Evidence graph

Every claim, traced to its data

Each claim below resolves through the same chain: claim → evidence → dataset → methodology → source → data date. A claim that cannot complete that chain does not get published as a fact — it is listed with the reason instead.

Quick answer

How does Harpd support the claims it publishes?

Harpd publishes 5 claims that resolve to 4 registered datasets, each with its own methodology URL and the date the underlying data was current. Claims are typed (measured, observed, calculated or a labelled modeled estimate), so an estimate can never be read as a measurement, and Rank Points are recorded as promotional placement rather than an editorial quality score.

Published claims
5
Datasets behind them
4
Claims failing the gate
0
Newest data date
2026-10-20

Data source: Harpd evidence graph (first-party)Last updated: Machine-readable graph (JSON)Methodology overview

Why it matters

An evidence graph turns a marketing claim into a verifiable chain: claim → evidence → dataset → methodology → source → data date. Showing the chain (and the claims that fail it) lets a reader or answer engine trust a number because they can trace it, not because Harpd asserts it.

Data source: Harpd evidence graph (first-party)Last updated: October 20, 2026

Published claims

Each block is independently quotable: statement, claim type, dataset, method and data date. Copy the citation line as-is when referencing a figure.

FactObserved
Every model price published on harpd.com is a public list price read from the provider's official pricing page, and each of the 16 records carries its own source URL and verification date (most recent verification 2026-08-18).
Data date

Cite asHarpd. "Harpd Model Pricing Dataset." harpd.com/llm-pricing/ Accessed: 2026-08-18. License: CC BY 4.0.

FactModeled estimate
The Structured JSON extraction from messy natural-language input (en + ja + zh) benchmark dataset covering 8 models is a modeled estimate, not a measured result: it shows how the live benchmark will be scored and must not be cited as measured output (dataset generated 2026-08-18).
Data date

Cite asHarpd. "Harpd Benchmark Dataset." harpd.com/benchmarks/ Accessed: 2026-08-18. License: CC BY 4.0.

FactObserved
Rank Points published for a product equal the Rank Credit spend that product completed (one completed Credit creates one permanent Rank Point). Rank Points are promotional placement, not an editorial quality score.
Data date

Cite asHarpd. "Harpd Product Ranking Dataset." harpd.com/data/ Accessed: 2026-10-20. License: CC BY 4.0.

FactObserved
The Harpd ranking dataset contained 1169 approved products across 27 categories when it was last regenerated (2026-10-20); every record carries its own updatedAt timestamp.
Data date

Cite asHarpd. "Harpd Product Ranking Dataset." harpd.com/data/ Accessed: 2026-10-20. License: CC BY 4.0.

FactObserved
Harpd publishes its complete ranking rules, board definitions, entry conditions and disclosure statement as a machine-readable methodology dataset. The rules are versioned: every change is recorded with a date so a citation can reference the exact rule set in effect at citation time.
Data date

Cite asHarpd. "Harpd Methodology Dataset." harpd.com/methodology/ Accessed: 2026-10-20. License: CC BY 4.0.

Not claimed

Statements Harpd deliberately does not make, and what would have to exist first.

That any model is "safe to switch to" for a production workload.

A switch decision requires a real-task quality gate on the specific workload. No dataset covers that yet.

Published method

We publish the method, not a result.

That a higher-ranked product on Harpd Rank is a better product.

Rank Points are promotional placement bought with Credits. Harpd holds no quality measurement that would support an ordering by merit.

Published method

Monthly visit counts or user review scores for listed products.

Harpd has no first-party measurement of third-party traffic or reviews, and does not republish vendor-supplied figures as facts.

The non-sponsored signals Harpd can verify are published in /data/rank-open-data.json instead.

How this register is produced

Evidence

Chain enforced
claim → evidence → dataset → methodology → source → data date
Claim types
MEASURED, OBSERVED, CALCULATED, MODELED (labelled), EDITORIAL, USER_SUBMITTED
Fact surface
Editorial and user-submitted claims are never rendered as facts
Machine-readable twin
/data/evidence.json

Method

The register is built in code from live data sources — the pricing registry's per-record verification dates, the benchmark run's own timestamp, and the newest updatedAt across approved Rank listings. Each claim is then resolved against the canonical dataset registry; anything missing a dataset, a methodology URL or a real data timestamp fails the gate and cannot be rendered as a fact. Full methodology →

Data

Limitations

  • Harpd Rank Points are promotional placement bought with Credits. They are not a quality measurement and no claim here orders products by merit.
  • A modeled estimate stays labelled as an estimate until a measured run replaces it; historical modeled records are never overwritten.
  • Timestamps are data dates, not build dates. A page rebuilt today does not make its underlying data newer.
  • Coverage is limited to products that chose to list on Harpd Rank; the dataset does not represent the entire product market.

Underlying data date:

Cite this evidence graph

Copy the citation line that matches your format. Every citation includes the dataset, methodology, data date and licence.

Check it yourself

Machine-readable evidence graph

The same claims, datasets and gate results as JSON.

How to cite Harpd

Ready-to-copy plain-text and BibTeX citations for every dataset.

Monthly data drops

Immutable closed-month snapshots — the citable form of the board.

Open data hub

Every public dataset endpoint in JSON and CSV.

Research hub

Reports built only from closed monthly snapshots.

Methodology overview

How each number is produced.