# Phase 0 — Environment, model and source audit (WP38)

Written 2026-11-01. All figures below were measured on the authorised snapshot or on the production node with read-only queries; nothing was modified in production.

# 0.1 Environment

# 0.1.1 Model inventory (official provider endpoints, tiny test calls)

See reports/model_inventory.md / logs/model_inventory.json (checked 2026-11-01T17:42Z). OpenAI GET /v1/models lists 140 IDs; gpt-5.5, gpt-5 and gpt-5-mini answered 200 (8-token budget fully consumed by reasoning, so empty text — expected for reasoning models with a tiny budget); gpt-5.1 returned 400 ("could not finish the message" under the 8-token budget). Anthropic lists 14 IDs; claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5 all answered "OK" in about 1 s. Total cost of the inventory: below 0.002 USD. LLMs will be used only for (a) second annotation of a product-matching validation sample and (b) optional category-reclassification audit; every call will be logged in logs/run_ledger.jsonl with its cost. Update after Phase 2: the category audit showed 46 % precision of the rule labels, so the primary category labelling is a classifier trained on cached model annotations (A-002); every category-based result is also reported under the rule labels. No result depends on a live model call.

# 0.2 Source identification: Fabri-Ka

Fabri-Ka (www.fabri-ka.com) is an aggregator built and operated by the author (Groupe Ka). It does not sell: it discovers Quebec online stores from certification registries, sector associations, regional agri-food brands and marketplaces (deliverables/research_report.md), verifies each domain, classifies the store by origin class A–E with retained textual evidence, and then syncs each store's public catalogue endpoint (Shopify /products.json, WooCommerce Store API, Wix GraphQL, Squarespace ?format=json, Square, Votresite, generic JSON-LD) roughly every 6 hours (PM2 process fabri-ka-sync). Compliance per source is documented in docs/CONFORMITE.md (reviewed 2026-08-19/21): public endpoints only, nothing behind authentication, polite rates, removal on request.

# 0.2.1 Access and authorisation

• Authorised access: the author owns the platform and the database. The research copy is a first-party export; no scraping of fabri-ka.com and no use of its /api/ (which robots.txt disallows for crawlers) beyond one control request to /api/stats on 2026-11-01T17:37Z to date-stamp the public counters (products 390 108, stores_live 1 569, stores_registry 4 082, regions 17). The mandate's "approximately 390 456 products, 1 572 stores" was therefore an interface figure from a slightly earlier moment; the counts drift by a few hundred per 6-hour cycle.
• Snapshot: sqlite3.Connection.backup() of M4M64a:~/fabri-ka/data/fabrika.db on 2026-11-01T17:38:30Z, application commit 689c7bb1…; tables users (2 rows) and inscriptions (1 row) — the only personal data — and the derived FTS index were dropped on the node before transfer. SHA-256 8d963ba9…98fb, 1.24 GB, stored in data/raw/ (git-ignored). The discovery registry stores.json and collections_taxonomy.json were copied with checksums (data/raw/README.md).
• Production untouched: all queries used mode=ro; the WAL file was not modified.

# 0.2.2 Schema (three tables kept)

There is no price-history table. The database upserts by content_hash and keeps only the current price per listing; sync_log records counts, not prices. Consequently every price analysis in WP38 is cross-sectional at the snapshot date; temporal analysis is limited to listing entry/exit (first_seen/last_seen) and store-level sync counts.

# 0.2.3 Temporal coverage

• first_seen runs 2026-08-12 → 2026-11-01 (81 days). 291 125 listings carry first_seen = 2026-08-12 (initial load) and a further 48 820 + 30 173 on 2026-08-19/23 (expansion waves); 1 266 stores were onboarded in ISO week 32 and 328 in week 33; only 15 stores were first observed later. Hence store onboarding ended by 2026-08-24, and product-level entry after that date reflects merchant catalogue changes (new listings, connector fixes) rather than platform expansion — this is the only usable "dynamic" margin and it must be audited connector by connector before any H4 statement.
• Sync cadence: median 6.3 h between successive syncs of a store (p10 1.2 h, p90 8.8 h; 200-store sample, last 30 days). A listing becomes active=0 after 2 consecutive misses (MISS_GRACE = 2, ≈ 12 h). 73 618 published listings are inactive at the snapshot (last seen Aug 6 970, Sep 31 642, Oct 35 222). A disappearance is not a sale (sold-out products usually stay listed with available=0: 48 997 active listings).

# 0.2.4 Field-level quality and missingness (active published listings, N = 390 108)

# 0.2.5 Merchant-level quality

# 0.2.6 Known selection mechanisms (from the platform's own documentation)

Discovery favours certified, association-member and search-indexed businesses; under-represents CMAQ artisans, industrial manufacturers selling direct, Instagram/Facebook-only micro-shops, Nord-du-Québec, Etsy sellers (excluded by Etsy ToS), ASC farms; Wix stores were only connected in a later wave; Ecwid storefronts cannot be connected; Montréal was deliberately topped-up via a dedicated source (acheterquebecois_mtl, modemtl). Connector caps (max_pages 200) truncate the very largest catalogues. These mechanisms are region- and category-correlated and are central to the representativeness audit (Phase 5).

# 0.3 Feasibility verdict by hypothesis

# 0.4 Blockers and decisions carried forward

1. No price history → temporal price analysis blocked (recorded in reports/limitations.md). A prospective collection plan is written but not executed.
2. Region unknown for 22 % of live stores → Phase 2 region-recovery procedure with quality flags; all regional results reported with and without recovered stores.
3. Origin class is store-level → wording throughout: "online assortment of Quebec-based merchants", never "Quebec production".
4. Category labels are rule-based → stratified audit (manual + LLM second annotation, logged cost) to estimate confusion rates and propagate uncertainty.
5. Public counters vs cleaned sample: 390 108 published active listings at the snapshot; the research sample will be smaller after exclusions (non-CAD, missing price, unknown region, duplicates) and the full attrition is reported in reports/sample_flow.md.