Calendar-defined periods
Only complete Monday-through-Sunday weeks inside the strict prior window are counted. Partial boundary weeks are excluded before eligibility is measured.
Confirmation Set v1 · Outcome-blind source capacity
The source can support 48 period-disjoint contexts without looking at model outcomes.
Complete Monday-through-Sunday weeks were evaluated under frozen data-quality and business-support rules. At this historical capacity checkpoint, no week was selected, no split was assigned, and no context manifest was created.
Observed complete weeks50
Required context slots48
Maximum matching48/48
Eligible families3 of 4
Capacity proof complete. Three allowed families each have 50 eligible periods for 16 required slots, and the minimum Hall-capacity slack is +2.
Checkpoint versus current state. The capacity proof and precision review preceded selection. A subsequent outcome-blind freeze now fixes 48 contexts and the 6/15/27 split. Question templates, prompts, model outputs, new labels, detector scores, and confirmation results still do not exist.
Frozen context grain
One complete calendar week can serve at most one final family.
Stronger separation
A final period may be assigned to one family only. Context periods would therefore be disjoint across all contexts, not merely across development and confirmation.
No assignment at this checkpoint
The capacity calculation discarded its period-to-slot matching. A later, separately committed procedure performed the one-time manifest freeze.
Complete calendar weeks51
Weeks with source rows50
Empty complete weeks1
Boundary rows excluded20,772
482,166 complete-period rows plus 20,772 excluded boundary rows reconcile to the 502,938-row strict source.
Source support
Even the least-supported observed week clears the frozen family rules.
| Weekly support measure | Minimum | Median | Maximum |
|---|---|---|---|
| Source rows | 3,238 | 8,646.5 | 19,111 |
| Valid net-revenue lines | 3,067 | 8,348 | 18,724 |
| Cancellation or return rows | 48 | 178.5 | 496 |
| Eligible products | 17 | 72 | 206 |
| Eligible countries | 3 | 9 | 15 |
Support statistics describe source capacity, not answer difficulty, correctness prevalence, or detector performance.
Family decision
Three families advance; customer concentration stays blocked.
| Question family | Status | Eligible periods | Required | Slack | Interpretation or block reason |
|---|---|---|---|---|---|
Net-revenue reconciliationnet_revenue_reconciliation_by_period | source-feasible | 50 | 16 | +34 | Supports deterministic gross-positive, negative, and net-revenue reconciliation. |
Product return comparisonproduct_return_rate_comparison | source-feasible | 50 | 16 | +34 | This is a recorded return-to-positive-sales unit ratio, not a customer-linked or original-sale-linked causal return rate. |
Country-product exposurecountry_product_exposure | source-feasible | 50 | 16 | +34 | Supports deterministic country-product concentration and exposure comparisons. |
Customer concentrationcustomer_revenue_concentration | blocked | 0 | 0 | 0 | The source audit records 100207 missing customer IDs and explicitly blocks customer-level questions. Feasibility cannot introduce a complete-case denominator after viewing source capacity. |
No complete-case shortcut. The source has 100,207 rows without Customer ID. The pre-existing source audit blocks customer-level questions, so this profile does not invent a denominator after seeing capacity.
Return metric limit. Product support uses a recorded return-to-positive-sales unit ratio within the same week. It is not an original-sale-linked or causal customer return rate; question wording and metric precision still require the pending review.
Capacity proof
Every family subset has enough unique periods.
| Family subset | Available unique periods | Required unique periods | Slack | Result |
|---|---|---|---|---|
| Net-revenue reconciliation | 50 | 16 | +34 | pass |
| Product return comparison | 50 | 16 | +34 | pass |
| Country-product exposure | 50 | 16 | +34 | pass |
| Net-revenue reconciliation + Product return comparison | 50 | 32 | +18 | pass |
| Net-revenue reconciliation + Country-product exposure | 50 | 32 | +18 | pass |
| Product return comparison + Country-product exposure | 50 | 32 | +18 | pass |
| Net-revenue reconciliation + Product return comparison + Country-product exposure | 50 | 48 | +2 | pass |
The independent maximum matching fills all 48 slots: 2 pilot, 5 development, and 9 confirmation contexts per eligible family. At this capacity checkpoint, the assignment itself was neither retained nor published.
Evidence and privacy boundary
The capacity checkpoint exposed no candidate or selected period.
- Exact-duplicate copies are excluded by the frozen analytical evidence flags.
- Complete calendar periods are row-disjoint; a source row belongs to only one period.
- Historical canonical and date-blind overlap with the exploratory source remain zero.
- The public report contains no candidate period list, invoice, customer, product, country candidate, row fingerprint, selected context ID, or split assignment.
- The local period profile remains Git-ignored and contains no matching assignment from this capacity calculation.
Claim limit: this proves source capacity for a same-retailer temporal internal replication. It does not validate question wording, establish independent labels, estimate confirmation performance, or support external-generalization claims.
Subsequent gate
The authorized manifest freeze is complete.
The outcome-blind precision review rejected the stronger comparison design and revised the plan to 6 pilot, 15 development, and 27 confirmation contexts. The later manifest gate froze exactly that allocation without inspecting model outcomes.
- No question, prompt, Qwen output, annotation target, verifier prediction, detector score, AUPRC, or F1 was created by either checkpoint.
- Two protocol gates are complete and the five downstream execution gates remain pending.
- The next gate is limited to deterministic question templates, gold calculations, and question-level evidence fingerprints.
- A later external dataset is still required for cross-company or cross-domain claims.