# DN Benchmark Denominator Lab v0.1

Published 4 November 2026. Ten invented tasks, three recorded attempts each. No model calls, provider tests, paid evaluation, external fetches or factual validation. All result and assistance labels are invented declarations. The sample is not drawn from a population.

First recorded attempt: 6/10 tasks qualify. Any recorded attempt: 9/10. Any attempt in the reported subset: 9/9 after task-10 is omitted. Any attempt with assistance declared none: 7/10. Across all attempts 14/30 pass; one of the other results remains unresolved. These rates answer different questions. None is a statistical pass@k estimator, a predicted deployment success rate, or independent evidence of autonomy.

There is one passing attempt labeled human-assisted and one with unknown assistance. Requiring none excludes both from the numerator; unknown remains unknown. Task order or array reordering does not choose a winner: explicit attempt numbers define first. Each task can have zero through three contiguous ordered attempts; a task with no attempts remains in the task denominator and does not qualify. An empty selected subset has rate null.

The fictional reported subset excludes the only task without a pass to expose denominator sensitivity. This demonstrates post-outcome selection; a predeclared legitimate subset can be appropriate if its scope and exclusions are disclosed. Infrastructure problems are not automatically evidence of model inability; retain unresolved rows and describe separate capability and operational measures.

Each recorded attempt has a modeled cost of 100 USD cents. The full ledger cost is always 3000 cents ($30) regardless of the selected view. This is the cost of all 30 recorded attempts, including failed and unresolved attempts and any outside the reporting subset. It is not a counterfactual first-attempt-only bill. Human work, infrastructure, grading and downstream selection costs are omitted.

Download ledger.json, results.json, summarize.mjs and reproduce.mjs to one folder; run `node reproduce.mjs`. The source hash in results.json identifies bytes, not publisher authenticity or proof of a test date. The JSON claim card is a blank planning template; null fields mean no evidence supplied.

Falsification: supply a supported ledger and reporting policy that the published summarizer miscounts. Changing a policy can legitimately change the answer. No result is a vendor ranking, fraud finding, certification or approval to act.

CC BY 4.0 for DN-authored research files. Preserve attribution and limitations. Third-party standards and benchmark sources retain their terms. Corrections: https://decentralised.news/contact . Version 0.1 is the initial release; material changes require a new version.