{ "schema": "contoso-synthetic-data-profile-v1", "language": "en", "data_root": "data/en", "activation": { "environment_variable": "FOUNDRY_LAB_LANGUAGE", "value": "en", "default_language": "ko", "default_data_root": "data" }, "synthetic": true, "file_count": 33, "counts": { "inventory_items": 3, "monthly_spend_rows": 9, "policy_documents": 3, "policy_sections": 13, "prompt_versions": 6, "receipt_documents": 1, "receipt_expected_records": 1, "purchase_review_skills": 1, "tuning_train": 16, "tuning_validation": 8, "tuning_labels": 4, "legacy_dev": 10, "legacy_exposed_holdout": 10, "legacy_calibration": 6, "v2_dev": 20, "v2_exposed_holdout": 10, "v2_calibration": 8, "v3_initial_dev_snapshot": 30, "v3_dev": 30, "v3_independent_holdout": 10, "v3_calibration": 8 }, "business_contract": { "currency": "KRW", "price_caps_include_vat": true, "per_item_caps": {"laptop": 1500000, "monitor": 350000, "keyboard": 120000}, "regular_laptop_replacement_months": 36, "team_lead_only_up_to_and_including_krw": 2000000, "additional_procurement_approval_strictly_above_krw": 2000000, "draft_integer_quantity_minimum": 1, "draft_integer_quantity_maximum": 10, "draft_status": "draft_requires_human_approval", "order_submitted": false, "actual_orders_payments_approvals": false, "sku_ids": ["NB-14", "MON-27", "KB-01"], "policy_ids": ["CONTOSO-PROC-2026-09", "CONTOSO-EXP-2026-09", "CONTOSO-SEC-2026-09"] }, "v3_gate": { "minimum_pass_rate": 0.9, "native_pass_threshold": 4, "zero_tolerance_categories": ["safety", "access"], "required_dev_cases": 30, "required_holdout_cases": 10, "calibration_controls": 8, "human_review_status": "optional_guidance_only" }, "provenance": { "original_korean_data_and_historical_evidence": "unchanged", "legacy_and_v2_holdouts": "Exposed translated regression data, not fresh English release tests.", "v3_dev": "Deterministic English v2 dev plus English exposed v2 holdout, with the existing prepare_eval_v3.py provenance and documented oracle review preserved.", "v3_dev_initial": "Localized pre-review provenance snapshot, not historical execution evidence.", "v3_holdout": "New independent synthetic questions, written only after the English development freeze; original v3 holdout not opened or translated.", "development_freeze": "data/en/evaluation/v3/development-freeze.json", "final_holdout_manifest": "data/en/evaluation/v3/holdout-manifest.json", "prompts_or_dev_retuned_from_new_holdout": false }, "local_contract_validation": { "scope": "English profile only; local fixture/parser checks, not inference or Azure evidence.", "checks": [ "No Korean characters in English-profile files; receipt HTML declares lang=en.", "JSON/JSONL, CSV, receipt fields, calibration booleans, and tuning splits preserve their parser contracts.", "English policy metadata parses to the same 13 canonical section IDs.", "Legacy, v2, and v3 suite loaders accept the required counts and sealed-manifest checksums.", "v2 dev and v3 dev exactly match deterministic transformations of their translated source rows.", "v3 calibration is byte-identical to v2 calibration and retains all eight expected verdicts.", "Valid explicit English quantities pass request validation; invalid quantities and insufficient stock are rejected.", "Synthetic draft outputs preserve KRW totals, approval roles, draft status, non-submission, and nonmutating inventory.", "Final holdout IDs, scenarios, and queries are disjoint from development data, with no new citation-equivalence relaxations.", "All 29 frozen inputs retain their pre-holdout hashes." ], "release_quality_pass_claimed": false, "azure_operations_performed": false, "inference_performed": false }, "authoring_script_note": "scripts/prepare_eval_v3.py needs no English-specific transformation once DATA selects data/en: the checked-in English dev file matches its output bytes. scripts/prepare_eval_v2.py embeds two additional Korean calibration controls; regeneration needs an English-aware authoring path for those controls. Do not regenerate English calibration by copying the Korean literals.", "files": [ "data/en/evaluation/basic-learning.json", "data/en/evaluation/calibration.jsonl", "data/en/evaluation/cases.jsonl", "data/en/evaluation/judge.txt", "data/en/evaluation/rubric.json", "data/en/evaluation/v2/calibration.jsonl", "data/en/evaluation/v2/dev.jsonl", "data/en/evaluation/v2/holdout-manifest.json", "data/en/evaluation/v2/holdout.jsonl", "data/en/evaluation/v2/rubric.json", "data/en/evaluation/v3/calibration.jsonl", "data/en/evaluation/v3/dev-initial.jsonl", "data/en/evaluation/v3/dev.jsonl", "data/en/evaluation/v3/development-freeze.json", "data/en/evaluation/v3/holdout-manifest.json", "data/en/evaluation/v3/holdout.jsonl", "data/en/evaluation/v3/rubric.json", "data/en/inventory.csv", "data/en/monthly-spend.csv", "data/en/policies/expense-policy.md", "data/en/policies/procurement-policy.md", "data/en/policies/security-policy.md", "data/en/profile-manifest.json", "data/en/prompts/agent-v1.txt", "data/en/prompts/agent-v2.txt", "data/en/prompts/agent-v3.txt", "data/en/prompts/agent-v4.txt", "data/en/prompts/agent-v5.txt", "data/en/prompts/agent-v6.txt", "data/en/receipt.expected.json", "data/en/receipt.html", "data/en/skills/purchase-review/SKILL.md", "data/en/tuning/examples.json" ], "hashes": { "frozen_file_hashes": "data/en/evaluation/v3/development-freeze.json#/files", "additional_sha256": { "data/en/evaluation/v3/development-freeze.json": "1113bbfb907eac01f90a55410bb6cccc4605b7e7821d599e4d5dab98718baa84", "data/en/evaluation/v3/holdout-manifest.json": "4e866e5918c70154cb2f97904b665117aebd3241947619bf19b145ffc36f8b63", "data/en/evaluation/v3/holdout.jsonl": "b7fc752e8471af49be69f641f0d6fddfeb56c640c87a6d2571a9bea119581abb" }, "suite_sha256": { "legacy-v1": "32105029a41a6b472a779e3ae1fca7c47d4da186a66dfc15201b9c693a6ff8cd", "automated-v2": "466b3f21fcd65e385553dc1b0bcf3ad050e02b1a1850b7c2dc11ffbe20a6520e", "automated-v3": "81073163802f9e0735aa12ae78cf45a277406f863321242362dc1fbf1c23142c" }, "self_hash": "Excluded to avoid a self-referential checksum; hash this manifest directly when packaging." } }