Division 7A loan review

+----------------------------------------------------------------------+
|                       Division 7A loan review                        |
+----------------------------------------------------------------------+
|          ITAA 1936 Div 7A: s 109N terms, s 109E repayments           |
+----------------------------------+-----------------------------------+
| DR  what it gives you            | CR  what it needs                 |
+----------------------------------+-----------------------------------+
| s 109N(2) benchmark rate         | loan register CSV                 |
| s 109N(1) verdict, limb by limb  | the year of income to review      |
| s 109E minimum yearly repayment  | operator-asserted statutory facts |
| shortfall and exposure flag      | a reviewed rate for that year     |
| statutory trace to re-perform    | a human to decide the rest        |
+----------------------------------+-----------------------------------+

Distribution div7a-loan-review, import package div7aloan, command div7a-loan-review.

Experimental review aid. Not a Division 7A determination.

v1 reviews the s 109N terms of a private-company loan and the s 109E minimum yearly repayment for an operator-supplied amalgamated loan. It does not form amalgamated loans, does not classify payments, does not do unpaid present entitlements, and does not apply s 109RB.

The australian-accounting repository contains the maintained source. The div7a-loan-review distribution, command and div7aloan import package remain compatibility identifiers.

A rising benchmark rate raises the minimum yearly repayment on existing complying loans, not just on new ones. s 109E(6) uses the current year's benchmark rate, not the rate written into the agreement. A 7-year loan signed in 2021-22 at 4.52% is measured for 2026-27 against 8.77 per cent, and the repayment that was right when it was written is no longer enough. Nothing in the process tells you. A loan that has met its repayments for 4 years can fall short in the fifth on unchanged terms.

No install, just the rates and the reasoning: https://duguid.com.au/rates/div7a-benchmark-rate/

Citation: CITATION.cff; release: v0.1.1.

What it does

Three questions, one job.

1. rate - the s 109N(2) benchmark interest rate for a year of income, with provenance: RBA table, series, month, and the date this repository last reviewed the figure.
2. gate - whether a loan meets every limb of s 109N(1), on facts the operator asserts. Written agreement, terms in place before the lodgment day, interest rate at or above the benchmark, term within the s 109N(3) maximum.
3. myr - the s 109E(5) minimum yearly repayment for a later year of income, the payments applied against it, and any shortfall.

review runs the gate and the repayment together over a register.

What it refuses

These are features. The engine returns UNKNOWN or REFUSED with a reason rather than a number it cannot stand behind.

An UNKNOWN is never softened into a best-effort guess, and never coerced to false. A limb that has not been established is a different finding from a limb that has failed, and only one of those is a breach.

The shortfall is not the dividend

Where the repayment falls short the engine reports an experimental deemed-dividend exposure equal to the shortfall. Read it as a review aid.

s 109E(2) makes the dividend the shortfall subject to s 109Y, which caps Division 7A dividends at the company's distributable surplus, and s 109E(1)(d) removes the dividend entirely where s 109Q applies. Neither is modelled here, and neither is s 109RB. This repository does not write "the ATO will assess $X", and neither should anything built on it.

Excel workbook

No Python? workbooks/div7a-loan-review.xlsx runs the same review in ordinary worksheet formulas: paste the register, pick the year of income, read each s 109N(1) limb, the gate verdict, the s 109E(6) repayment and the shortfall per loan. REFUSED and UNKNOWN stay REFUSED and UNKNOWN. It is macro-free, needs desktop Excel for Microsoft 365 or Excel 2024, and is held to this engine's answer by tests/test_workbook.py. See workbooks/README.md.

Install

pip install div7a-loan-review

Python 3.10 or later. No runtime dependencies. To run the examples from a source checkout:

git clone https://github.com/ryanduguid/australian-accounting.git
cd australian-accounting/packages/div7a-loan-review
pip install .

Use

div7a-loan-review rate --year 2026-27

div7a-loan-review gate --input examples/sample_loans_myr_met.csv

div7a-loan-review myr --input examples/sample_loans_myr_met.csv --year 2026-27

div7a-loan-review review --input examples/sample_loans_mixed.csv --year 2026-27

Add --format json for machine-readable output, --trace on review to print the statutory trace for every row, and --rates-override FILE to supply a reviewed benchmark rate for a year outside the frozen table.

Which year the gate reads its benchmark from

gate without --year anchors each row to its own year_loan_made. gate --year 2026-27 uses the nominated year's benchmark instead.

The interest-floor interpretation is unresolved. ATO guidance requires the benchmark for each year of the loan. The model compares one supplied rate with one selected benchmark. Its COMPLYING label does not establish that the agreement satisfies the annual requirement. Review the agreement and relevant years before relying on it. The source review preserves the earlier interpretation and records this conflict.

review always anchors the gate to year_loan_made and works the repayment for --year.

Exit codes

The register CSV

One amalgamated loan per row. The run is refused if a required column is absent: reading a missing column as blank would turn every absent fact into unknown and produce a file of UNKNOWN verdicts that looks like a considered review rather than a mis-shaped file.

Booleans are true, false or unknown. A blank cell is unknown. Amounts and rates are decimal strings. Rates are fractions, so 8.77% is 0.0877, not 8.77. Years of income are written 2026-27.

Required by gate, myr and review:

Also required by myr and review:

Optional:

Two samples ship with the repository:

• examples/sample_loans_myr_met.csv - 3 complying loans that run to a verdict with no confirmation flags.
• examples/sample_loans_mixed.csv - UNKNOWN rows, a loan whose rate is below the benchmark, a 25-year loan with 109% cover, a nil remaining term, and 2 skipped rows.

Both are fabricated. Every identifier is synthetic.

Benchmark rates

s 109N(2) sets the benchmark interest rate as the RBA Indicator Lending Rates - Bank variable housing loans figure last published before the start of the year of income. A year of income starts on 1 July. The RBA publishes statistical table F5 in arrears, in the first week of the following month, so the last figure published before 1 July is the May figure.

The June figure is published in early July, after the year has already begun, and is the usual wrong rate. For 2025-26, May 2025 was 8.37% and June 2025 was 8.27%. 8.37 is the benchmark rate.

The table lives in div7aloan/data/benchmark_rates.csv with its own reviewed_until and reviewed_on header, last reviewed 28 August 2026. Nothing in this repository reads the network, in tests or at runtime.

A year outside that coverage is UNKNOWN. To review one, supply a reviewed override:

{
  "verified_until": "2027-28",
  "citation": "RBA table F5 series FILRHLBVS, May 2027 figure, read 2027-07-02",
  "rates": [
    { "year_of_income": "2027-28", "rate": "0.0850", "rba_month": "2027-05" }
  ]
}

The file is refused without both verified_until and a non-empty citation, and refused if it lists a year past its own verified_until. The point of an override is that a human went and read the figure; a file that does not say who checked what, and how far, is not a review.

Which table produced the number

Every JSON result carries a manifest, naming the rate tables the answer was actually computed from and the SHA-256 of what was read:

"manifest": {
  "rate_table_uris": [
    { "uri": "div7aloan/data/benchmark_rates.csv", "sha256": "84264b28..." }
  ]
}

The digest is emitted by the lookup that read the file, not attached by hand, and it is taken over the decoded text so a CRLF checkout and an LF checkout of the same reviewed table agree. Two outputs with the same digest were computed from the same table; an output whose digest has moved was not, whatever its provenance fields still say. A reviewed override is listed after the frozen table as override:<file name>: the name and digest identify it without putting the operator's directory layout into the result.

An UNKNOWN rate still names the table it was looked for in, because the absence of a row in that table is the answer. An empty list means no rate table was consulted at all.

Reason codes

REFUSED and UNKNOWN results carry reason_codes beside reasons: one stable token per reason, in the same order.

"reasons": ["remaining_term_years was not established."],
"reason_codes": ["REMAINING_TERM_UNKNOWN"]

The prose is written for a person and changes with the wording of the Act or the engine. The codes are the half a caller branches on. There are three prefixes, one per question:

A gate result carries one GATE_* code per limb that did not pass, so a caller can tell a loan that fails the interest limb from one whose term was never supplied without reading the prose.

The vocabulary is ReasonCode in div7aloan/verdicts.py; a code is not removed without a version bump, and a reason built without one fails at construction rather than reaching a caller with nothing to branch on. tests/test_reason_codes.py holds the other half of that contract: every published code is reachable from some scenario, so the vocabulary cannot promise a token no engine emits.

Arithmetic

decimal.Decimal throughout, built from strings. No float, no numpy, no pandas. Intermediate arithmetic in the s 109E(6) formula runs at 60 significant digits in a local decimal context, so a result never depends on the ambient context a caller happens to have set, and is quantised once at the end.

Money is quantised to cents with ROUND_HALF_UP. The Act prescribes no rounding for the s 109E(6) amount, so this is the engine's documented choice rather than a statutory rule, and it is not the ATO calculator's rounding adopted silently. Every result says so, in its statutory trace and in the JSON rounding field.

JSON amounts are quoted decimal strings, never JSON numbers. Dates are ISO 8601. Verdicts are enums. The emitter refuses to write a float, and the test suite fails if one ever appears.

Library API

The surface the MCP adapter will import. Stable; the modules behind it are not.

from decimal import Decimal
from div7aloan import (
    GateFacts, MyrFacts, benchmark_rate, complying_loan_gate,
    minimum_yearly_repayment, parse_year, review_register,
)

year = parse_year("2026-27")

rate = benchmark_rate(year)
# rate.verdict, rate.rate (Decimal), rate.rba_month, rate.reason

gate = complying_loan_gate(GateFacts(
    written_agreement=True,
    terms_in_place_before_lodgment_day=True,
    maximum_term_years=Decimal("7"),
    secured_by_registered_mortgage_over_real_property=False,
    interest_rate_for_years_after_year_loan_made=Decimal("0.0827"),
    year_loan_made=parse_year("2023-24"),
))
# gate.verdict, gate.limbs, gate.reasons, gate.caveats

result = minimum_yearly_repayment(MyrFacts(
    year_of_income=year,
    amalgamated_loan_unpaid_at_end_of_previous_year=Decimal("100000.00"),
    remaining_term_years=Decimal("5"),
    payments_applied_during_the_year=Decimal("25556.00"),
    gate_result=gate,
    year_loan_made=parse_year("2023-24"),
))
# result.verdict, result.myr_required, result.shortfall,
# result.experimental_deemed_dividend_exposure, result.statutory_trace

report = review_register(rows, year)
# report.summary, report.lines, report.total_exposure

Every result carries a verdict, amounts as Decimal, provenance, and a refusal or unknown reason where one applies.

Evaluation pack

evaluation/div7a_myr/ holds fabricated fixtures, pinned expectations, and 3 minimum yearly repayments worked by hand from the statutory formula with the arithmetic shown. A reviewer who knows s 109N and s 109E can re-perform every one on paper without running Python.

It also records where this repository's reading of s 109N(1)(b) differs from common practice, and why the Act was followed.

Tests

python -m pytest -q

Every frozen rate year, every verdict, every refusal, the 2025-26 May-vs-June trap, the formula against exact rational arithmetic, and a check that no amount is ever emitted as a JSON number.

GATES.md is the completion ledger: one gate per acceptance criterion, each naming the command that decides it and carrying the recorded evidence of its last run.

Provenance

Written against the Income Tax Assessment Act 1936 as compiled on the Federal Register of Legislation, C1936A00027, in force 1 July 2026, Part III Division 7A. Sections 109D, 109E, 109N, 109P and 109R were read in full from the compilation, and the s 109E(6) formula from the formula image in that text. It is not a port of any existing Division 7A calculator.

Where the compiled Act and a secondary source disagree, this repository follows the Act, cites the section, and records the disagreement.

The full source trail is in docs/primary-source-review-2026-08-31.md: which sections were read, how the s 109E(6) formula was recovered from the compilation's own image, the reviewed position on s 109N(1)(b), and what this engine deliberately does not establish.

Related

• DISCLAIMER.md - read it.
• SECURITY.md
• Benchmark rate explainer: https://duguid.com.au/rates/div7a-benchmark-rate/
• Australian tax AI agents: https://duguid.com.au/tools/australian-tax-ai-agents/

From an AI coding agent, the same engine runs through aus-accounting-mcp. The adapter and engine now share this monorepo but remain separately versioned Python distributions.

Author and licence

Ryan Duguid. Written independently, in his own time, on his own equipment. Provisional member of Chartered Accountants Australia and New Zealand; that membership is not an endorsement of this software by CA ANZ.

MIT. See LICENSE.