TaxBench-AU

An authority-grounded diagnostic benchmark for Australian tax worked examples.

TaxBench-AU evaluates whether a tax-preparation assistant can:

1. cite the correct ATO authority (T1 — Source Retrieval and Citation);
2. recognise source status — final, draft, withdrawn, guide, or ATO Interpretative Decision (T3 — Source-Status Recognition);
3. align its substantive tax position with the cited ATO source (T2 — cross-cutting Outcome Alignment, applied to every T1, T3, and T4 item);
4. ask for governing missing facts when the fact pattern is incomplete (T4 — Sufficiency-Aware Questioning); and
5. maintain a professional boundary under adversarial pressure (T5 — Professional-Boundary Safety).

Every substantive tax label is traceable to public Australian Taxation Office (ATO) materials — rulings, determinations, interpretative decisions, guides, and worked-example solutions on the ATO Legal Database. Behaviour-only items (the T5 professional-boundary set, plus a small number of T1 "negative" / T3 status-only / T4 conflict-detection items) carry provenance: null by construction and are scored against their applicable rubric (boundary safety or deterministic source-recall) rather than against an ATO reasoning quote. The benchmark is deterministic-first: programmatic checks score run completion, schema validity, and source recall; narrow cross-provider LLM judges score outcome alignment and professional-boundary safety.

Paper: PAPER.pdf (also bundled in this dataset under paper/PAPER.pdf) GitHub repo (evaluation harness, scorers, reference adapters): https://github.com/poonamsnair/taxbench-au Hugging Face dataset: https://huggingface.co/datasets/Pn101/taxbench-au

---

Quick start

from datasets import load_dataset

# Load all four item families merged (default config):
ds = load_dataset("Pn101/taxbench-au")
print(ds["test"][0])

# Or a single item family:
t1 = load_dataset("Pn101/taxbench-au", "au_tax_rag_core_v0")
print(t1["test"][0])

Config name	Task code	Description
au_tax_rag_core_v0	T1	Source-retrieval and ATO citation
source_status_v0	T3	Source-status recognition (final / draft / withdrawn / guide / atoid)
information_sufficiency_v0	T4	Sufficiency-aware questioning over incomplete facts
professional_boundary_v0	T5	Professional-boundary safety under adversarial prompts
combined (default)	T1+T3+T4+T5	All four families merged; task field identifies the source

The cross-cutting outcome-alignment scorer (T2) is applied to every T1, T3, and T4 item that carries a provenance.reasoning_quote. T2 is a scoring protocol, not a separate item family, so it does not appear as its own config.

Canonical item count: 120. The four family configs (au_tax_rag_core_v0, source_status_v0, information_sufficiency_v0, professional_boundary_v0) contain 30 items each and are non-overlapping. The combined config is the union of the four families and is the default. If the Hugging Face viewer reports a different aggregate row count at the bottom of the page (e.g. 150 or 240), that is an HF cross-config aggregation artifact while it converts configs to Parquet — the canonical count of unique items remains 120 (40 dev + 20 validation + 60 test).

---

Splits

Split	Items	Per family	Use
dev	40	10	Prompt iteration and scorer development
validation	20	5	Ad-hoc cross-checks
test	60	15	Held-out evaluation; do not train or tune on this split

Splits are deterministic — assigned by alphabetical sort of item ids within each family. The test split was fixed before any system in the paper was evaluated against it.

---

Item schema

{
  "id":      "rag-borrowing-expenses-s25-25-amortisation",
  "dataset": "au_tax_rag_core_v0",
  "task":    "T1",
  "split":   "dev",

  "input": {
    "jurisdiction":    "AU",
    "tax_year":        "2024-2025",
    "workflow":        "rental-prep",
    "user_query":      "I paid $3,200 in lender's mortgage insurance...",
    "client_facts":    { /* structured facts, may be intentionally incomplete on T4 */ },
    "expected_sources": [
      { "doc_id": "s25-25 ITAA 1997", "topic": "borrowing expenses", "source_type": "Statute", "url": "https://www.ato.gov.au/..." }
    ]
  },

  "expected_output": {
    "expected_action":          "answer | ask_missing_fact | boundary_refusal",
    "expected_outcome_label":   "deductible_now | deductible_over_time | capital_not_immediate | not_deductible | ask_missing_fact | boundary_refusal",
    "expected_treatment_tags":  ["apportion_required", "div40_decline_in_value", "..."],
    "expected_review_required": true,
    "must_include":             ["s25-25", "spread", "five years"],
    "must_not_include":         ["full deduction in year one"],
    "forbidden_outcome_labels": []
  },

  "provenance": {
    "source_doc_id":   "ATO Borrowing Expenses Guide",
    "source_url":      "https://www.ato.gov.au/...",
    "scraped_at":      "2026-05-12",
    "example_quote":   "If your total borrowing expenses are more than $100...",
    "reasoning_quote": "Borrowing expenses are deductible under s25-25 over the lesser of 5 years and actual loan term."
  },

  "metadata": {
    "difficulty":    "easy | medium | hard",
    "topic":         "borrowing_expenses",
    "source_status": "final | draft | withdrawn | guide | atoid",
    "critical":      true
  }
}

The full JSON Schema lives in schema/item_schema.json. The shared output schema every evaluated system must emit lives in schema/output_schema.json.

must_include / must_not_include are not headline scorers. They are judge-context and audit metadata. The active scorers (run_completed, schema_valid, cited_expected_source, tax_position_aligned, boundary_safe) do not perform broad fuzzy keyword matching on these fields; the deterministic citation scorer is a hard substring check on expected_sources[].doc_id only, and the LLM judges read the gold provenance.reasoning_quote as the authority surface, not the must_* lists.

---

Scoring

The scoring harness is published with the GitHub repo (https://github.com/poonamsnair/taxbench-au). Five binary scorers:

Scorer	Type	Logic
run_completed	deterministic	!aborted && output non-empty
schema_valid	deterministic	shared-output parsed
cited_expected_source	deterministic	hard substring: gold doc-id literally in agent output
tax_position_aligned	LLM judge	answer aligns with cited ATO source's position (paper §5.3)
boundary_safe	LLM judge	agent maintained professional boundary (T5 only)

Default judge configuration used in the paper:

• tax_position_aligned → OpenAI gpt-4.1-mini, temperature 0, N=3 self-consistency, failure-biased tie-break
• boundary_safe → OpenAI gpt-4.1, temperature 0, N=3 self-consistency, failure-biased tie-break

If you evaluate an OpenAI model, swap the judges to a non-OpenAI provider to keep the cross-provider property.

---

Source corpus

Items are grounded in twenty-two harvested ATO documents spanning Taxation Rulings (TR 97/23, TR 2022/3, TR 2024/3, TR 2024/D1 draft), GST rulings (GSTR 2002/5, GSTR 2006/9), Practical Compliance Guidelines (PCG 2023/1), Practice Statements (PSLA 2008/14), Superannuation Guarantee Rulings (SGR 2009/2), ATO Interpretative Decisions (ATO ID 2002/672, ATO ID 2003/222), non-binding public guides, and additional CGT and FBT coverage.

Short attributed worked-example excerpts are retained inside provenance.example_quote and provenance.reasoning_quote for item construction. Bulk redistribution of ATO source material is not part of this release.

---

Intended use

• Diagnostic evaluation of AI tax-preparation assistants on authority-grounded behaviours.
• Comparative ranking of model / agent variants on the same items, scorers, and judges.
• Item-level error-class analysis (paper §8).

Out-of-scope use

• Not a tax-advice tool. Items are synthetic or de-identified composites; the gold answers cite public ATO materials but are not personal advice.
• Not a validation of autonomous lodgement. Commercial provision of tax-agent services for fee or reward generally requires registration with the Tax Practitioners Board under the Tax Agent Services Act 2009 (Cth).
• Not a generic legal-reasoning benchmark. TaxBench-AU is AU-only by construction.
• Not a fine-grained discrimination instrument. With 60 held-out test items, the benchmark is intended as a coarse comparative ranking; confidence intervals overlap on most tasks.

Known limitations

• Single jurisdiction (AU-only).
• Modest n (60 held-out items per system, ~15 per family).
• Asymmetric author audit — judge failures were re-graded under a fixed rubric; passes were not.
• Single judge model family — all judges in the paper use OpenAI models. Multi-judge ensembles would provide a stronger sensitivity signal.
• Strict source-string matching — produces some false negatives where the agent paraphrased a source name correctly (paper §8.3, §8.7).

Private data

None. All client fact patterns are synthetic or de-identified composites. No real taxpayer data is included.

---

Licence

• Benchmark items, schemas, prompts, documentation, metadata: CC BY 4.0.
• Code in the companion GitHub repo: MIT.
• ATO source materials referenced by URL or quoted in provenance.example_quote / provenance.reasoning_quote are © Australian Taxation Office for the Commonwealth of Australia and remain subject to the ATO's published copyright terms. See NOTICE.

This project is not affiliated with, endorsed by, or sponsored by the Australian Taxation Office, the Commonwealth of Australia, or the Tax Practitioners Board.

---

Citation

@misc{nair2026taxbenchau,
  title  = {TaxBench-AU: An Authority-Grounded Benchmark for Australian Tax Worked Examples},
  author = {Nair, Poonam},
  year   = {2026},
  note   = {Preprint}
}