Show the work.
Every number needs a source, a denominator and an account of what the method can miss.
Define the corpus
The October snapshot contains 35,897 NBER working papers issued from 1973 to 2026. NBER metadata identifies papers, authors, programs and recorded publications. OpenAlex supplies working-paper citation counts, references and affiliations.
Authorship records cover 35,892 papers. There are 5 indexed papers outside that coverage. Missing records are not zero-valued observations.
Keep the samples separate
The text-reading sample contains 333 papers from the most-cited group across years, 681 from the most-cited group within issue years, and 810 from random draws within years. Highly cited samples are selected deliberately and do not estimate the whole series.
Text was extracted successfully from 1,660 papers. 709 yielded checkable table entries. The existing extraction generated 137,326 entries. Download failures, unusable text and tables the parser cannot interpret are separate coverage limits.
Describe authorship and citations
Authors are keyed by NBER IDs, not matched solely on names. An author credit is one author-paper association. Citation rankings in the source pipeline are calculated within issue year. The graph includes only references from one NBER working paper to another.
Version splitting, missing references, incomplete affiliations and tied citation counts can affect rankings. A paper's working-paper count may differ from the count for its published version. The saved rankings have not been recomputed for this release.
Check arithmetic; adjudicate flags
The deterministic pipeline compares printed coefficients, standard errors and significance stars, with checks for rounding and table conventions. One-sided tests, alternative inference, mislabeled quantities or extraction mistakes can explain apparent inconsistencies.
An automated flag is a request to inspect the original page and statistical method. It is not a finding of error or misconduct. No individual-level misconduct conclusion is made on this site.
Record model limitations
The retirement study has 184 model-read papers and 134 retained empirical records. Its report includes quote verification and a second-coder sample. Agreement between models is not a substitute for human ground truth.
A planted-error calibration set, independent manual adjudication and a versioned correction register remain future work. This release does not claim Casey Handmer's benchmark performance or costs for our own pipeline.
Make the snapshot inspectable
The aggregate JSON records the SHA-256 of the checked-in public report used to build this site. Team data and saved author rankings are available as CSV.
Source systems: NBER metadata and OpenAlex. Read the original report and its detailed methods.
Release record: November 3, 2026 — reorganized existing October results, added descriptive charts and data downloads. No new retrieval, model reading or audit run.
Corrections
To report a problem, send the page URL, the number or claim, the source paper and page number where applicable, and the proposed correction to t@thomas.tc. Proposed corrections require source verification; this site does not automatically turn submitted claims into findings.