OOF Campaign Fixtures

This document defines the first reproducible OOF campaign fixtures owned by dag-ml. The fixtures are intentionally tiny and identity-heavy: they prove fold and prediction semantics before any real model execution exists.

Fixture Set

Directory: examples/fixtures/oof_campaign/

Fixture	Purpose
uc6_oof_success_predictions.json	UC6-shaped positive stacking case
uc11_train_prediction_refusal.json	UC11-shaped leakage refusal case

Shared Sample Universe

Use six sample ids and three folds:

Fold	Validation samples	Train samples
fold0	S001, S004	S002, S003, S005, S006
fold1	S002, S005	S001, S003, S004, S006
fold2	S003, S006	S001, S002, S004, S005

Rows inside prediction blocks should be deliberately shuffled to force joins by sample_id, never by input row position.

UC6 Success

Required top-level fields:

Field	Shape
fold_set	fold assignments by absolute sample ids
join_policy	allow_train_predictions_as_features=false, join_on="sample_id"
requested_sample_order	["S001", "S002", "S003", "S004", "S005", "S006"]
prediction_blocks	three validation-only producers

Prediction block fields:

Field	Requirement
prediction_id	stable fixture id
producer_node	e.g. branch:b0.model:pls
partition	exactly validation
fold_id	fold0, fold1, or fold2
sample_ids	validation samples for that fold, possibly shuffled
values	one regression column per sample
target_names	["y"]

Assertions:

• join succeeds with default safe policy;

• output sample order equals requested_sample_order;

• output columns are producer namespaced;

• each sample has exactly one validation prediction per producer;

• no train/leakage flags are present.

UC11 Refusal

Required difference from UC6:

• at least one prediction block has partition="train";

• join_policy.allow_train_predictions_as_features=false;

• join_policy.include_partitions=["train", "validation"].

Assertions:

• execution refuses before building meta features;

• error kind is structured OOF leakage;

• payload includes node_id="merge:pred";

• payload lists every train violator with producer, partition and fold id;

• remediation text tells the user to use validation-only OOF predictions or to explicitly opt into the unsafe policy.

Boundary

dag-ml owns folds, prediction blocks, OOF joins, leakage refusal and campaign fingerprints. dag-ml-data owns schema/model-input/adapter-plan fixtures only.

Stacking REFIT Coverage Contract

The runtime distinguishes three stacking REFIT outcomes:

Case	Required policy	Outcome
full validation-OOF coverage for the refit sample universe	default require_full_coverage	meta-model REFIT may consume OOF
incomplete but otherwise well-formed validation OOF	cv_only or skip_refit_on_incomplete_oof on metadata.stacking_oof_refit_contract.policy	stacking REFIT is skipped
incomplete OOF without explicit policy, train/final/test prediction input, missing fold ids, unknown folds, fold/sample mismatch or duplicate validation rows	none	rejected with a stable cause such as partial_oof_without_policy

D8 Conformance Pack

docs/contracts/conformance_pack.v1.json pins canonical digests for uc6_oof_success_predictions.json and uc11_train_prediction_refusal.json. D8 uses them for the stacking_oof_contract.v1 and row_vs_sample_selection_mismatch.v1 scenarios, so changes to these fixtures must update the pack and rerun python3 scripts/validate_contracts.py.