1# Limitations ledger — WP3823| id | limitation | consequence | handling |4|---|---|---|---|5| L-01 | No price history in the source (current price per listing only) | no temporal price analysis; H5 is cross-sectional | declared ex ante (D-005); prospective collection plan in the paper |6| L-02 | Listings are posted offers, not sales; disappearance ≠ sale; sold-out listings remain published | no demand, revenue or employment inference | wording rule; availability flag reported |7| L-03 | Merchant location ≠ product origin; origin class is a store-level prior | H6 untestable; "production" never inferred | D-004; descriptive certification contrast labelled as not H6 |8| L-04 | 229 live stores (71 692 listings, 19 % of the assortment sample) have no recoverable location, including the largest catalogues | regional statistics describe located merchants only | reported as "Inconnue"; excluded from inference; coverage audit |9| L-05 | Region assignment combines three sources (municipality lookup, platform label, FSA centroid); rural FSA centroids are imprecise near borders | some misassignment near borders (30 flagged conflicts) | A-001; sensitivity with PIP-first and conflict-free samples |10| L-06 | Rule-based category labels have 46 % precision against annotator consensus | category-based statistics under rule labels are noisy | A-002 classifier validated on a held-out audit; all category results under both labellings; confusion-reassignment robustness |11| L-07 | Annotators are LLMs (two independent models) and adjudication was done by the research team, not by independent human coders | label validity rests on inter-model agreement (95 %) and taxonomy definitions | disclosed; disagreement files published |12| L-08 | Discovery is a non-probability process favouring certified/association merchants; coverage per 1 000 businesses varies six-fold across regions | regional comparisons confound economy and discovery | representativeness audit; certification-source-only and no-Montréal variants |13| L-09 | 17 regions → very low power for H2 and H7-V3; ECI unstable at 17×20 | H2 inconclusive; complexity not validated | preregistered criteria; components only |14| L-10 | Matching without GTIN; tier precision 0.98–1.00 among consensus pairs (lower bounds 0.90–0.92 counting non-consensus as failures) for A, B | few false matches in H5 | precision reported; comparable-variant restriction; sensitivity with C |15| L-11 | Region effects in H5 identified only through groups spanning regions; one store = one region | wide intervals; cannot separate region from store | collinearity note; wild cluster bootstrap; stores per region reported |16| L-12 | Connector caps (max 200 pages) truncate the largest catalogues; Wix/Ecwid coverage incomplete | catalogue sizes of the largest stores under-measured | noted; trimmed scheme T |17| L-13 | Single snapshot; onboarding ended 2026-08-24; 10 weeks of listing dynamics | H4 exploratory only | temporal holdout; placebo |18| L-14 | Power-law vs log-normal comparison uses a continuous log-normal on the discrete tail | approximate model comparison | stated; not a hypothesis |19| L-15 | External covariates mix vintages (ISQ 2025/2024, Census 2021 on 2021 ER boundaries for Estrie/Montérégie) | small concordance error for density | ISQ columns used wherever available; manifest records vintages |20| L-16 | H5 geographic joint test is dominated by one region/one reseller (Outaouais, 65 of 164 Wald points) and is sensitive to the bootstrap implementation | no claim about regional price levels | corrected bootstrap, LOO decomposition, status 'not supported' |21| L-17 | Audit sample (500) was drawn before the mirror-domain rule: one listing belongs to a later-excluded sibling store; 92 % of audit stores also appear in the classifier training set (no merchant holdout) | audit accuracy may be slightly optimistic for unseen merchants | disclosed; labelled CSVs committed and hashed |22