Generic Price Cliff · method and verification

How was the Generic Price Cliff checked, and what did the independent check find?

A separate reader recomputed the figures from the raw files and matched them with no mismatches. The method was written and hashed on 12 November 2026 before any figure was published, and 1 dated amendment records what the checking found.

What was fixed before any figure was published?

The method was frozen on 12 November 2026. The file's SHA-256 hash begins e59f5eeba242, and the original, before the amendment, began b7749c62cc15. The tool refuses to build if the method file no longer matches its hash.

What does the independent check do?

  1. A separate reader (Python, no shared code) recomputes the eligible set and the curve straight from the stored NADAC snapshot rows and the raw Orange Book files; the published curve and the eligible product count must equal its result.
  2. A mutation control: moving a product's year zero by one year must change its index path (the control must be able to fail).
  3. Every dropped product is counted with its reason (no unique match, first generic outside 2013 to 2022, fewer than 3 surveys, reference price with fewer than 2 NDCs), and the counts reconcile with the products considered.

What did it find?

MeasureResult
Eligible products compared200
Curve years compared14
Mismatches0

A separate Python reader rebuilt the eligible product set and the whole curve from the stored NADAC snapshot rows and the raw Orange Book files. Its first run found a matching defect in the builder, recorded as protocol amendment 1; after the fix the two agree.

The checking led to one amendment to the method, dated and recorded:

2026-11-12 (amendment 1, before any publication of the cliff statistics): the independent control in section 4.1 found that the first version of the matching let "AMOXICILLIN 250 MG CAPSULE" match the three-drug combination group "amoxicillin; clarithromycin; lansoprazole", because the single-ingredient amoxicillin capsule group has no dated brand approval and so the combination was the only candidate. The matching rule in section 1 is clarified: an Orange Book group matches only if its ingredient text is a single ingredient (no semicolon; a salt or ester suffix after the NADAC drug name is allowed, a second active ingredient is not), and its form text contains exactly one of TABLET or CAPSULE (kits and mixed-form groups are excluded). A NADAC product whose only candidate is a combination or mixed-form group is dropped as "no unique match". With the clarified rule the independent reader and the builder agree. The original protocol hash is kept in PROTOCOL-GENERIC-CLIFF-2026-11-12.original.sha256.

What is still not covered?

Matching is by name and form kind only, not by strength; a product on a strength the first generic did not cover can enter later. The Orange Book dates the first approval, not the first sale. NADAC year-end snapshots are one week in 52. Survivorship: products priced in later years are the ones still on the market. The selection is tablets and capsules with a unique match, a minority of all products.

How can you check it yourself?

The source data is public: see the FDA generic drug facts. The dataset, the method and the code that rebuilds it are published at doi:10.5281/zenodo.24182324. Open the Generic Price Cliff to see the result.

Questions about how the check works

What does "frozen before the first publication" mean?

The method was written, hashed and saved before any figure from the data was published. A change after that is an amendment with a date, and the original hash is kept, so anyone can see what was decided and when.

Why does an independent reader matter?

A tool can agree with itself and still be wrong. A separate reader, written without the builder's code, recomputes the figures from the raw files; where the two disagree, one of them has a defect.

Does a clean check mean the tool is always right?

No. It means the numbers match an independent recomputation under the stated method. The limits section lists what the method does not cover.

Further reading: Zenodo, where the datasets are deposited; the FAIR data principles; the National Academies on reproducibility; the author's ORCID record; the FDA generic drug facts.

Compiled by Vincent Wesley Couey (ORCID 0009-0005-6869-308X).