Corp_Taxes — Expert-Authored US Corporate Tax Reasoning (Pilot Sample)

Seven evaluation tasks, authored and peer-reviewed by credentialed US corporate tax professionals.

Produced by Human Edge as a public demonstration of our expert-data methodology.

• Curated by: Human Edge
• Language: English
• License: CC BY 4.0
• Repository: HumanEdgeAI/Corp_Taxes

---

Why this dataset exists

Seven corporate tax scenarios, each written from scratch by a practicing US tax professional, paired with a ground-truth answer, a stepwise reasoning trace, and the statutory sources the expert relied on. Every task then passed through a review pipeline in which independent domain experts ruled on the answer's correctness, scored difficulty and real-world representativeness, and recorded written reasoning for both scores.

Frontier models handle textbook tax questions well. They are far less reliable on the multi-step, statute-dependent reasoning that practitioners do daily — where the answer turns on which provision governs, how basis and boot are allocated, and what the current-year rules actually say. Evaluating that requires people who do the work, not annotators following a rubric.

Why the review data is the point

Most public evaluation datasets ship a question and a gold answer. Whether qualified experts actually agreed on that answer, and on how hard the question is, is discarded.

This dataset keeps it. Every task carries its two final peer reviews in full — 14 reviews across 7 tasks — each with a ground-truth verdict, a difficulty score, a representativeness score, and the reviewer's written justification for each score.

These are the reviews that closed the task out, so all 14 affirm the ground truth. That unanimity is the output of the process rather than evidence that review was frictionless. Across the same seven tasks, four drew a ground-truth challenge at some point during review, four required a second full round after revision, and two required adjudication by a third expert. Those are overlapping but distinct groups of four — a challenge did not always trigger a second round, and a second round was not always preceded by a challenge. The superseded rounds are held internally; what ships here is the settled position and the reasoning behind it.

Where the two final reviewers still disagreed, both positions stand as recorded. They split on difficulty for 4 of 7 tasks and on representativeness for all 7.

Contents

Tasks	7
Columns	21
Domain	US corporate tax
Final peer reviews	14 (2 per task)
Tasks that required a second review round	4 of 7
Tasks that required adjudication	2 of 7
Reasoning steps	34 across all tasks
Prompt length	643–2,763 characters
Language	English

Question types — six categories, each represented:

Type	Tasks
Statutory Interpretation	2
Comparative Treatment	1
Rule Application	1
Multi-Issue Analysis	1
Numeric Reasoning	1
Recent Developments	1

Difficulty — all 14 final-round reviewer scores, on a 5-point scale calibrated to a senior CPA/JD baseline:

Rating	Reviews
Trivial	0
Easy	0
Moderate	8
Hard	4
Very Hard	2

Representativeness — how often a practitioner encounters this scenario, across the same 14 reviews:

Rating	Reviews
Rarely Encountered	1
Uncommon	2
Somewhat Typical	6
Typical	5
Core Activity	0

Each task contributes exactly two scores, so these distributions describe reviewer judgments rather than tasks — 14 judgments across 7 tasks.

Who wrote and reviewed these tasks

The pilot cohort were US corporate tax specialists averaging roughly 18 years of practice — CPAs, Enrolled Agents, and tax attorneys, drawn from Big 4, national, and regional firm backgrounds, in Tax Partner, Director, and Senior Manager roles. Credentials in advanced taxation, law, and accounting.

Admission was gated: CV screening, a domain-knowledge assessment calibrated to professional-exam difficulty, and a three-module training program on model behavior, prompt construction, and peer-review standards. Experts who did not clear the assessment threshold were not assigned tasks.

These seven tasks are a sample of that cohort's output; the figures above describe the cohort, not the seven authors individually.

Quality assurance

Six stages. The first two and the last apply to every task; adjudication, revision, and second-round review are triggered only when review surfaces a problem.

1 · Automated quality gate. Every task. Before any human review, an LLM evaluator independently solves the case and compares its result to the submitted ground truth, distinguishing material errors — a wrong rate, a wrong statutory conclusion — from formatting and rounding differences. It separately audits whether the reasoning trace supports the stated answer, since reasoning that contradicts a correct answer is still a defect. Every submission is also screened for AI-generated text and plagiarism. All seven tasks passed; the result is recorded in AI Check.

2 · Dual independent peer review. Every task. Two domain experts review the task blind, without visibility into each other's assessment. Each rules on ground-truth correctness, evaluates whether the prompt is unambiguous and answerable, and scores difficulty and representativeness with written reasoning.

3 · Adjudication. 2 of 7 tasks. A third expert is brought in when reviewers disagree on the ground-truth verdict, or when their difficulty scores diverge sharply. The adjudicator issues the binding ruling. Flagged per task in Adjudication in Final Round. Because the triggering disagreement usually sits in a round that is not published, the flag cannot be reconstructed from the two reviews in this file.

4 · Revision. Where review required it. The original author addresses the feedback — clarifying ambiguous constraints, correcting statutory or arithmetic errors, resequencing or completing reasoning steps — and resubmits.

5 · Second-round review. 4 of 7 tasks. Independent experts re-evaluate the revised task. Renewed disagreement dismisses the task rather than forcing consensus. Recorded per task in Peer Review Rounds.

6 · Final approval. Every task. An additional correctness check, on the reasoning that these carry the most benchmark signal and the least tolerance for a wrong answer. A project lead approves the batch for release.

Field reference

21 columns in three groups. No column is empty and no row has a missing value.

Task content

Column	Type	Description
Task ID	string	Stable task identifier. Values are non-contiguous — they are the original corpus identifiers.
Question Type	string	One of the six categories above
Prompt	string	The tax scenario as written by the expert, including the practitioner persona and constraints
Reasoning Steps	string	The expert's stepwise derivation, serialized as a list of {step_id, reasoning_step} objects
Ground Truth	string	The authoritative answer
AI Check	string	Result of automated AI-authorship and plagiarism screening. Pass for all seven tasks.
Sources and Material	string	IRC sections, regulations, rulings, and publications relied upon

Review process

Column	Type	Description
Peer Review Rounds	int32	1 or 2 — whether the task cleared review on the first pass or required revision and a second round
Adjudication in Final Round	bool	Whether a third expert was brought in to settle a disagreement

Final peer reviews

Two blocks, Review 1 - and Review 2 -, holding the two reviews that closed the task out. Both are fully populated for every task.

Field	Type	Description
Ground Truth Examination	string	Correct / Incorrect verdict. Correct throughout — see Limitations.
Prompt Feedback	string	Assessment of scenario clarity and answerability, often including the reviewer's own working
Difficulty	string	Reviewer's difficulty rating
Difficulty Reasoning	string	Written justification for the difficulty rating
Representativeness	string	Reviewer's frequency rating
Representativeness Reasoning	string	Written justification for the frequency rating

The two blocks are independent reviewers rather than two rounds. On the two tasks where Adjudication in Final Round is true, one of the blocks is the adjudicator's binding ruling rather than a peer assessment — the blocks are not interchangeable on those rows.

Personal and sensitive information

The dataset contains no personal data. Scenarios are constructed for evaluation and describe hypothetical taxpayers, not real ones; cited authorities are public statutes, regulations, and rulings. Author and reviewer identities are not published.

Intended uses

Built for: evaluating frontier-model reasoning on statute-dependent corporate tax problems · inspecting how expert-authored evaluation data is constructed, reviewed, and adjudicated · studying how experienced practitioners calibrate difficulty and representativeness, including where they disagree · assessing whether this methodology fits your evaluation needs.

Not built for: statistically significant model benchmarking at this sample size · training a tax-advice system · tax advice, filing positions, or professional guidance of any kind.

Nothing here is tax advice. These scenarios are constructed for model evaluation. Do not rely on them for any real filing position.

Limitations

Stated plainly, because a dataset card that hides its own caveats is not worth trusting.

• Seven tasks. Five of the six question types have exactly one task; only Statutory Interpretation has two. Treat the taxonomy as a description of coverage, not a balanced distribution, and do not draw statistical conclusions at this sample size.

• Only the final round is published. Each task carries the two reviews that closed it out. First-round reviews, superseded assessments, and adjudicator reasoning from earlier rounds are held internally. The consequence is that all 14 published verdicts read Correct: the dissent that drove revision on four of the seven tasks is not in this file. Peer Review Rounds and Adjudication in Final Round record that friction occurred without reproducing it.

• Three reviewer fields from the source export are absent. The written justification for the ground-truth verdict is N/A in all 14 final-round reviews, so it was dropped. The step-level critique of the reasoning trace, and a yes/no flag summarizing it, survive in only 2 of the 14 — both on Review 1, for tasks #14 and #28 — and were dropped rather than shipped as columns that would be null for six of seven rows. The reasoning trace itself is present in Reasoning Steps; what is missing is the reviewer's audit of it.

• AI Check carries no variance. It is Pass for all seven tasks, so it documents that screening happened rather than differentiating between rows.

• There is no single task-level difficulty label. Author self-assessments are excluded, so difficulty and representativeness exist only as individual reviewer judgments. Deriving a task-level rating — by majority, mean, or adjudicated verdict — is left to the user, deliberately, because the aggregation choice is itself a modeling decision.

• Reviewer ratings are expert judgment, not ground truth. They are calibrated to a senior practitioner baseline, and the two reviewers on a task frequently diverged — on difficulty for 4 of 7 tasks and on representativeness for all 7. That disagreement is preserved rather than averaged away.

• Reviewer commentary is substantively unedited. Written justifications appear as submitted, including abbreviations, shorthand computations, and informal phrasing; no wording was rewritten or tidied, because the raw record is more useful than a polished one. The only processing applied was trimming leading and trailing whitespace and normalizing empty and N/A cells.

• Reasoning Steps is a serialized string, not a native nested column. Parse it before use; it will not render as expandable structure in the dataset viewer.

License

Released under the Creative Commons Attribution 4.0 International license (CC BY 4.0).

You are free to share and adapt this dataset for any purpose, including commercially, provided you give appropriate credit to Human Edge, link to the license, and indicate whether changes were made.

The license covers Human Edge's contribution — the scenarios, ground-truth answers, reasoning traces, and review ratings. It does not grant rights in the third-party statutes, regulations, and rulings cited within the tasks; those remain governed by their own terms. It does not extend to the remainder of the corpus.

Citation

If you use this dataset, please cite it:

BibTeX:

@misc{humanedgeai2026corptaxes,
  title     = {Corp\_Taxes: Expert-Authored US Corporate Tax Reasoning (Pilot Sample)},
  author    = {{Human Edge}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/HumanEdgeAI/Corp_Taxes}
}

APA:

Human Edge. (2026). Corp_Taxes: Expert-Authored US Corporate Tax Reasoning (Pilot Sample) [Data set]. Hugging Face. https://huggingface.co/datasets/HumanEdgeAI/Corp_Taxes

About Human Edge

Human Edge builds expert human data for AI development — SME-based evaluation, benchmarking, and reinforcement learning from expert feedback in domains where correctness requires professional judgment: finance, legal, healthcare, and tax.

This dataset is a sample of the pilot phase of a larger program. The production methodology scales the cohort, the review pipeline, and the volume well past what is shown here.

To discuss an evaluation or benchmarking engagement: humanedgetech.ai