Quick answer: Encoding prescriptive, numeric code clauses into an executable rule set and running building models against them is generally unlikely to be a core R&D activity on those facts, subject to the activity's own facts and the statutory tests. Where a performance-based provision resists formalisation and nobody can say in advance whether a machine-checkable representation preserves its intent, an experimental activity may exist — but paragraph 355-25(2)(f) is in issue throughout, and you self-assess.
18 August 2026 — this article describes the current rules. The 2026–27 Federal Budget announced proposed R&DTI reforms for income years starting on or after 1 July 2028; those changes are not yet law.
Automated compliance checking sits close to an important exclusion, but the exclusion applies at the activity level. Paragraph 355-25(2)(f) of the Income Tax Assessment Act 1997 excludes activities associated with complying with statutory requirements or standards from being core R&D activities. It does not automatically exclude every activity involved in developing a compliance-checking product. Under the R&D Tax Incentive (R&DTI), each activity must be assessed on its own facts.
I write as a Registered Research Service Provider (RSP000047) at Lot Fourteen in Adelaide. We supply research capability and structure R&D activities; we do not give tax advice, and the rules below belong to the ATO, AusIndustry and the legislation, not to us.
Scope note: this article is about rule formalisation — turning code text into a representation a machine can evaluate, and evaluating models against it. It does not cover clash detection, model coordination or project controls (our separate Insight on BIM and clash detection), generative or parametric design, where the open question is the design space rather than the rule (our Insight on generative and parametric design), or physical certification testing of a building system (our Insight on NCC compliance testing versus developing the system you test). The property and construction page maps how they fit together.
Reading a Clause and Writing It Down Is Not Generating New Knowledge
Most of a rule engine is transcription. A clause says a corridor must be at least a stated width, a stair riser sits between two numbers, a room of a given class needs a light or ventilation area of a given fraction of floor area. Turning those numbers into predicates over model attributes is careful, high-volume work — and the outcome of each translation was determined before the work began, by the person who read the clause. Scale does not change that: two thousand encoded clauses are two thousand instances of one known operation, not an experiment with two thousand data points.
Nor does the technology decide anything — including where a language model reads the clause text. AusIndustry's AI-related activities sub-guide to the software development sector guide is direct: "Using an AI model or technique that is new to you does not, by itself, mean the activity is eligible for the program." The same guidance treats acceptance or functionality testing by established methods, where the expected outcomes are already known, as usually not core — which describes regression-testing a rule engine against clauses whose right answer is printed in the code.
The Statutory Test, Stated in Full
Section 355-25(1) defines core R&D activities as experimental activities:
• whose outcome cannot be known or determined in advance on the basis of current knowledge, information or experience, but can only be determined by applying a systematic progression of work that (i) is based on principles of established science and (ii) proceeds from hypothesis to experiment, observation and evaluation, and leads to logical conclusions; and
• that are conducted for the purpose of generating new knowledge, including new knowledge in the form of new or improved materials, products, devices, processes or services (legislation.gov.au; business.gov.au).
That is the test as the Act states it. The phrase "competent professional" is not in s 355-25(1); it is AusIndustry guidance wording, framing whether an outcome could be determined in advance by reference to what a competent professional in the field could have known.
Established science
A team can run a disciplined loop — encode, evaluate, measure agreement, revise — with no basis in established science at all, and still not satisfy the test. What can ground it is the science of knowledge representation: formal logic, the semantics of defeasible and conditional norms, decidability, and the measurement of inter-rater agreement. That theory is what makes a prediction possible about whether a representation can express a kind of provision.
Logical conclusions
A progression has to conclude something. "Version 7 agreed with the certifier more often" is a score. "A monotonic conjunction of predicates cannot represent this provision, because the determinations behave as a general norm defeated by prioritised exceptions" is a conclusion — it rules something out and dictates the next trial.
Both limbs are applied to a particular activity. Neither the platform nor the release is the unit of assessment.
Paragraph (f) Sits over the Whole Project, and Two Other Exclusions Reach It
Subsection 355-25(2) lists activities that cannot be core R&D activities:
Paragraph (f) — Statutory requirements and standards
Covers 'activities associated with complying with statutory requirements or standards', including: (i) maintaining national standards; (ii) calibrating secondary standards; and (iii) routine testing and analysis of materials, components, products, processes, soils, atmospheres and other things. Current business.gov.au guidance states that activities conducted to meet a requirement or standard contained in legislation, including activities directed by a regulator under legislation, cannot be core R&D activities. It also states that R&D activities are not excluded merely because they are conducted in a regulated environment. Whether a particular activity is associated with complying with a statutory requirement or standard is therefore assessed on its own facts, activity by activity, and self-assessed by the company. See the department's built environment sector guide.
Paragraph (g) — Reproduction of a commercial product or process
Excludes "any activity related to the reproduction of a commercial product or process: (i) by a physical examination of an existing system; or (ii) from plans, blueprints, detailed specifications or publicly available information". Both elements are required — reproduction, and of a commercial product or process. Rebuilding a competitor checking product's rule library from its published documentation may engage the exclusion where the activity is in fact directed to reproducing a commercial product or process.
Paragraph (e) — Commercial, legal and administrative aspects
Excludes commercial, legal and administrative aspects of patenting, licensing or other activities — rule libraries are frequently licensed, and that work is not part of any experiment.
None of this puts the work outside the program. Under s 355-30(1), supporting R&D activities are activities directly related to core R&D activities. Under s 355-30(2), where an activity (a) is an activity referred to in s 355-25(2), or (b) produces goods or services, or (c) is directly related to producing goods or services, it is a supporting R&D activity only if it is undertaken for the dominant purpose of supporting core R&D activities. Where running the engine over live projects or producing client reports produces, or is directly related to producing, services, the additional dominant-purpose test applies if the activity is being assessed as supporting R&D.
Where the Unknown Can Genuinely Sit
AusIndustry's guidance notes that AI-related activities "may meet the requirements of a core R&D activity where a technical hurdle exists and an expert in the field considers that only experimentation will determine if a proposed solution, or the way to develop a solution, can resolve the technical hurdle". Three things here can be open in that sense, in a way the transcription is not:
Whether the provision is representable at all: Performance-based provisions state an outcome and factors to have regard to, with no threshold to extract; whether any machine-checkable representation preserves that intent across cases can be genuinely open.
Which class of representation the norm belongs to: Monotonic rules, defeasible rules with priorities and hybrid schemes make different, testable predictions about which case patterns they can reproduce.
Where decidability stops: If a provision is only partly decidable from model facts, the boundary — the facts that must come from outside the model — is itself a finding.
A Worked Example — Hypothetical and Illustrative Only
The following scenario is entirely hypothetical. It illustrates where a technical boundary falls; it is not a statement that any of this work would be eligible, and eligibility would be self-assessed by the company against the statutory tests.
An Adelaide building-technology firm is developing a checking engine that evaluates digital building models against a code. Its prescriptive, numeric clauses were already encoded, agreeing with a registered certifier's recorded determination on 96% of a reference corpus. The work in question concerns a single performance-based provision expressed as an outcome plus factors to have regard to.
Baseline and target, fixed in March before any trial: Corpus: 240 assessed design cases across 18 buildings and three building classes, each with a registered certifier's recorded determination and stated basis, split 160 development / 80 held-out. Baseline: the code consultant's first hand-written encoding agreed with the certifier on 61% of cases and returned "non-compliant" on 22% of cases the certifier had passed. Target, fixed in advance: at least 90% agreement on the held-out 80 where a determination is returned, false "non-compliant" at or below 2%, a stated factor for every disagreement, and one representation holding across all three building classes with no per-class rules. Failure was defined in advance as missing that band, or reaching it only by adding case-specific exceptions.
Held constant across every trial: The corpus and its split, the model-extraction pipeline (validated separately), the code edition and jurisdictional variations applied, the certifier determinations as ground truth, and the evaluation harness. Varied: the knowledge representation only.
Why the outcome was not determinable in advance: Published work on automated code checking addresses prescriptive clauses, where a number in the text becomes a number in a predicate. This provision has no threshold to extract, and whether the determinations behave as a function of model-extractable facts at all — and if so, which representation class expresses them — had no predictive basis in existing knowledge.
Trial 1 — Monotonic first-order encoding (failed): The provision was translated into a conjunction of hard predicates, each "have regard to" factor a required condition. Adding that condition as a further conjunct fixed 19 of them but turned 14 previously matching cases into disagreements, raising net agreement to approximately 69%. What it ruled out: a monotonic, conflict-free rule set as a representation of this provision — the determinations behave as a general norm defeated by prioritised exceptions, not as a conjunction of conditions — and with it the working assumption that adding rules would converge.
Trial 2 — Defeasible representation with explicit priorities: Explicit priorities between the general norm and its exceptions. Held-out agreement rose to 84%, false "non-compliant" fell to 6%. But 11 of the 13 residual disagreements clustered on cases where the certifier's stated basis referred to a fact no model in the corpus carried — an operational management measure recorded in a fire-safety strategy document. What it ruled out: that the residual gap was a logic problem. It is a data-availability boundary: the provision is not a function of model-extractable facts alone.
Trial 3 — Defeasible representation + explicit "not determinable from model" outcome: Returned with the specific missing facts rather than forcing a compliant/non-compliant answer. On the held-out 80: 91% agreement where a determination was returned, false "non-compliant" 1%, 14% undetermined, and the missing-fact list matched the certifier's stated basis in 11 of 11 sampled undetermined cases — across all three classes, with no per-class rules.
Result actually reached: For this provision, a defeasible representation with an explicit undetermined outcome reaches the band, and the limit of machine decidability is characterised by an enumerable list of facts absent from the model schema. Two further performance requirements with longer factor lists returned undetermined on more than 40% of cases in first trials, and the year closed without establishing whether that is the same boundary or a different one.
Where the boundary falls: Of roughly 1,300 hours: about 290 sit in the hypothesis, the three representation trials and the evaluation that ruled things out; about 150 in assembling and normalising the corpus, the split and the harness, which exist only to feed those trials, so whether they attach turns on s 355-30 including the dominant-purpose test; about 430 in the extraction pipeline, interface, reporting and deployment, applying known methods to known outcomes; about 230 in encoding the prescriptive numeric clauses and running the engine over live projects to produce compliance reports — activities associated with complying with statutory requirements or standards under s 355-25(2)(f), which also picks up routine testing and analysis at (f)(iii); and about 200 in rebuilding a competitor product's rule library from its published documentation, potentially engaging s 355-25(2)(g) where the activity constitutes reproduction of a commercial product or process. Paragraph (f) has to be weighed against the 290 trial hours too, on their own facts. How each stream is treated in a registration and in the company's tax return is for the company and its registered tax agent to determine.
Where an RSP Fits
Qualifying expenditure incurred to a non-associate RSP may still form part of the offset where total notional deductions are below the usual $20,000 threshold, provided the services are within a research field for which the RSP is registered (business.gov.au) — see claiming R&D under $20,000. Using an RSP does not guarantee eligibility — you still self-assess. Offset rates, the refundable and non-refundable tiers and the entitlement rules are covered separately in our refundable vs non-refundable offset article; the general framing sits in our software and AI and what does not qualify pages. Judging whether only experimentation can resolve a hurdle is a question about the state of knowledge in a field, and the kind an RSP is registered to provide R&D services within specific research fields and may assist companies to plan and conduct R&D activities; eligibility remains self-assessed by the company.
Frequently Asked Questions
Q: Is automated code compliance checking eligible for the R&D Tax Incentive?
A: Encoding prescriptive, numeric clauses into an executable rule set and evaluating models against them is generally unlikely to be a core R&D activity on those facts, subject to the activity's own facts and the statutory tests — the right answer to each translation was determined by reading the clause. Separately, s 355-25(2)(f) excludes activities associated with complying with statutory requirements or standards from being core, and its reach over a compliance-checking project is a question of fact for each activity. You self-assess.
Q: Does s 355-25(2)(f) exclude compliance software from the R&DTI entirely?
A: No. Paragraph 355-25(2)(f) says such activities are not core R&D activities. Such an activity may still qualify as a supporting R&D activity where it is directly related to a core R&D activity and, because s 355-30(2) applies, is undertaken for the dominant purpose of supporting that core activity. Whether a given activity is "associated with complying" is assessed activity by activity.
Q: Can formalising a performance-based provision be a core R&D activity?
A: It may be, depending on the facts. Where a provision states an outcome and factors rather than a threshold, whether any machine-checkable representation preserves its intent — and which representation class can express it — may not be determinable in advance from current knowledge, information or experience. Section 355-25(1) also requires a systematic progression of work based on principles of established science that proceeds from hypothesis to experiment, observation and evaluation and leads to logical conclusions, conducted to generate new knowledge, and paragraph (f) remains in issue.
Q: We rebuilt our rule library from another product's published documentation. Does that matter?
A: Paragraph 355-25(2)(g) excludes any activity related to the reproduction of a commercial product or process by physical examination of an existing system, or from plans, blueprints, detailed specifications or publicly available information. Both elements are required: reproduction, and of a commercial product or process. A published academic method, a standards body's machine-readable schema or an open-source rule format is not by itself a commercial product or process, so whether the paragraph applies turns on the facts of the particular activity.
Sources & Further Reading
legislation.gov.au — Income Tax Assessment Act 1997 — Div 355, incl. ss 355-25 and 355-30
ATO — Tax Reform: better targeting the R&D Tax Incentive — the 2026-27 Federal Budget announcement of 12 May 2026, the proposed changes, and the ATO's statement that the measure is not yet law
industry.gov.au — Research and Development Tax Incentive — the proposed commencement: changes will apply to income years starting on or after 1 July 2028
Related: property and construction · software and AI · what does not qualify · what an RSP is · claiming R&D under $20,000 · refundable vs non-refundable offset · R&D Tax Incentive in Adelaide
Talk to Ignition Research before you register a year of automated compliance-checking work — as a Registered Research Service Provider at Lot Fourteen in Adelaide, we work with construction-technology and regulatory-technology teams on what is genuinely unknown about representing a provision, what the evaluation is actually measuring, and what a progression from hypothesis to logical conclusions looks like in knowledge representation. We are not a registered tax agent: your company self-assesses and remains responsible for its own claim, with advice and lodgement handled by your tax adviser. Get in touch.
This article is general information from a Registered Research Service Provider about the R&D Tax Incentive. It is not tax, legal or financial advice; eligibility depends on your circumstances and you should self-assess and seek your own advice.
Thinking about a project like this?
If you're weighing up an AI, software or technical improvement project and can't tell yet whether it's implementation or research, start with a quick read on where it sits.