Confirmation Dataset Source Audit v1

Source quality and historical record separation are verified.

UCI Online Retail II is acquired, hashed, structurally inspected, and profiled across the 502,938-row strict prior-period window. Canonical and date-blind record comparisons found zero overlap, and the revised outcome-blind capacity proof supports 48 unique complete-week context slots across three allowed families.

Dataset gateComplete

Record overlap0 rows

Strict rows502,938

New resultsNone

Execution remains blocked. The raw files and detailed context manifest remain local and Git-ignored. The 48-context inventory and 6/15/27 split are frozen under a public SHA-256 commitment, but no question, prompt, model output, annotation target, detector decision, or Confirmation Set result exists.

Decision

Prefer unseen prior-period transactions before changing domains.

Selected window

2009-12-01T00:00:00 inclusive through 2010-12-01T00:00:00 exclusive.

Research role

prospective_temporal_internal_replication. This is not an independent company or cross-domain replication.

Why it fits

Select the strict prior-period portion of UCI Online Retail II for a same-retailer temporal internal replication. It preserves the required business fields through three explicit header aliases, is conditionally suitable after documented quality controls, has zero canonical plus date-blind normalized record overlap with the current source, and supports the revised 48 unique complete-week contexts across three allowed question families without using model outcomes.

Why execution stays blocked

The dataset_source_selected_and_audited gate is complete because acquisition, hashing, workbook structure, strict-window completeness, duplicate/grain handling, cancellation/value rules, monthly coverage, zero normalized historical record overlap, outcome-blind precision review, and revised capacity for 48 unique complete-week contexts are verified. This does not authorize question generation or experiment execution.

External path remains open. Complete Journey is the preferred external-replication shortlist candidate after the corrected protocol is tested on the lower-risk prior-period source.

Acquisition evidence

The immutable file and strict date boundary now reconcile.

Official ZIP

45622418 bytes

572e36277c2390fbfde10664750731e0a86f55e33470d91919085f0408e67bfb

Extracted workbook

45622278 bytes

bcbe73b35f5b7babf197fb0cb983a11f5d9ff929078d4aa53d171b1f2df2e980

Workbook reconciliation

1067371 data rows across 2 sheets; documented instance delta: 0.

Strict window

502938 rows from 2009-12-01 07:45:00 through 2010-11-30 19:35:00.

Schema drift recorded. The workbook uses Invoice, Price, and Customer ID; ingestion must map them to InvoiceNo, UnitPrice, and CustomerID before shared business rules run.

Strict-window quality evidence

Aggregate use is conditionally suitable after documented controls.

Completeness

2,821 descriptions and 100,207 customer IDs are missing. Customer-level questions remain blocked; product and valid-net evidence requires a description.

Duplicate control

6,544 normalized exact duplicate extra rows are removed before aggregation. InvoiceNo + StockCode is not treated as a unique key.

Cancellation control

2,057 negative-quantity rows lack the C prefix. Net revenue retains negative quantities instead of relying on invoice prefix alone.

Reconciled net evidence

492,887 valid net-revenue lines reconcile to GBP 9,266,060.76 across all 12 expected months.

Public boundary. The line-level table remains local and Git-ignored because it contains source invoice and customer identifiers. Read the full quality profile. Read the aggregate-only overlap proof. Read the aggregate-only capacity proof.

Acceptance criteria

Analytical quality, record separation, and aggregate capacity are controlled.

Local evidence confirms file integrity, workbook shape, date coverage, deterministic aliases, completeness, duplicate behavior, business-rule validity, monthly reconciliation, zero canonical plus date-blind historical record overlap, outcome-blind precision review, and capacity for 48 unique complete-week contexts. The context inventory and split are now frozen; exact question-level evidence payload fingerprints remain a next-gate check.

Criterion	Status	Requirement	Current evidence
`row_level_deterministic_gold`	pass_metadata	Row-level evidence supports deterministic gold calculations.	UCI documents transaction-level invoice, stock, quantity, date, price, customer, and country fields.
`documented_time_entity_quantity_value`	pass_metadata	Time, entity, quantity, and value fields are documented.	InvoiceDate, InvoiceNo, StockCode, CustomerID, Country, Quantity, and UnitPrice are documented on the official UCI page.
`completeness_and_duplicates_auditable`	pass_local_profile_with_controls	Source completeness and duplicate handling can be audited.	The 502,938-row strict window has 2,821 missing descriptions, 100,207 missing customer IDs, and 6,544 normalized exact duplicate extra rows. The frozen controls exclude missing descriptions from auditable product/value evidence, keep customer-level questions blocked, and remove only normalized exact duplicates before aggregation.
`permitted_research_use`	pass_metadata	License or terms permit local research and public documentation.	UCI lists CC BY 4.0 and provides DOI 10.24432/C5CG6D.
`no_sensitive_personal_data_required`	conditional_pass	No sensitive personal data is required for the public artifact.	The source uses nominal customer identifiers rather than names or contact details. Public outputs must remain aggregated or use non-reversible study IDs instead of exposing source customer identifiers.
`minimum_48_disjoint_contexts`	pass_outcome_blind_capacity_profile	At least 48 disjoint evidence contexts can be formed without outcome-based filtering.	Fifty observed complete calendar weeks support 48 unique period-to-family slots across three allowed families. Generalized Hall capacity has minimum slack 2 and an independent maximum matching fills 48 of 48 slots; no assignment was retained.

Candidate matrix

Keep internal replication, external replication, and domain transfer separate.

Candidate	Role and decision	Field fit	Strengths	Risks	Recommended use
UCI Online Retail, current rows with new contexts`current_online_retail_new_contexts`	protocol_development_onlynot_selected_for_primary_confirmation	time: strong entity: strong quantity: strong value: strong returns_or_cancellations: strong	already downloaded, cleaned, and reconciled locally exactly matches existing BizHallu metric definitions supports customer, product, country, invoice, quantity, and value evidence	reuses the same transaction population as the exploratory study cannot establish unseen-transaction generalization official metadata says no missing values, while the local audit found 1,454 missing descriptions, 135,080 missing customer IDs, and 5,268 duplicate rows	Use only for protocol development, unit tests, and non-confirmation verifier implementation.
UCI Online Retail II, strict prior-period window`uci_online_retail_ii_prior_period`	prospective_temporal_internal_replicationselected_source_audited_context_manifest_pending	time: strong entity: strong quantity: strong value: strong returns_or_cancellations: strong	same required business fields after deterministic Invoice, Price, and Customer ID header aliases prior periods allow transaction-level temporal separation small enough for the local Windows and 5070 Ti workflow the strict prior-period window contains 502,938 date-valid rows before any outcome-based filtering	same retailer and source lineage, so this remains internal replication the first workbook sheet extends through 2010-12-09, so 22,523 rows in that sheet and all 541,910 rows in the second sheet must remain outside the strict window actual workbook headers differ from UCI's canonical field names for InvoiceNo, UnitPrice, and CustomerID and require an explicit ingestion mapping 100,207 customer IDs and 2,821 descriptions are missing, so customer-level questions stay blocked and product/value evidence requires descriptions 6,544 normalized exact duplicate extra rows require exact-row deduplication before aggregation 2,057 negative-quantity rows lack the C invoice prefix, so cancellation logic cannot rely on prefix alone customer concentration remains blocked because 100,207 Customer IDs are missing; the verified 48-context capacity covers three other families	Use the documented strict-window controls, completed precision review, estimation-only claim amendment, and revised aggregate capacity proof; next freeze only the deterministic 6/15/27 context manifest and seeded split.
84.51 Complete Journey via completejourney`completejourney_external`	external_replication_shortlistpreferred_external_candidate_pending_profile	time: strong entity: strong quantity: strong value: strong returns_or_cancellations: unclear	independent grocery-retail source with household, store, basket, product, quantity, sales value, discount, week, and timestamp fields large enough for context-grouped sampling CC0 package documentation and reproducible programmatic access	returns or cancellation semantics are not documented in the transaction reference sales value and three discount fields require a new accounting metric map full data acquisition relies on the package repository rather than the current Python pipeline	Preferred Phase 2 external replication candidate after a separate acquisition, license-lineage, value-definition, and return-semantics review.
Brazilian E-Commerce Public Dataset by Olist`olist_external`	external_relational_alternativesecondary_external_candidate	time: strong entity: strong quantity: implicit_item_rows value: strong returns_or_cancellations: partial_order_status	real anonymized multi-marketplace order data supports order status, item price, payments, freight, customer, product, seller, and delivery evidence strong operations and supply-management extension	requires multi-table joins and new integrity tests no direct quantity field comparable to Online Retail line quantities cancelled order status is not equivalent to item-level returns or refunds noncommercial share-alike license and Kaggle access add distribution constraints	Keep as a relational e-commerce alternative, especially for operations questions, but do not use as the first confirmation source.
NYC TLC Yellow Taxi Trip Records`nyc_tlc_domain_transfer`	operations_domain_transferdomain_transfer_only	time: strong entity: moderate_zone_vendor quantity: not_retail_quantity value: strong returns_or_cancellations: not_applicable	official public operations data with itemized monetary fields many disjoint monthly contexts and a detailed data dictionary strong JHU operations-analytics narrative	not a retail transaction source and does not preserve current question families very large monthly files increase processing cost TLC explicitly does not guarantee source accuracy or completeness schema changes across years require version-specific controls	Reserve for a separately labeled operations domain-transfer study, not Confirmation Set v1 internal replication.
Future permitted JHU course, capstone, operations, or healthcare source`jhu_future_domain_extension`	future_domain_extensionunavailable_not_auditable	time: unknown entity: unknown quantity: unknown value: unknown returns_or_cancellations: unknown	could align the project with coursework, faculty, or capstone needs	no specific dataset or permission exists yet cannot be audited or selected now must not delay the near-term internal replication	Reopen only after a concrete source, owner, permission, and schema are available.

Metadata conflict matters. The UCI Online Retail page reports no missing values, but BizHallu's local source audit found 1,454 missing descriptions, 135,080 missing customer IDs, and 5,268 duplicate rows. Online metadata alone is therefore not sufficient to clear the new source.

Local acquisition gate

Nine checks are complete; public privacy stays active.

#	Check	Status	Requirement
1	`official_acquisition_and_hash`	completed	Download online_retail_II.xlsx only from the official UCI source, record retrieval date, byte size, and SHA-256, and keep the workbook out of Git.
2	`workbook_structure`	completed	Record workbook sheet names, dimensions, headers, inferred types, and parsing warnings before filtering.
3	`strict_prior_period_filter`	completed	Retain InvoiceDate >= 2009-12-01 and < 2010-12-01; report excluded rows on both sides of the window.
4	`completeness`	completed	Measure null and blank rates overall and by month for every required field, especially Description and CustomerID.
5	`duplicates_and_grain`	completed	Measure exact duplicate rows and candidate invoice-line key duplication; define rather than assume the analytical grain.
6	`business_rule_validity`	completed	Profile negative, zero, and extreme Quantity and UnitPrice values; compare cancellation-prefix rows with negative-quantity rows.
7	`historical_overlap`	completed	Normalize shared fields and prove zero canonical and date-blind record fingerprints against the current Online Retail lineage after the strict cutoff. The public proof reports aggregate counts only.
8	`monthly_coverage`	completed	Report monthly row, invoice, customer, product, country, cancellation, and monetary totals to identify sparse or abnormal periods.
9	`context_feasibility`	completed	Outcome-blind capacity is verified for 6 pilot, 15 development, and 27 confirmation contexts using unique complete calendar weeks across three allowed families. Final evidence fingerprints remain a manifest-freeze check.
10	`public_privacy_boundary`	active	Keep raw customer and invoice identifiers out of public artifacts; use aggregates or study-local opaque identifiers.

Decision rules

Reject the source if independence or deterministic evidence fails.

Accept only if

the official file hash and lineage are recorded
the strict prior-period window has zero exact normalized record overlap with the current source
required fields remain sufficiently complete for deterministic gold calculations
cancellation and value rules can be documented without outcome-dependent exclusions
at least 48 disjoint evidence contexts satisfy the frozen split policy

Reject or redesign if

record overlap remains after the strict cutoff
missingness or duplicate ambiguity prevents deterministic evidence calculations
fewer than 48 eligible contexts can be formed without looking at generated answers
the proposed period split produces materially sparse or incomparable business scopes

Source register

Every desk-audit claim points to a primary or maintainer source.

Source	Type	Checked
UCI Online Retail	official_dataset_page	2026-09-01
UCI Online Retail II	official_dataset_page	2026-09-01
completejourney dataset documentation	maintainer_dataset_documentation	2026-09-01
Complete Journey transactions reference	maintainer_data_dictionary	2026-09-01
Brazilian E-Commerce Public Dataset by Olist	owner_published_dataset_page	2026-09-01
Kaggle dataset metadata API for Olist	official_host_metadata_api	2026-09-01
NYC TLC Trip Record Data	official_agency_dataset_page	2026-09-01
NYC TLC Yellow Taxi Trip Records Data Dictionary	official_agency_data_dictionary	2026-09-01
NYC Open Data FAQ	official_open_data_terms_summary	2026-09-01
`data/processed/data_quality_report.json`	repo_local_audit	2026-09-01

Next authorized action

Validate question templates and gold calculations next; do not generate.

Define and validate deterministic question templates, scope-entity selection rules, gold calculations, question IDs, and question-level evidence payload fingerprints. Do not generate prompts or run the model yet.

Still prohibited: creating confirmation prompts, running Qwen, viewing answer correctness, selecting annotation targets, tuning detectors, or reporting Confirmation Set v1 performance.