.	Not applicable: no conditional or study-design inputs.	Wide and declared long net-benefit rows; no perspective-split fixture.

This matrix is a fixture-support boundary, not a statement that the omitted methods or shapes are unsupported by voiage generally. It prevents these small deterministic examples from being presented as evidence for inputs they do not contain.

The input descriptors and their SHA-256 digests live in tests/fixtures/standardized_ingestion/. Use canonical-decision.manifest.json to verify exact source artifacts before reproducing a result.