How to Write an R&D Tax Incentive Technical Narrative

TLDR: An R&D tax incentive "technical narrative" is not one essay. The AusIndustry registration (online customer portal, current version released 15 August 2025) splits every core activity into separate fields: sources investigated, why a competent professional could not know the outcome, hypothesis, experiment, evaluation, conclusions and new knowledge.

Draft each field to its own question and its own character minimum, anchor every claim to evidence that existed at the time, and lodge by 30 April for a 30 June year end.

What a technical narrative actually is

It is the set of activity descriptions that make up your AusIndustry registration. There is no separate document called a technical narrative, and there is no free-text box where a story about your product will do the work.

The statutory unit is the core R&D activity, defined in section 355-25 of the Income Tax Assessment Act 1997. Eligibility is self-assessed per activity, never per project or per product.

That matters for drafting. A registration built around one knowledge gap per activity typically lands at 3 to 7 core activities. Fewer usually means the activities are too broad to describe an experiment; more usually means you have registered individual experiments as activities.

The fields you are writing into

The registration is a field-by-field questionnaire structured as Projects, then Core activities, then Supporting activities. The project layer carries commercial context. The core activity layer carries the science.

What the assessor is testing in each field

An assessor is checking two things, and each field carries one of them. The first limb of s 355-25(1) asks whether the outcome was genuinely unknowable in advance to a competent professional, and whether the work ran as a systematic progression from hypothesis to experiment, observation, evaluation and logical conclusions.

The sources investigated, why-unknown, hypothesis, experiment, evaluation and conclusions fields all serve that limb.

The second limb asks whether the activity was conducted for the purpose of generating new knowledge. That is what the new knowledge field carries, which is why it must describe knowledge rather than a shipped feature. Write to the statutory language rather than paraphrasing it away, because paraphrase is what loses defensibility under review.

How to write the hypothesis

A hypothesis is a specific, testable prediction with a stated basis in established science or engineering principle, capable of being validated or invalidated by the work you actually did.

In GQHC [2024] AATA 409 the tribunal rejected hypotheses it described as clumsy or vague because they were commercial aims rather than scientific propositions. A statement like "we hypothesised that a better matching engine would increase retention" is a business objective, not a hypothesis.

Use this shape:

1. State the mechanism you believed would work.
2. State the measurable outcome and the threshold that would count as confirmation.
3. State the conditions under which it was to be tested.
4. State the established principle it was based on.

So: "We hypothesised that quantising the ranking model to 8-bit integer weights with per-channel calibration would hold ranking quality within 1% NDCG@10 of the float32 baseline while reducing p99 inference latency below 40ms on our CPU-only inference fleet."

That version can be proved wrong. That is the test.

Timing matters more than wording

Active Sports Management [2023] treats the contemporaneous formation of a hypothesis as the essential starting point of the activities. A hypothesis written in April for work done the previous August is not evidence of a systematic progression: it is a description of one.

Write hypotheses when you start the work, in whatever tool your team already uses, then draft the registration field from that record.

Sources investigated and the worldwide knowledge test

These two fields are where most registrations are weakest, and they are what an assessor reads first.

Sources investigated

Name what you actually searched: academic literature, patent databases, vendor documentation, open-source implementations, standards, conference talks, clinical trial registries. Give enough specificity that the search is reproducible, and say what each source did and did not resolve.

In Absolute Vision Technologies [2022] AATA 2319 the absence of any literature review meant there was no demonstrated knowledge gap, and documents prepared afterwards for an internal review were not accepted as contemporaneous.

Why a competent professional could not know the outcome

The benchmark is a competent professional in the field with access to knowledge that is publicly available or reasonably accessible anywhere in the world. "New to our team" never satisfies s 355-25.

Structure the answer in three moves: what the public knowledge covered, precisely where it stopped, and why the gap could not be closed by reasoning or by applying an existing solution.

Novelty of context can help. Moreton Resources Ltd v ISA [2019] FCAFC 120 confirmed that applying an existing technology in a new context can be a core activity where the outcome was genuinely unknown and the purpose was to generate new knowledge. Coal of Queensland [2021] FCAFC 54 is the limit: applying known methods without a genuine unknown and a documented systematic progression still fails.

Describing the experiment

Describe the experiment as a method, not as a work log. An assessor is checking whether the work was a systematic progression from hypothesis to experiment, observation, evaluation and logical conclusions, or whether it was iteration until something shipped.

Include, at minimum:

• The independent variable you deliberately varied, and its levels
• The variables held constant, and how
• The dependent variable measured, and the instrument or metric
• The control or baseline condition
• Sample size, duration, or number of runs
• The pre-stated threshold for accepting or rejecting the hypothesis

Active Sports Management is the warning here: trial and error without isolable variables is not a repeatable, verifiable experiment.

If the activity ran several cycles, describe the cycle structure and note how each result changed the next hypothesis. That linkage is also your grouping test: experiments whose results inform each other belong in the same core activity.

Evaluation, conclusions and new knowledge

These three fields are distinct, and collapsing them is a common drafting failure.

Evaluation is what the data showed and how you analysed it. GQHC is explicit that observation and evaluation must be evidenced, not asserted, so reference the actual results: benchmark tables, run logs, statistical outputs.

Conclusions state whether the hypothesis held. Write negative results plainly. A disproved hypothesis still satisfies s 355-25, and a registration in which every experiment succeeded reads as reconstruction.

New knowledge states what is now known that was not before, framed as knowledge rather than as a shipped feature. "We now know that per-channel 8-bit quantisation degrades ranking quality beyond tolerance above 200 candidate items, and that a hybrid float fallback above that threshold preserves quality at 61ms p99" is knowledge. "We launched the new ranking service" is not.

Writing the supporting activity description

A supporting activity description must tie the work to a named core activity and explain the relationship. Under section 355-30 of the ITAA 1997, supporting activities are directly related to core R&D activities. Where the activity is in an excluded category, or produces goods or services, it must also be undertaken for the dominant purpose of supporting the core activity.

Group supporting activities into 2 to 4 broad categories per core activity, not one per ticket. Useful categories include:

• Prior-art or clinical-literature review
• Environment, infrastructure and test-harness setup
• Data acquisition, preparation or labelling
• Evaluation frameworks, benchmarking and instrumentation
• Ethics applications, protocol development or clinical governance
• Project management directly tied to the experiments

Titles like "Built dashboard for experiment 3" reveal per-ticket drafting and invite questions. See core vs supporting R&D activities for the boundary in more detail.

Language that weakens a narrative

Weak narratives fail on phrasing before they fail on substance. Marketing adjectives, undefined effort verbs and unfalsifiable claims give an assessor nothing to test against s 355-25, while statutory phrasing names a mechanism, a measurement and a threshold. Rewrite each of these patterns as you draft.

Two more habits to drop. Do not describe the product roadmap in the hypothesis field, and do not use "we needed to" where you mean "we predicted that", because need is a commercial state, not a scientific proposition.

A worked example: software core activity

The example below is a compressed version of a full registration, one field per line. For the eligibility boundary in software specifically, see the eligibility criteria for software companies.

Activity: Determining whether streaming incremental re-indexing can sustain sub-second search freshness at 50M documents.

Sources investigated: Lucene segment-merge documentation, published benchmarks from two open-source vector search projects, three ACM papers on incremental index maintenance, and vendor documentation for two managed search services. None addressed freshness under a write pattern of 4,000 updates per second against a corpus of this size with per-tenant filtering.

Why unknown: Published results assume append-heavy workloads. Our workload is update-heavy with high-cardinality tenant filters, and the interaction between segment merge scheduling and filter-cache invalidation was not documented in any accessible source.

Hypothesis: Deferring merges below a segment-count threshold while maintaining a per-tenant delta index will hold p95 query latency under 300ms and index-to-visible lag under 1 second at 4,000 updates per second.

Experiment: Four merge policies tested against a fixed 50M-document corpus and a replayed production write trace, hardware and query mix held constant, five runs each, measuring p95 latency and lag.

Conclusion: The hypothesis was rejected below a delta-index size of 2M documents and held above it, contrary to the predicted direction, and the crossover was traced to filter-cache thrash.

A worked example: clinical or health-tech core activity

Health-tech uncertainty often sits in the clinical outcome rather than the code. The app may be routine to build while the question of whether it changes a health outcome is a genuine gap.

Activity: Determining whether a personalised low-FODMAP reintroduction schedule delivered in-app reduces IBS symptom severity compared with a fixed schedule.

Hypothesis: Adaptive scheduling driven by logged symptom response will produce a mean IBS-SSS reduction at least 50 points greater than the fixed protocol at 12 weeks.

Experiment: Randomised parallel-group assignment, pre-registered primary endpoint (IBS-SSS at 12 weeks), stratified by baseline severity, with adherence tracked as a covariate.

A structured study measuring a clinical or behavioural endpoint against a pre-stated hypothesis is not excluded market research, even where it runs on A/B infrastructure. The discriminator is the pre-specified endpoint and hypothesis. The identical experiment measured on conversion, churn or revenue would be excluded under s 355-25(2)(a).

Clinical activity also draws integrity scrutiny. Taxpayer Alerts TA 2023/4 and TA 2023/5 target arrangements where the claimant is not genuinely the entity for whom the R&D is conducted, or where the spend is not at risk.

Keep clinical streams separate from technical build streams, because each is assessed against a different competent professional and a different literature.

How long should each field be?

Long enough to answer the field's question, and no longer. Each field carries its own character minimum in the portal, and a few hundred words per field is typical for a well-scoped activity.

Map the evidence before you write

Draft from records, not from memory. Every field should point at something that existed while the work was happening: prior-art searches before the activity started, hypothesis notes before the experiments, run logs and issue-tracker entries during, evaluation notes after each cycle.

Royal Wins Pty Ltd [2020] AATA 4320 held that without contemporaneous documentation that a hypothesis was developed and tested, the activities were not eligible regardless of how much work was done. Records should be kept for at least five years, consistent with ATO record-keeping guidance.

The narrative anchors the money

Suppose four engineers logged 1,240 of 3,800 total hours to one core activity, giving apportioned salary and on-costs of $310,000, and total notional R&D deductions across all registered activities of $600,000.

Assuming the activities are registered and eligible and the expenditure passes the at-risk and other integrity rules, a base-rate entity with aggregated turnover under $20M would claim the refundable offset at its 25% tax rate plus 18.5 percentage points: 43.5% of $600,000, or $261,000.

Every one of those hours must tie to a named registered activity.

This is where Rand does the mapping work: it builds each activity from the evidence trail your team already generates in GitHub and Jira, allocates hours against it, and drafts the registration fields from that trail rather than from a year-end interview.

Common drafting mistakes that trigger review

1. Registering the project instead of activities. One 900-word narrative covering a whole platform gives an assessor nothing to test against s 355-25.
2. Commercial hypotheses. If the prediction cannot be invalidated by data, it is not a hypothesis (GQHC).
3. No prior-art record. Absolute Vision failed on exactly this, and after-the-fact documents did not rescue it.
4. Only successes. Registrations without a single rejected hypothesis read as retrospective.
5. Mismatch between records and registration. Amending a deficient registration with documents prepared well after the income year is unlikely to be accepted.
6. Mislabelled excluded work. Internal administration software under s 355-25(2)(h), and reverse-engineering an existing product under s 355-25(2)(g), can never be core activities.

For the surrounding compliance picture, see how to document R&D activities for AusIndustry and what triggers an ATO review.

Before you lodge

Registration is due 10 months after the end of the income year, which is 30 April for a 30 June year end, and it must precede claiming the offset. Activity eligibility is decided by Industry Innovation and Science Australia and its delegates at AusIndustry; the expenditure and offset sit with the ATO.

Check the current form and guidance on the business.gov.au R&D Tax Incentive pages and confirm your dates against the registration deadline guide.

Written from contemporaneous records, the narrative takes hours. Written from memory in April, it takes weeks and defends poorly. Rand exists to close that gap by turning the evidence your team already produces into activity descriptions your Expert Reviewer can stand behind.