# Nameworthy research counts, version 1

Published November 7, 2026. This package supports the corrected engine-overlap and cross-date durability reports. It contains anonymous count primitives, not raw customer records.

Files: `research-counts-v1.json`, `research-summary-v1.json`, `recompute.py`.

To reproduce: place the counts at `public/data/research-counts-v1.json` and run `python3 recompute.py` from that directory's project root. Python standard library only. The script writes the summary beside the counts. Website source also contains `scripts/research/export.py`, which needs the private archive and does not alter it.

## Units

- `overlap`: one panel/question group and engine pair; intersection and union of pooled name sets. Require both sets nonempty and union >=3. A published pair mean needs >=12 groups. Overall 12.4% is the unweighted mean of 24 pair means; weighting by groups gives 11.9%.
- `repeat`: two nonempty runs of one question and engine, with union >=3. Published engine means require >=20 pairs. Runs participate in multiple comparisons; pairs are dependent.
- `transitions`: successive measured dates within a panel/question/engine. Distinct first names and all names are pooled separately on each date. `first_held/first_total` and `presence_held/presence_total` sum within engine. Engine tables require >=8 first-position opportunities. Empty competitor lists are omitted by the original source algorithm, a potential upward bias in persistence. Gaps are observed calendar-day differences, not survival times.
- `lane`: sequential opaque grouping identifier within each analysis, not a client ID. No mapping to names is provided. IDs in transitions and overlap are unrelated.
- `runs` and `lanes`: counts from the readable indexed archive at rebuild time. Not all qualify for each table.
- `source_snapshot_sha256`: digest of input data and source-code bytes used by the export. It identifies the source snapshot, not a claim of independent validation.

API and app labels remain distinct. Extraction, normalization, missing outcomes, unequal repetitions, and convenience sampling limit interpretation. No inference of causality, unmanaged status, survival time, revenue lift, or representative market prevalence is supported.

No names, locations, full prompts, source URLs, exact observation dates, answer text, credentials, or client reports are included. Count-only public reproducibility cannot independently validate private name extraction. Cite Nameworthy, *Research counts v1*, November 7, 2026, https://nameworthy.ai/data/README.md. No DOI is claimed.

## Source-class counts

`offsite-counts-v1.json` contains only industry totals, citing-run denominators and integer source-class counts reconstructed from the retained August 30 artifact. `offsite-summary-v1.json` is regenerated by the same Python script. Obtain both counts files before running it. These reproduce all seven rows of the public source table: 3,194 runs, 1,647 with citations. A run can cite multiple classes, so percentages need not sum to 100. Public counts reproduce the arithmetic; independent domain-classification validation would require the private source evidence. Original classifier limitations and redirect-title handling remain limitations, not newly validated facts.