Add listing_enrichment (typesafe/jev-1.13 labels, 19 Sep 2026) and README section

README.md

@@ -36,6 +36,8 @@ configs:
     data_files: category_count_entries.parquet
   - config_name: fetch_log
     data_files: fetch_log.parquet
+  - config_name: listing_enrichment
+    data_files: listing_enrichment.parquet
 ---
 
 # DrugHub Market Snapshot, September 2026
@@ -44,7 +46,9 @@ A complete, text-only capture of the public listing, vendor, and review pages of
 DrugHub, a Monero-only darknet market operating since 2023. Everything here was
 visible to any visitor without an account. Doesn't include any images.
 
-Collected 16-17 September 2026.
+Collected 16-17 September 2026. Enriched with model-derived labels
+(`typesafe/jev-1.13`) on 19 September 2026; see the `listing_enrichment`
+table and the Enrichment section below.
 
 ## What's in it
 
@@ -59,6 +63,7 @@ Collected 16-17 September 2026.
 | `vendor_listing_titles` | 35,793 | Every listing title each vendor has been reviewed on, including delisted ones. |
 | `rosters`, `roster_entries` | 91 / 145,422 | Snapshots of the sidebar vendor roster with feedback percentages, deduplicated by content. |
 | `category_counts`, `category_count_entries` | 16 / 1,888 | Snapshots of the category tree with listing counts. |
+| `listing_enrichment` | 22,621 | Model-derived labels per listing (language, substance, form, purity claim, and 12 other fields). See "Enrichment" below. |
 | `fetch_log` | 35,402 | Every HTTP fetch made, including failures, with timestamp, status, and content hash. Page bodies are not included. |
 
 Every parsed row carries a `fetch_id` that points into `fetch_log`, and a
@@ -124,3 +129,48 @@ rates at fetch time are on every row (`xmr_usd`, `xmr_eur`, ...).
   10,000-character maximum.
 - **Category is the market's own tree**, not a normalised taxonomy. About 10% of
   listings are outside the Drugs category.
+
+## Enrichment (model-derived, not ground truth)
+
+`listing_enrichment` holds one row per listing with labels produced by
+`typesafe/jev-1.13`, a classification-only model that picks from options we
+define and returns a probability distribution. Every enumeration includes an
+escape option (`other_listed_class`, `not_stated`, `unknown`, ...), and each
+choice column has a companion `<field>_p` with the probability of the chosen
+option. `answers_json` carries the full distributions. Treat these as
+annotations to be thresholded, not facts.
+
+| Field | Type | What it answers |
+|---|---|---|
+| `language` | choice | Primary language of the description (ISO code, `mixed`, `other`, `unknown`) |
+| `substance` | choice | Substance sold, from a per-subcategory vocabulary (Drugs only) |
+| `product_type` | choice | Product kind for non-drug categories |
+| `form` | choice | Physical form (pill, powder, crystal, flower, vape, blister_pack, ...) |
+| `purity_claim` | choice | Stated purity/potency bucket (`none`, `under_50`, `50_79`, `80_94`, `95_plus`) |
+| `lab_tested` | 0-1 | Claims lab testing |
+| `pharma_claim` | choice | Presented as genuine pharmaceutical, replica/pressed, or not applicable |
+| `counterfeit_risk` | 0-1 | Imitates a pharmaceutical product (branded press, fake blister, RC substitution) |
+| `quantity_stated` | choice | Whether the price maps to a stated quantity |
+| `wholesale` | 0-1 | Bulk/wholesale orientation |
+| `stealth_described` | 0-1 | Packaging/stealth described |
+| `tracking` | choice | Tracked, untracked, both, not stated |
+| `fe_required` | 0-1 | Vendor asks for Finalize Early / early escrow release |
+| `external_contact` | choice | Mentions an outside channel (see caveat) |
+| `cross_market` | 0-1 | Mentions history on another darknet market |
+| `harm_reduction` | 0-1 | Contains dosing guidance or safety warnings |
+| `fentanyl_mention` | choice | `none`, `claims_free`, `contains`, `warning` |
+| `text_style` | choice | `terse`, `handwritten`, `marketing`, `template`, `other` |
+
+Notes:
+
+- `prompt_version` is `2` for most rows and `3` for Prescription Drugs,
+  Peptides, Steroids, and Benzos - Pills, where the substance vocabulary was
+  expanded after a first pass. All other questions are identical across
+  versions.
+- `substance` vocabularies were built from title-token frequencies per
+  subcategory. Prescription Drugs uses therapeutic-class buckets for the long
+  tail; about 12% of that subcategory still lands in `other_listed_class`.
+- `external_contact` has low precision: its `other_platform` value often
+  reflects a mention of another site or market rather than an actual handle.
+- The model saw the title, category, unit, origin, and the first 2,500
+  characters of the description.

listing_enrichment.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1738d658aa2c868a223f06de09a7b323f75b14ace795a8f1a53eec3edd1a17b
+size 7968242