PharmaCliffMethodology

Every source named. Every signal weight published.

The competitor set treats data provenance like a moat. PharmaCliff treats it like a checklist. If a verdict goes in front of a Director of IP, the ingest path is owed to them.

Patents indexed
30,142

Pharma + biotech, CPC-pre-filtered against 14 therapeutic-area anchors.

Orange Book matched
329

Linked to live ANDA opportunities — real generic-launch candidates.

Jurisdictions
US · EP · JP · KR · WO

INPADOC family linkage adds IN, CN, BR reach when a family member is present.

Send window
Tue 06:00 UTC

Indexes refresh daily. Digest assembled Monday, AI-scored, subscriber feedback tunes the next rubric.

FDA Orange Bookfda.gov/cder/obPatent + exclusivity listings tied to approved NDAs and ANDA opportunity routes.refresh · daily
USPTO PAIRpatentcenter.uspto.govApplication file wrappers, term adjustments, maintenance status. Bulk weekly + on-demand re-pull.weekly + on-demand
EPO INPADOCepo.org/inpadocCross-jurisdiction patent families. The piece that makes US/EP/JP/KR alignment honest.weekly
Google Patents BigQuerypatents-public-dataHistorical backfill and CPC normalization. Used to seed the index, not to score.monthly · historical

01Curation pipeline

Four stages, fully automated, no human pre-send review:

  1. CPC pre-filter. Every new patent is matched against 14 therapeutic-area anchors (CPC class prefixes). Non-matches drop out.
  2. AI scoring. Each matched patent gets scored against the five signals (see §02). Temperature 0 — deterministic.
  3. Verdict bands. Score 0–100 maps to HIGH (≥ 75), MEDIUM (50–74), LOW (< 50).
  4. Subscriber feedback. Thumbs-down clicks on email rows or dashboard rows are reviewed and used to tune the next rubric version. No silent re-scoring of past digests.
Honest

PharmaCliff is curated by AI scoring, not by hand. The founder is not personally reviewing every HIGH-verdict row before send — claims to the contrary on any page are wrong. The honest framing: AI scores, subscriber feedback corrects, rubric improves.

02Verdict logic

Five signals, weighted into a 0–100 score:

SignalWeightWhat it captures
Orange Book ANDA route25Patent is associated with an NDA + has a visible ANDA pathway (generic submission slot).
INPADOC family convergence20How many jurisdictions in the family cliff within the same 12-month window.
Claim breadth20Composition-of-matter (broad) ranks higher than method-of-use or formulation (narrow).
Time-to-expiry20Closer expiries score higher — the digest is for decisions you can act on this quarter or next.
TA-anchor match strength15Strong CPC-anchor matches in your tracked areas outrank tangential matches.

Holder name is removed from the scoring prompt to prevent bias. Same patent + same rubric version → same score, always.

03Known limitations

  • Beta AI. Rubric weights are still being tuned. First 20–50 subscribers will see imperfect rankings before convergence.
  • No real-time USPTO grant feed. New grants land in the index within a week of publication, not within minutes.
  • No human pre-send review. AI scores every row autonomously. Subscriber feedback is the correction loop.
  • Excludes medical devices, diagnostics, veterinary pharma, agrochemicals, cannabis, and consumer-health formulations without a CPC anchor match.

04Data handling

Personal data stored: email + therapeutic-area selections + jurisdiction selections. Hosted on Neon Postgres in Frankfurt (EU). Email delivery via Resend (also EU infrastructure when possible). GDPR data export and deletion endpoints in account settings. Full details in the privacy policy.