Campus classifier changelog

Rendered from classification_runs/CHANGELOG.md at build time.

Versions of classify_locations_jev.py. Each run is recorded under classification_runs/<version>/<timestamp>/:

run.json: version, model, threshold, taxonomy sources, full choice list, label counts.
location_report_jev.csv: the per-message results (the web/ status site shows the latest one).

Both are committed.

Bump CLASSIFIER_VERSION in the script and add an entry here whenever the taxonomy, question, or model changes.

v3 (2026-09-27): origin facts in the input

Input: each message's state now starts with an Origin block from src/mail_origin.py (rules over headers/body, no model): whether the vendor sent it directly, sent it via an invoicing platform, replied in a thread, or an Alpha/Trilogy employee forwarded or wrote it; the original sender inside a forwarded block; and whether billing@ was in To or only Cc. To/Cc headers are included too. Prompted by the mail↔QuickBooks join (docs/qbo-investigation.md): of the 9 messages provably tied to Miami bills, v2 labelled 5 not_campus_specific — all forwards/threads where the employee sender was taken for the vendor.
Taxonomy, question, model and threshold unchanged from v2, so v2→v3 diffs isolate the effect of the origin block.
Result (run 20260927T114857): no improvement. 99/100 labels unchanged; the block mostly moved confidence — 6 messages fell below 0.7, 4 rose above (21 uncertain vs 19). On the 8 messages the mail↔QuickBooks join ties to a company: v2 4 right / 2 wrong / 2 uncertain, v3 3 / 2 / 3. The two confident-wrong cases are identical in both: an employee's one-line reply to a Mason Security reminder, where the only campus evidence is the invoice number — no input framing can fix that; a QuickBooks DocNumber lookup can. The one new confident answer is the known 75010 collision: "Invoice for Sports Academy 75010 LLC" is now a confident Alpha Carrolton (wrong). Keep v3 as the current run (the origin facts are still useful on the site), but the next version should add deterministic pre-passes rather than prompt changes: invoice number → QuickBooks bill → company, and vendor → company from the per-company vendor lists.

v2 (2026-09-25): QuickBooks companies as the taxonomy

Taxonomy source: the 73 QuickBooks companies Stephen accepted invites to on 2026-09-25 (records/qbo_invites_2026-09-25.tsv), replacing the alpha.school/locations scrape. Labels are now slugs of the QBO company name (e.g. alpha_school_78746_llc_austin_k_8 rather than alpha_austin), so each label maps to a company an expense can actually be posted to.
Addresses: each Alpha-branded company is enriched with its street address from alpha_locations.json, matched by ZIP, then by alias (Carrolton, Jamaica, OKC), then by campus name. Sister brands (Nova, Sports Academy, gt.school, ...) get no scraped address.
75 choices (73 companies + other_alpha_campus + not_campus_specific), up from 49.
Output CSV gains a classifier_version column.

v1 (2026-09-24): alpha.school campus list

Taxonomy source: 47 campuses scraped from https://alpha.school/locations/ (alpha_locations.json), plus other_alpha_campus and not_campus_specific.
Model jev-latest, confidence threshold 0.7.
Runs 20260924T133942 (100 messages; 18 below threshold) and 20260924T151708 (100 messages; 19 below threshold), recorded retroactively from the former alpha-billing-shared snapshots.