Confirmation Set v1 · Source QA

The prior-period window is usable only after explicit controls.

The 502,938-row strict historical window passes core schema, date, type, monthly coverage, and revenue reconciliation checks. Missing descriptions, optional customer IDs, exact duplicates, nonpositive prices, and cancellation encoding still require frozen analytical controls.

Strict rows502,938

Months present12 / 12

Missing CustomerID19.92%

Duplicate extra rows1.30%

Controlled net revenueGBP 9,266,060.76

Quality decision: conditionally suitable for aggregate business analysis after documented controls.

The dataset gate is still pending. This profile does not prove historical non-overlap, 36-context feasibility, split independence, or confirmation performance.

Dataset and grain

A complete twelve-month transaction window, without a stable source primary key.

The intended analytical grain is one normalized source invoice-line row before exact-row deduplication. InvoiceNo + StockCode is not unique; the local-only lineage key is source_sheet + source_row_number.

Invoices27,728

Stock codes4,621

Descriptions4,624

Customers observed4,336

Countries40

Completeness

Core transaction fields are complete; two analytical fields are not.

Canonical field	Null	Missing total	Missing rate
`InvoiceNo`	0	0	0.00%
`StockCode`	0	0	0.00%
`Description`	2,821	2,821	0.56%
`Quantity`	0	0	0.00%
`InvoiceDate`	0	0	0.00%
`UnitPrice`	0	0	0.00%
`CustomerID`	100,207	100,207	19.92%
`Country`	0	0	0.00%

Control: CustomerID remains optional for month, country, product, and cancellation questions. Description is mandatory for product-name and valid-net evidence.

Duplicates and grain

Only normalized exact rows are removed before aggregation.

Candidate grain	Extra rows	Extra rate	Affected rows	Groups	Max group
Raw exact row	6,544	1.30%	12,666	6,122	6
Normalized exact row	6,544	1.30%	12,666	6,122	6
InvoiceNo + StockCode candidate key	12,740	2.53%	24,280	11,540	10

InvoiceNo plus StockCode is not unique. Preserve source-row grain locally and remove only repeated normalized exact rows before aggregation.

Cancellation and value controls

Negative quantity is broader than the invoice-prefix signal.

Cancel-prefix rows9,877

Negative-quantity rows11,933

Negative without C2,057

C without negative qty1

Return/cancel rows after dedup11,910

Metric	Value
Gross positive revenue	GBP 9,834,505.21
Negative revenue	GBP -568,444.45
Net revenue	GBP 9,266,060.76
Gross + negative - net	GBP 0.00
Merchandise net revenue	GBP 9,157,800.84

Frozen analytical policy

Cleaning rules are explicit and auditable.

Layer	Rule
Exact Duplicate	keep first normalized exact row across the eight canonical business fields
Positive Sales	deduplicated, no cancel prefix, Quantity > 0, UnitPrice > 0, Description present
Valid Net Revenue	deduplicated, UnitPrice > 0, Description present; negative quantities and cancel invoices retained
Merchandise Net Revenue	valid net revenue plus the existing no-digit and special stock-code exclusion

Tail policy: IQR and percentile flags are diagnostic only. Large quantities, prices, or adjustments are not deleted automatically because that would introduce outcome-dependent filtering.

Monthly coverage

Every frozen month is present and reconciled.

Month	Rows	Invoices	Customers	Stock codes	Countries	Negative qty	Duplicate extras	Net revenue
`2009-12`	45,228	2,330	1,045	3,141	24	1,105	506	GBP 796,648.50
`2010-01`	31,555	1,633	786	2,809	19	814	321	GBP 622,479.50
`2010-02`	29,388	1,969	807	2,971	19	1,023	330	GBP 531,265.37
`2010-03`	41,511	2,367	1,111	2,928	23	1,034	526	GBP 763,247.24
`2010-04`	34,057	1,892	998	2,635	23	701	410	GBP 641,521.05
`2010-05`	35,323	2,418	1,062	2,790	20	1,352	415	GBP 613,270.72
`2010-06`	39,983	2,216	1,095	2,784	24	933	514	GBP 677,073.87
`2010-07`	33,383	2,017	988	2,648	21	822	391	GBP 617,365.48
`2010-08`	33,306	1,877	964	2,727	23	666	367	GBP 654,774.39
`2010-09`	42,091	2,375	1,202	2,755	28	905	484	GBP 851,105.96
`2010-10`	59,098	2,965	1,577	2,903	27	1,159	849	GBP 1,080,611.48
`2010-11`	78,015	3,669	1,683	3,137	29	1,419	1,431	GBP 1,416,697.20

Findings and controls

Known data issues are documented rather than hidden.

mediumhigh confidence

Optional Customer Id Missingness

Risk: Customer-level questions and customer-count comparisons may be biased if missing IDs are treated as observed customers.

Control: Keep CustomerID optional for aggregate month, country, product, and cancellation analyses; prohibit customer-level confirmation questions until a separate coverage rule is frozen.

mediumhigh confidence

Missing Product Descriptions

Risk: Rows without descriptions cannot support auditable product-name claims.

Control: Exclude missing-description rows from product and valid-net evidence while retaining their counts in the quality report.

mediumhigh confidence

Exact Duplicate Source Rows

Risk: Uncontrolled duplicates would overstate quantities, gross revenue, cancellation impact, and ranking values.

Control: Use first-occurrence exact-row deduplication on the eight normalized canonical fields before revenue aggregation.

mediumhigh confidence

Cancellation Encoding Is Not Prefix Complete

Risk: Using only InvoiceNo prefix C would miss negative-quantity adjustments or returns.

Control: Retain negative quantities in net revenue and define cancellation/return review rows as cancel-prefix OR negative-quantity after exact deduplication.

mediumhigh confidence

Nonpositive Prices Require Control

Risk: Nonpositive prices do not represent ordinary merchandise revenue and can distort value-based questions.

Control: Require UnitPrice > 0 for positive-sales and valid-net evidence; retain excluded-row counts for auditability.

lowhigh confidence

Long Tailed Numeric Values

Risk: Large legitimate wholesale orders and manual adjustments can dominate rankings; automatic outlier deletion would create outcome-dependent filtering.

Control: Do not remove tail values automatically. Review only evidence contexts selected by frozen, outcome-blind rules.

Research boundary

Quality profiling does not authorize model execution.

No current-source overlap comparison was performed in this step.
No candidate contexts or split assignments were generated.
No prompt, Qwen answer, annotation target, verifier output, or detector metric was created.
The next authorized action is only the normalized record-overlap proof.