How CM-PATRA works, and what is measured

Software agents do the repetitive work; a person reviews exceptions. No language model sits in the publishing path, so the same input always gives the same output and every label can be traced.

Architecture

Scout. Ingests Google Patents Public Data (BigQuery, CC BY 4.0) and the SATYABHAMA ongoing-projects list. Next: IPO Patent Journal.
Classifier. CPC rules first, then title/abstract keyword rules, then reviewed overrides. Records the method on every label.
Resolver. Merges applicant-name variants to one canonical organisation and tags country and type.
Auditor. Deterministic gate: identity present, id formats, date order. A failure blocks the build; odd source dates are flagged, not hidden.
Analyst. Counts by mineral, stage, year and applicant; builds the technology-gap matrix against funded projects.
Alerts. Fixed rules (empty minerals, funded-but-no-patent, recent Indian activity). Each alert states its rule.

Measured on this build (2026-10-13)

Coverage and provenance, computed from the data. Not classification accuracy: that needs a hand-labelled test set and has not been measured yet.

Patent families kept / ingested	346 / 438	Classifier output vs scout input
Kept records with an English title	55%	190 of 346
Kept records with a named applicant	100%	346 of 346
Records with a specific mineral (not “general”)	75%	Mineral resolved by CPC or text
Labels decided by CPC rules alone	86%	Trace method = cpc-rules
Labels needing the text rules	14%	Trace method = cpc-rules+text
Labels set by a reviewer override	0	Each carries its reason
Low-confidence records (no English text)	30	Flagged on the record page
Audit gate	Passed	568 records checked
Funded projects naming an India-list mineral	62 / 130	Title rules; Ministry of Mines S&T list

Pilot: Indian Patent Office Journal 41/2026

The latest weekly journal (9 Oct 2026) lists 1,472 newly published applications. A broad filter on classification and title words flagged 49; reading them, 6 are mineral-processing technology and 0 have an Indian applicant. This shows the recent-filings feed works; it is one issue, kept out of the main counts until more issues are ingested.

METHODS OF RECYCLING ALKALI METAL ION BATTERIES · 202617071376 · NATIONAL UNIVERSITY OF SINGAPORE · H01M10/54, C25B1/16, C22B3/44
METHOD FOR RECYCLING NEGATIVE ELECTRODE MATERIAL GENERATED AS PROCESS WASTE · 202617085954 · ABR CO., LTD · H01M10/54, B03B5/02
PROCESS FOR LEACHING MAGNESIUM FROM A MATERIAL · 202617088708 · PRAYON · C01F5/22, C22B3/08
VANADIUM EXTRACTION · 202617087593 · IMPERIAL COLLEGE INNOVATIONS LTD · B01D11/02, B09B3/70
METHOD FOR TREATING SULFATE IN INDUSTRIAL PROCESSES · 202647116323 · UMICORE BATTERY MATERIALS FINLAND OY · C01D5/08, H01M10/54
SHELF DETECTION AND CONTROL FOR VACUUM ARC REMELTING · 202647116473 · TITANIUM METALS CORPORATION · F27D11/08, C22B9/04

Limitations and risks

Google Patents has few Indian-office filings for 2016–2026; recent Indian filings are under-counted until more IPO Journal issues are ingested (one is piloted).
R&D coverage is one scheme with titles and status only; no institutions or outcomes.
Keyword and CPC rules can mislabel; overrides exist but the accuracy figure is not yet measured.
Counts are lower bounds and are not legal advice or a freedom-to-operate opinion.

Roadmap to 30 October

Ingest more IPO Patent Journal issues (one issue is piloted above).
Research-paper outputs (OpenAlex) alongside patents and projects.
Hand-labelled test set and a measured accuracy figure.
MCP tools so AI assistants can query the tracker.
Watchlist alerts by mineral or applicant.

Cost to run: the dataset is a static file served with the site, so there is no per-query cost. Refreshing it from BigQuery costs per scan (the seed extract scanned about 258 GB per wide query).