docs/REPRODUCIBILITY.md

Reproducibility and Validation for the Artifact Mirror

This Hugging Face dataset is a lightweight mirror, not the source-code archive. It supports inspection and validation of released reports, strict-JSON result artifacts, schemas, provenance, and citation metadata.

Local artifacts

• results/move1/MOVE1_REPORT.md — canonical fact-locked report;
• results/move1/RESULT_layerA.md — Norman and Tahoe Layer-A details;
• results/move1/MOVE1_3TIER_RESULT.md — scoped three-tier cost sweep;
• results/move1/EXPERIMENT_SELECTION_RESULT.md — goal-dependent result;
• results/move1/RETIRED_ANALYSES.md — removed claims and reasons;
• results/move1/layerA_norman.json and layerA_tahoe.json — strict JSON;
• artifact_manifest.json — HF-specific local artifact and source-link map.

Strict JSON checks

python3 -m json.tool artifact_manifest.json >/dev/null
python3 -m json.tool schemas/artifact_manifest.schema.json >/dev/null
python3 -m json.tool schemas/layerA_norman.schema.json >/dev/null
python3 -m json.tool schemas/layerA_tahoe.schema.json >/dev/null
python3 -m json.tool results/move1/layerA_norman.json >/dev/null
python3 -m json.tool results/move1/layerA_tahoe.json >/dev/null

Reject non-standard JSON constants:

python3 -c 'import json,pathlib; [json.loads(p.read_text(), parse_constant=lambda x: (_ for _ in ()).throw(ValueError(x))) for p in map(pathlib.Path,["artifact_manifest.json","results/move1/layerA_norman.json","results/move1/layerA_tahoe.json"])]'

JSON Schema checks

python3 -m pip install jsonschema
python3 - <<'PY'
import json
from pathlib import Path
from jsonschema import Draft202012Validator

pairs = [
    ("artifact_manifest.json", "schemas/artifact_manifest.schema.json"),
    ("results/move1/layerA_norman.json", "schemas/layerA_norman.schema.json"),
    ("results/move1/layerA_tahoe.json", "schemas/layerA_tahoe.schema.json"),
]
for instance_path, schema_path in pairs:
    instance = json.loads(Path(instance_path).read_text())
    schema = json.loads(Path(schema_path).read_text())
    Draft202012Validator.check_schema(schema)
    Draft202012Validator(schema).validate(instance)
print("schema validation OK")
PY

Source-code reproduction

Analysis code is maintained in jang1563/causalatlas:

• eval/move1_layerA*.py, move1_gears_cv.py, and move1_synergy_cv.py for Layer A;
• eval/move1_3tier_*.py for the three-tier arm;
• eval/move1_replogle_prep.py and move1_expsel*.py for experiment selection;
• scripts/validate_public_release.py for full path, link, identity, and stale-claim validation.

The source repository requires upstream Norman, Replogle, GEARS, and Tahoe/Arc inputs. This mirror does not support full reruns by itself.

Evidence boundaries

• The Norman JSON contains the original split and threshold sweep; the held-out CV regime summary is reported in RESULT_layerA.md and generated in the source repository.
• Large CV predictions, LLM runs, and experiment-selection intermediates are not mirrored.
• Exploratory competence/verification scripts remain source-repository provenance only and do not support public claims.

Mirror invariants

• no raw third-party data or large intermediate predictions;
• strict JSON and valid schemas;
• every local Markdown link resolves;
• one public identity, JangKeun Kim;
• final report and retired-analysis registry remain consistent.