| --- |
| license: cc-by-4.0 |
| pretty_name: Reliability dataset |
| language: |
| - en |
| size_categories: |
| - 10K<n<100K |
| tags: |
| - fact-checking |
| - claims |
| - public-figures |
| - reliability |
| configs: |
| - config_name: subjects |
| data_files: v1.1.2/subjects.csv |
| - config_name: scores |
| data_files: v1.1.2/scores.csv |
| - config_name: claims |
| data_files: v1.1.2/claims.csv.gz |
| - config_name: changes |
| data_files: v1.1.2/changes.csv |
| --- |
| |
| # Reliability dataset |
|
|
| Reliability scores for public figures and organisations, with every claim, verdict and evidence link behind them, |
| built by the Reliability API (https://epistacklabs.com). A subject's public statements are sampled, the checkable |
| claims are checked against the public record, and each claim carries a verdict linked to its quote and its |
| evidence. The score is the share of resolved claims that held up, for each kind of claim (factual claims, |
| predictions, promises, company guidance), with an 80% interval and a confidence level. Predictions are scored on |
| their own. The overall row mixes kinds of claim in different proportions for different subjects; compare subjects |
| within one kind of claim, and only at medium or high confidence. |
|
|
| Verdicts are made automatically and are not individually reviewed by a person. Two model judges from different |
| model families judge each claim; `judges_disagree` marks where they differed. Two judges agreeing is consistency, |
| not accuracy. The method, its measured results and its caveats: https://epistacklabs.com/methodology. |
|
|
| ## Releases |
|
|
| Each release is a folder (`v1/`, `v1.1/`, `v1.1.1/`, `v1.1.2/`, ...) with: |
|
|
| - `subjects.csv`: one row per subject, with the latest report and how to reply (`respond_url`). |
| - `scores.csv`: one row per subject and claim type (and overall): score, interval, resolved count, confidence. |
| - `claims.csv.gz`: one row per claim: quote, source, verdict, rationale, evidence links, whether the judges |
| disagreed, how many evidence items are fact-checks. |
| - `claims.parquet` (from v1.1): the claims table as Parquet, the same columns and cells. |
| - `changes.csv`: verdict changes (old and new verdict, when, why, evidence added). |
| - `claims.jsonl.gz`: every record in full, one JSON object per line. |
| - `infographics.html`: the release in charts, one self-contained page; `infographics.json` holds its figures. |
| - `datapackage.json` (the columns) and `README.md` (what the release is, its method version, measured results, |
| caveats, scope and licence). |
|
|
| `CHANGELOG.md` lists the releases and what changed between them. The latest is `v1.1.2`. |
|
|
| ## Look a subject up |
|
|
| No key needed: |
|
|
| curl "https://epistacklabs.com/v1/lookup?name=Jensen%20Huang" |
| |
| Or read every claim in a browser at https://epistacklabs.com/explore. |
|
|
| ## Scope and right of reply |
|
|
| Public figures and organisations only, from their public professional output; nothing about private life; no |
| inferred motives; neutral wording throughout. Politicians and officials are covered like anyone else, and |
| party-political claims are tagged `political` and scored like any other claim (from release v1.1; v1 left some |
| out of scores, as its README says). A subject, or someone acting for them, can reply to any claim with a correction |
| or counter-evidence, with the form at the foot of the claim's own page on https://epistacklabs.com (the |
| `respond_url` on their row explains how replies work). |
|
|
| ## Licence |
|
|
| CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quotes are short, attributed and linked to their |
| source; no page text is included. Cite as: Reliability dataset v1.1.2, epistacklabs.com, 2026. |
|
|
