CryptoTaxEdge/crypto-tax-classification-2026h1

Crypto tax transaction classification, 2026 H1 labelled benchmark

388 public blockchain transactions on Ethereum, Polygon, and Arbitrum, each labelled with a transaction category and the default US tax treatment that category maps to under an open, versioned classification schema. Published under CC BY 4.0 together with the schema itself.

Canonical landing page: https://cryptotaxedge.com/research/benchmarks/2026-h1/ Schema: https://cryptotaxedge.com/standard/ DOI: https://doi.org/10.5281/zenodo.21911546

Not tax advice

A category and its default treatment are a starting point for professional review, not a tax position. Classification of any specific transaction depends on facts and elections, and several categories carry documented alternative positions a firm can elect instead of the default. Verify with a qualified tax professional before filing.

Fields

field	description
hash	Transaction hash, a public on-chain identifier. Fetch the transaction, receipt, and logs from any public node to re-derive the evidence.
chain	ethereum, polygon, or arbitrum.
stratum	The sampling stratum the row was drawn from, recorded verbatim from the frozen corpus.
category	Ground-truth category from a closed 17-category set fixed before any model was called. 13 occur here.
treatment	disposal, income, non_taxable, or expense. Derived from category, so it carries no information beyond it.

388 rows: 262 Ethereum, 77 Polygon, 49 Arbitrum. By treatment: 213 disposal, 150 non_taxable, 22 expense, 3 income.

How the labels were made

Under a protocol pre-registered on 2026-07-22, before execution. Contested rows were labelled by two independent blind model lenses reading the raw receipt and logs, a net-token-flow lens and a protocol-event lens, with every system's output relabelled first so neither lens knew which system had claimed what. The label is their unanimous verdict; rows without unanimity were dropped rather than tie-broken. The remaining rows carry a label two independent classification systems agreed on under an equivalence map fixed in the scorer before adjudication.

Per-row provenance is not distributed: one of the two agreeing systems is a third party whose outputs are not redistributed.

Limitations

• Model-adjudicated, not CPA-labelled. No human tax professional adjudicated this key.
• Answerable by construction. 187 hard rows where the reference implementation abstained have no truth key and are absent, along with 12 duplicate hashes, 14 rows outside the closed taxonomy, and 176 unadjudicated disagreement rows. A score here is a score on the answerable subset.
• Superseded as an accuracy instrument, which is why it is open. Six classification rules added on 2026-08-13 were discovered by mining this key, so the reference implementation is fitted to it and no current head-to-head number may be computed from it. It stays useful for taxonomy work, error analysis, and evaluating systems that had no part in building it. The replacement held-out corpus is not published and will not be until its own study publishes.
• A label consensus is not an oracle.

Citation

@misc{cryptotaxedge2026standard,
  author       = {CryptoTaxEdge},
  title        = {CryptoTaxEdge Classification Standard v1.1.0 and the 2026 H1 labelled benchmark corpus},
  year         = {2026},
  doi          = {10.5281/zenodo.21911546},
  howpublished = {Zenodo, https://doi.org/10.5281/zenodo.21911546},
  url          = {https://cryptotaxedge.com/research/benchmarks/2026-h1/}
}

Trademarks

The text and data are CC BY 4.0; the marks are not licensed. A modified or independently maintained document may not be presented as the Standard or as a version of it, and must indicate that it is modified.