1# Decision log — WP3823| id | date | decision | rationale |4|---|---|---|---|5| D-001 | 2026-11-01 | Data = owner-authorised read-only `.backup` of `fabrika.db` (2026-11-01T17:38:30Z); personal tables dropped on the node; no scraping, one control request to the public `/api/stats` | mandate rules 3–4; `robots.txt` disallows `/api/` for crawlers |6| D-002 | 2026-11-01 | Unit of analysis = **published active listing** of an enabled store at the snapshot; variants collapsed to the listing (min price), with variant counts kept as a covariate | the platform's own publication rule; variants are not separate products |7| D-003 | 2026-11-01 | Region recovery: official MERN administrative-region polygons + FSA centroid point-in-polygon for stores with coordinates; registry postal/city fields otherwise; flags `region_source ∈ {directory, site, fsa_map, pip_centroid, registry_postal, unknown}` | 22 % of live stores lack a region; results reported with/without recovered stores |8| D-004 | 2026-11-01 | H6 declared **untestable** at product level; a store-certification contrast is reported as a labelled proxy only | no product-level provenance in the source (mandate H6 clause) |9| D-005 | 2026-11-01 | All price analyses are cross-sectional at the snapshot; H4 exploratory only | no price history table; store onboarding ended 2026-08-24 |10| D-006 | 2026-11-01 | Seed 20261101 for every bootstrap/permutation; outputs hashed in `logs/manifest.json` | reproducibility |11| D-007 | 2026-11-01 | Preregistration frozen: `research/preregistration.md` sha256 `b7cd3f177510af0b29d5a67c84f14562b068c36c789de1f5ea078ef4697e180b` | mandate Phase 3; frozen before any Phase 4 computation |12| A-001 | 2026-11-01 | **Amendment** to preregistration §2 / D-003: region assignment = municipality lookup of the FSA sector label (urban FSAs only, official MERN municipality table, current boundaries) > Fabri-Ka label (directory/site/FSA map) > point-in-polygon of the FSA centroid in the MERN administrative-region layer (V2026-10) > unknown. A point-in-polygon-first variant is the sensitivity sample. | After freezing we found (i) a first attempt with the StatCan 2021 economic-region layer produced spurious Estrie/Montérégie conflicts because it predates the 2021 transfer of Brome-Missisquoi and La Haute-Yamaska to Estrie — the official MERN layer and municipality table are used instead; (ii) 99 apparent conflicts were a double-hyphen spelling artefact; (iii) remaining disagreements (139) are 69 % rural FSAs whose centroids fall near borders, so the directory label is preferred when present. Region recovery adds 116 stores that had no label. |13| A-002 | 2026-11-01 | **Amendment** to preregistration §2 (categories "used as labelled"): the 500-listing audit found the platform's keyword rules agree with annotator consensus on only 46 % of listings (two LLM annotators agree with each other on 95 %). A supervised classifier (OpenAI text-embedding-3-small + one-hot rule label, multinomial logistic regression) is trained on 6 481 listings labelled by claude-sonnet-5-5 and validated on the audit consensus sample (held out). **All category-based results (H1, H2, H4, H7, QDACI, robustness) are reported under both labellings; the model-based labelling becomes the primary one only if its audit accuracy exceeds the rule labelling by at least 20 points; otherwise the preregistered rule labelling stays primary.** | measurement validity; decided before seeing any result under the model labelling |14| D-008 | 2026-11-01 | Title changed from the mandate's provisional « Digital Production » to « Digital Retail Supply » | referee 2: the provisional title contradicts the wording rule D-004 (listings are not production); the mandate labelled the title provisional |15| D-009 | 2026-11-01 | Corrections after the three referee reports (see `reports/referee_responses.md`): H5 wild-cluster bootstrap re-projects the disturbance off the group fixed effects; H3 log-normal fitted by truncated discrete MLE; H1 cell intervals calibrated against the merchant-reshuffle null; H2 bootstrap combines region and rarefaction draws; matching precision computed among consensus pairs; H4 uses all published listings including those inactive at the snapshot; region conflicts defined against the polygon and the 2021 Brome-Missisquoi/Haute-Yamaska transfer applied to platform labels; S_R0 = platform-label-only sample and a point-in-polygon-first variant; a province-wide-sources-only variant replaces the misnamed certification-only one as the answer to directory-driven discovery; V2 pass requires both spectral statistics; H1 status follows the preregistered rule (no « partially supported ») | all re-run; statuses updated in the registry |16