The decision of the Administrative Appeals Tribunal (AAT) in H2O Exchange Pty Ltd and Innovation and Science Australia (Taxation) [2019] AATA 4195, together with its subsequent affirmation on appeal by the Federal Court of Australia in H2O Exchange Pty Ltd v Innovation and Science Australia [2021] FCA 11, stands as a watershed authority in Australian taxation jurisprudence governing the eligibility of software development projects under the Research and Development Tax Incentive (RDTI)1. The dispute arose from an application by H2O Exchange Pty Ltd (H2O), an Australian software and market-platform enterprise, to register activities associated with the development of an online water exchange platform for the purposes of claiming notional deductions and the corresponding R&D tax offset3.
H2O sought to develop an integrated digital marketplace capable of facilitating the purchase and sale of statutory water rights—specifically permanent water entitlements and temporary water allocations—across the Murray-Darling Basin spanning the jurisdictions of New South Wales, Victoria, and South Australia3. In its registration application submitted to AusIndustry, H2O characterised the architectural design and software development of its core trading platform and underlying database infrastructure as an eligible Core R&D Activity3. Specifically, H2O registered two core activities for each of the 2013-14 and 2014-15 income years—Activity 1.1 (“development of core trading platform and data base, CRM and various APIs”) and Activity 1.2 (“Integration of matrix rules”), the latter concerned with embedding the complex and frequently changing trading rules of the various water authorities so that only complying trades could proceed—and the expert evidence broke Activity 1.1 down into five functional elements: first, a web-based trading platform accessible directly by retail customers or intermediaries (brokers) to list buy and sell orders for water entitlements and water allocations; second, a centralised membership database and Customer Relationship Management (CRM) system engineered to record client identity data and detailed inventories of their statutory water assets; third, a binding legal framework whereby the placing of buy and sell orders incorporated pre-acceptance of a binding legal agreement, or a binding contract would be formed when compatible buy and sell orders were matched; fourth, a real-time financial settlement and transaction fee payment processing engine; and fifth, complete end-to-end automation of the statutory water conveyancing and title registration lifecycle, encompassing automated electronic lodgement and updating of title registries and water accounting systems administered by state water authorities across New South Wales, Victoria, and South Australia (an element described in the Objectives section of the application rather than in the Activity 1.1 description itself)3.
Although the respondent registered the activities on lodgement (on 23 July 2014 for the 2013-14 income year and on 14 July 2015 for the 2014-15 income year), Innovation and Science Australia (ISA, subsequently renamed Industry Innovation and Science Australia (IISA)), assisted by AusIndustry, later examined the registrations and on 20 May 2016 made a finding under section 27J of the IR&D Act that Activities 1.1 and 1.2 were not core R&D activities, and therefore did not constitute eligible “R&D activities” within the meaning of Division 355 of the Income Tax Assessment Act 1997 (Cth) (ITAA 1997)1. Following an unfavourable internal review decision made by a delegate of ISA on 24 April 2017 under section 30D(2) of the IR&D Act—which confirmed that the activities were not core R&D activities and also found that they were not supporting R&D activities—H2O applied to the Administrative Appeals Tribunal for merits review under section 30E of the IR&D Act3. In H2O Exchange Pty Ltd and Innovation and Science Australia [2019] AATA 4195, delivered on 14 October 2019 by Deputy President B W Rayment OAM QC, the Tribunal affirmed the reviewable internal review decision in its entirety3. H2O subsequently appealed the Tribunal’s decision on questions of law—ultimately pressing three grounds: that the Tribunal acted irrationally and illogically in accepting the respondent’s expert evidence, that it denied H2O natural justice by treating the presence or absence of artificial intelligence (AI) research as a relevant issue without notice, and that it denied natural justice by proceeding with the review despite a purported variation application—to the Federal Court of Australia under section 44 of the Administrative Appeals Tribunal Act 1975 (Cth)2. In H2O Exchange Pty Ltd v Innovation and Science Australia [2021] FCA 11, Stewart J dismissed the appeal with costs on 22 January 2021, finding that none of the three grounds disclosed an error of law and that the Tribunal was entirely justified in accepting the expertise of the respondent’s expert witness2.
The Australian R&D Tax Incentive operates under a co-regulatory statutory architecture split between the Industry Research and Development Act 1986 (Cth) (IR&D Act) and Division 355 of the Income Tax Assessment Act 1997 (Cth) (ITAA 1997)1. Under Part III of the IR&D Act, the Board of Innovation and Science Australia (assisted by AusIndustry within the Department of Industry, Science and Resources) administers the registration of R&D activities and holds statutory authority to make binding findings as to whether registered activities constitute “core R&D activities” or “supporting R&D activities”3. Under section 355-705 of the ITAA 1997, those findings bind the Commissioner of Taxation (Australian Taxation Office), who administers the calculation and substantiation of eligible notional deductions and the resulting refundable or non-refundable tax offsets under Division 355 of the ITAA 19974.
The substantive legal threshold at the heart of H2O Exchange is codified in section 355-25(1) of the ITAA 1997, which defines “core R&D activities” through a cumulative, multi-limb statutory test3. To qualify as a core R&D activity, an activity must be an experimental activity that satisfies two overarching statutory limbs3. Under the first limb, codified in section 355-25(1)(a), the activity must be one 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, pursuant to subparagraph (i), is based on principles of established science and, pursuant to subparagraph (ii), proceeds from hypothesis to experiment, observation and evaluation, and leads to logical conclusions3. Under the second limb, codified in section 355-25(1)(b), the experimental activity must be conducted for the purpose of generating new knowledge, including new knowledge in the form of new or improved materials, products, devices, processes, or services4. Furthermore, to retain eligibility as a core R&D activity, the activity must not fall within any of the statutory exclusions enumerated in section 355-25(2) of the ITAA 199711.
Complementing this primary definition, section 355-30 of the ITAA 1997 defines “supporting R&D activities” as activities directly related to core R&D activities, or, where the activity falls within the section 355-25(2) exclusions or produces goods or services, activities undertaken for the dominant purpose of supporting core R&D activities5. Because supporting R&D activities are legally derivative of core R&D activities, a taxpayer must establish the existence of at least one valid core R&D activity under section 355-25(1) as a condition precedent to claiming any supporting expenditure5. Current legislative reform proposals announced in the Federal Budget propose narrowing the incentive from 1 July 2028 by abolishing eligibility for supporting R&D activities altogether while increasing the offset rates on core R&D expenditure by 4.5 percentage points (a measure that is not yet law), further magnifying the legal imperative for claimants to satisfy every element of section 355-25(1)5.
In affirming the internal review decision, the Tribunal was not satisfied that the registered activities associated with the development of the online water exchange platform met the statutory criteria of section 355-25(1)(a) of the ITAA 19973. The Tribunal determined that, as to Activity 1.1, the outcome of the web-based trading platform, membership database and CRM, binding legal framework and real-time settlement elements was known to be achievable—parts of the activities being “simple and easily accomplished by a competent professional” and, to a large extent, involving choosing between a suite of known software tools and incorporating them into the platform, even if that work was detailed or complicated—while complete automation of conveyancing was known not to be achievable because New South Wales and South Australia did not offer online registration of transfers, an uncertainty stemming from government procedures outside H2O’s control rather than uncertainty of outcome under section 355-25(1)(a); as to Activity 1.2, the Tribunal was not satisfied on the evidence that H2O had embarked during the relevant income years on experimentation with AI software that had an uncertain outcome3.
Crucially, the Tribunal relied on the expert evidence of Mr Harding, an industry specialist called by Innovation and Science Australia who had worked at competing water marketplace Waterfind Pty Ltd from July 2014 to March 2017, first as a conveyancing officer and then as manager of its Market Services arm; his evidence was that the Waterfind platform already automated the matching of compatible buy and sell orders, contract generation and execution and the creation of a binding legal framework, that a trading-rules matrix could readily be systematised in a database, and that complete automation of conveyancing was not achievable because of external government procedures, and the Tribunal found that H2O’s own expert, Dr Fearn, was in general agreement with Mr Harding on the outcome question3. Furthermore, although H2O’s registration applications stated hypotheses for each activity, the Tribunal observed that no information had been provided about what experimental activities using AI software were embarked upon, whether they were based on principles of established science, whether a systematic progression of work was involved, and what hypotheses were considered; H2O also called no witness from Veritec, the technology partner that actually performed the work, and produced only limited documentation from it3.
The foundational legal defect in H2O’s claim was its conflation of commercial novelty and functional system integration complexity with genuine scientific or technological uncertainty under section 355-25(1)(a) of the ITAA 19973. H2O’s registration applications asserted that there was no “off the shelf” product of any kind that could achieve its platform objectives, and the respondent’s own expert accepted that there may not have been an existing service combining all of the elements of Activity 1.1—specifically, uniting live order-book matching, CRM asset tracking, automated contract execution, real-time fee settlement, and automated multi-state title conveyancing across New South Wales, Victoria, and South Australia3.
The Tribunal rejected the proposition that commercial differentiation or market-first functionality satisfies the statutory test for a Core R&D Activity3. As the Tribunal’s legal rationale demonstrates, the mere fact that a software product or service is new, bespoke, or functionally different from existing commercial offerings in the marketplace does not establish that its technical outcome cannot be known or determined in advance3. Under section 355-25(1)(a), the relevant inquiry is not whether a specific commercial system has previously been assembled in that exact configuration, but whether a competent professional in the relevant field could deduce or achieve the technical outcome using extant baseline knowledge, information, or experience without resorting to scientific experimentation3.
In affirming this principle, the Tribunal drew a doctrinal distinction between the difficulty or complexity of a task and genuine uncertainty of outcome, holding that a task which is difficult will not necessarily involve uncertainty of outcome, and that uncertainty stemming from an absence of control by the applicant (such as dependence on governments changing their registry procedures) is not uncertainty of outcome consistent with section 355-25(1)(a); the Federal Court’s appeal judgment was confined to questions of law and did not revisit this analysis4. Complex software systems integrating disparate external databases, state water registries, and regulatory rules inevitably exhibit implementation friction, edge-case bugs, and performance variability under load3. Resolving those implementation challenges requires skilled software architecture, business-rule mapping, and iterative performance tuning, yet the underlying computer science principles governing database transactions, API communication, and state-machine workflows remain entirely known4. Genuine R&D uncertainty under section 355-25(1)(a) arises only where established computer science principles, algorithms, or mathematical models are incapable of predicting whether or how a technical objective can be achieved due to conflicting technical constraints or uncharacterised variable interactions3.
Because four of the five elements of Activity 1.1 relied on known software tools and established technologies (such as PostgreSQL and Java, which H2O itself intended to use, and trading, CRM and settlement functions already found in online share trading, CRM products and e-commerce platforms), the Tribunal concluded that their outcome was known to be achievable, whereas the fifth element—complete automation of conveyancing—was known in advance not to be achievable because two of the three States required manual processing, so that in neither case was the outcome unknown3. Complicated engineering work, regardless of scale or commercial difficulty, does not equate to uncertainty of outcome3.
A pivotal evidentiary dimension of H2O Exchange [2019] AATA 4195—and the primary ground of H2O’s subsequent appeal in [2021] FCA 11—centred on the qualification required of an expert witness to opine on what a “competent professional” could accomplish3. The statute does not define “current knowledge, information or experience”; the Tribunal observed that the word “current” directs attention to the knowledge, information and experience available when the project was initiated (or, for the second income year, at the beginning of that year), and in applying the test it found that parts of H2O’s activities were “simple and easily accomplished by a competent professional” in the relevant domain3.
To challenge H2O’s assertion of uncertainty, Innovation and Science Australia adduced expert evidence from a witness holding a Bachelor of Finance degree who possessed extensive commercial and operational experience trading water rights in Australia3. H2O challenged this witness’s expertise in relation to software development (without otherwise cross-examining him on the contents of his report), relying on his acceptance in cross-examination that he had not studied software design, had never been hired as a software engineer, and had never written a line of code3. H2O contended that an individual incapable of writing software code could not legally qualify as a “competent professional” to opine on whether the outcome of a software development activity was determinable in advance3.
The Tribunal relied on Mr Harding’s evidence, and the Federal Court rejected H2O’s contention that it was irrational or illogical to do so3. The Federal Court examined the substance of the witness’s experience, noting that in 2006 he had assisted with the design and testing of a replacement CRM program (including its database component), had designed and project-managed software platforms at Global Commodities, and from December 2014 managed the Market Services arm of Waterfind Pty Ltd, covering its conveyancing and policy teams, later performing user acceptance testing on new features of the Waterfind platform; the Court held that this experience provided a proper basis for specialised knowledge3. Crucially, the Waterfind platform was an existing online water marketplace that already performed automated matching of compatible buy and sell orders to form binding legal contracts3. Furthermore, the witness possessed an exhaustive command of Murray-Darling Basin water market regulations, state-by-state conveyancing rules across New South Wales, Victoria, and South Australia, participant trading behaviour, and market clearing systems3.
The acceptance of this non-programmer’s expert opinion on what a “competent professional” could have known exposes a profound strategic vulnerability for R&D claimants in Australia3. H2O brought this evidentiary defeat upon itself through the manner in which it framed its registered Core Activity3. By defining its Core Activity at the functional and business-logic layer—listing orders on a web portal, recording client water assets in a CRM, forming binding contracts upon order match, settling fees, and automating state conveyancing workflows—rather than isolating a deep computer science or algorithmic problem at the code or systems-architecture layer, H2O placed the dispute squarely within the functional domain of water-market operations3. Once the activity was framed in functional terms, a domain specialist who had administered, and performed user acceptance testing on new features of, a competing platform (Waterfind) was fully qualified to testify that those exact functional mechanics were already known, understood, and straightforward to specify and implement using standard software development teams3.
Even if a taxpayer identifies a complex technical challenge, section 355-25(1)(a)(i) and (ii) of the ITAA 1997 mandate that the outcome must only be determinable by applying a “systematic progression of work” that is based on principles of established science and proceeds sequentially from hypothesis to experiment, observation, and evaluation, leading to logical conclusions3. In H2O Exchange, the Tribunal decided the matter principally on the outcome question, but observed that, even had it been satisfied that H2O embarked on uncertain AI experimentation for Activity 1.2, no information had been provided about what experiments were undertaken, whether they were based on principles of established science, whether a systematic progression of work was involved, and what hypotheses were considered3.
In the context of Division 355, a valid hypothesis is not a commercial aspiration or a high-level engineering statement that a proposed software system will satisfy user or regulatory requirements3. Rather, a compliant hypothesis must be a concrete, falsifiable proposition founded on principles of established science (such as computer science, formal logic, queuing theory, or computational mathematics) that postulates a specific causal relationship between independent and dependent technical variables3. Although H2O’s applications did state hypotheses—for Activity 1.1, that an internet-based application able to trade water entitlements interactively could be created so that all settlement and title processes occur in real time, and for Activity 1.2, that real-time updates of trading rules could populate the platform so that only eligible trades occur—the Tribunal noted that the registered activities mentioned only a desired result, with no indication of how it would be attempted to be achieved3. Instead, H2O operated on the broad developmental objective that an integrated water trading and automated conveyancing platform could be constructed to interface with state water registries3.
Consequently, no hypothesis-driven experimental protocol was established on the evidence: neither expert knew the detail of the work done by Veritec in the relevant years, both had to give evidence involving some speculation as to how it might have been achieved, and H2O’s own description of its process—each component being tested and retested as it was created and integrated—was consistent with a standard commercial Software Development Lifecycle (SDLC)4. As R&D tax commentary following H2O Exchange has emphasised—although the point was not decided by the Tribunal or the Federal Court in that case—repeated cycles of prototyping, functional testing under load, identifying design flaws, algorithm optimisation, and performance tuning are inherently directed toward improving or validating a known system4. Iterative engineering and trial-and-error refinement do not constitute a systematic progression of experimental work, nor do they satisfy the section 355-25(1)(b) requirement of generating new scientific or technical knowledge4.
On merits review of a finding under the IR&D Act, the applicant must satisfy the Tribunal that its registered activities meet the statutory eligibility criteria under Division 3553. In H2O Exchange, the Tribunal observed that the word “current” directs attention to the knowledge, information and experience available when the project was initiated, and that for the second income year it may refer to what was known at the beginning of that year (including knowledge derived by the applicant from the project itself); it also noted that whether an outcome already achieved by others is “known” depends on whether the methodology was published or could be replicated using readily available software tools3.
H2O did not produce evidence of the actual work performed in the relevant income years: it called no witness from Veritec, the technology partner that did the work, explaining that it could not locate those involved, and produced only limited documentation from Veritec, so that both experts had to speculate about how the work might have been done3. Technical tax commentary on the ruling (rather than the Tribunal’s reasons, which do not refer to any such protocol) recommends a disciplined, five-stage pre-experimentation proof protocol that H2O did not follow3. Under Stage A of this protocol, the claimant must explicitly document the specific technical features of the development outcome that differ uniquely from existing platforms in the marketplace, such as Waterfind3. Under Stage B, the claimant must catalogue precisely what is already known in computer science and industry practice regarding how to engineer those differentiating features3. Under Stage C, the claimant must isolate what cannot be known in advance and why, by identifying the competing technical factors, architectural constraints, and independent variables whose non-linear interactions prevent a competent professional from deducing the design outcome prior to testing3. Under Stage D, the claimant must translate those variable interactions into granular, falsifiable technical questions and hypothesised causal answers requiring structured experimentation3. Finally, under Stage E, the claimant must record this entire analytical baseline and experimental test protocol in formal engineering repositories before commencing the experimental coding and testing3. In the absence of such evidence, the Tribunal was not satisfied that the outcome of any part of the registered activities could not be known or determined in advance3.
Furthermore, H2O was bound by the broad terms in which it had registered its activities: the Tribunal proceeded on the basis that the relevant activities were those the respondent registered, and refused an adjournment to pursue a purported March 2016 variation request because, under clause 5.2 of the Industry Research and Development Decision-making Principles 2011, the respondent could not vary a registration while the decision was under external review (and could ordinarily do so only within 10 months after the end of the income year)3. When a taxpayer bundles an entire commercial platform build into broad registered activities, the Tribunal evaluates the activities as registered; in H2O Exchange, it noted that even if the only possibility favourable to H2O (uncertain AI experimentation for Activity 1.2) had been made out, it would have affected only a small part, if any, of the registered activities3.
The following table maps each statutory and evidentiary requirement examined in H2O Exchange Pty Ltd and Innovation and Science Australia [2019] AATA 4195 (and affirmed in [2021] FCA 11) against the specific deficiencies in H2O’s claim and the governing legal standard of proof applied by the Tribunal and the Federal Court3.
| Statutory Provision / Legal Requirement | Taxpayer’s Specific Failure in H2O Exchange | Tribunal & Federal Court Standard of Proof |
| Outcome Cannot Be Known or Determined in Advance
(ITAA 1997, s 355-25(1)(a)) |
H2O conflated commercial novelty—building an all-in-one online water exchange combining trading, CRM, binding contracts, real-time settlement, and multi-state conveyancing—with technological uncertainty3. H2O assumed that because the integrated platform was new to the market and complex to assemble, its outcome was uncertain3. | The applicant must prove on the balance of probabilities that the technical outcome could not be deduced in advance based on global current knowledge, information, or experience3. If the outcome is known to be achievable (parts of the activities being “simple and easily accomplished by a competent professional” or involving choosing between known software tools), or is known in advance not to be achievable because of external constraints such as government procedures, the test is not met, however complicated or labour-intensive the work3. |
| Current Knowledge, Information or Experience & The “Competent Professional” Test
(ITAA 1997, s 355-25(1)(a); Evidence Evaluation) |
H2O failed to establish the baseline of existing industry knowledge (ignoring existing platforms like Waterfind) and challenged the expertise of ISA’s expert witness on the basis that he had not studied software design, had never been hired as a software engineer and had never written a line of code (cross-examination also established that he held a Bachelor of Finance)3. H2O framed its Core Activity functionally rather than algorithmically3. | The applicant must satisfy the Tribunal that the outcome could not be known or determined in advance on the basis of the knowledge, information and experience current when the project was initiated (prior-art and literature reviews are a practical means of doing so)3. Where a claim is framed around functional system capabilities, a domain professional with relevant experience in collaborating on software, database and CRM design and in market regulatory mechanics may give expert evidence on what a competent professional could have known, without coding credentials (as the Federal Court held in dismissing the first ground of appeal)3. |
| Principles of Established Science
(ITAA 1997, s 355-25(1)(a)(i)) |
H2O focused on encoding statutory water trading rules, legal contract formation requirements, and state water registry conveyancing procedures across NSW, Victoria, and South Australia, rather than investigating computer science principles3. | The investigation must be grounded in principles of established science (e.g., computer science, formal logic, mathematics, queuing theory)3. On the evidence, systematising publicly available, clearly defined trading rules into a matrix (for example, as time-stamped database entries) was an outcome known to be achievable, and the Tribunal had no information on whether any AI-based automation of rule changes was based on principles of established science3. |
| Falsifiable Hypothesis
(ITAA 1997, s 355-25(1)(a)(ii)) |
H2O’s applications stated only high-level hypotheses (that its platform could trade water entitlements with settlement and title processes occurring in real time, and that real-time rule updates could ensure only eligible trades), describing a desired result without indicating how it would be attempted to be achieved3. Its overarching proposition was merely that a functional online water exchange could be built3. | The claimant must articulate a concrete, falsifiable hypothesis before experimentation begins, postulating a specific causal relationship between independent and dependent technical variables that cannot be resolved without empirical testing3. |
| Systematic Progression of Work (Experiment, Observation, Evaluation, Logical Conclusions)
(ITAA 1997, s 355-25(1)(a)(ii)) |
H2O produced no evidence of structured experimental protocols: it called no witness from Veritec (which actually did the work), produced limited documentation, and described its process as testing and retesting each component as it was created and integrated, consistent with standard Software Development Lifecycle (SDLC) processes4. | Repeated testing cycles, prototyping, system variability under load, and iterative algorithm optimisation do not constitute experimentation4. The taxpayer must prove a structured protocol with defined control/variant benchmarks, quantitative observation metrics, and formal evaluation proving or disproving the hypothesis3. |
| Purpose of Generating New Knowledge
(ITAA 1997, s 355-25(1)(b)) |
H2O’s activities were conducted for the commercial purpose of assembling an operational trading platform by configuring and optimising known software components and database architectures3. | The dominant purpose of the experimental activity must be the generation of new scientific or technical knowledge (not merely operational learning, bug resolution, or commercial product delivery)4. |
| Activity Segregation & Supporting Activity Nexus
(ITAA 1997, s 355-25(1), s 355-25(2), s 355-30) |
H2O registered two broad activities—Activity 1.1 (encompassing the web platform, CRM database, legal framework, payment settlement and conveyancing automation) and Activity 1.2 (integration of the trading-rules matrix)—and could not vary those registrations once the decision was under external review3. | Claimants must strictly unbundle core experimental modules from supporting activities (e.g., test environment build, data preparation, prior-art review) and exclude ineligible BAU tasks (e.g., UI design, standard CRM configuration, routine bug fixing)5. |
To demonstrate how Australian taxpayers and technical tax advisors can operationalise the legal lessons of H2O Exchange Pty Ltd and Innovation and Science Australia [2019] AATA 4195, the following two industry case studies examine realistic commercial projects that initially exhibit the exact structural, evidentiary, and technical vulnerabilities identified by the Tribunal. Each case study provides a comprehensive, step-by-step remedial architecture designed to satisfy Division 355 of the ITAA 1997 and withstand joint AusIndustry (IISA) and ATO compliance scrutiny.
ClearSettle Australia Pty Ltd (“ClearSettle”) is a Melbourne-based FinTech and PropTech enterprise developing an automated digital exchange and real-time settlement platform for secondary trading of tokenised Australian Carbon Credit Units (ACCUs), biodiversity certificates, and agricultural water and land encumbrances across state registries in New South Wales, Victoria, Queensland, and Western Australia.
In its initial project scoping, ClearSettle’s executive team intended to register a single omnibus Core R&D Activity mirroring H2O Exchange’s five-part functional claim: a web-based broker and institutional order portal, an integrated participant CRM and asset-holding ledger, automated smart-contract legal execution upon order matching, real-time settlement via the New Payments Platform (NPP), and automated multi-state registry conveyancing and encumbrance updating3. Had ClearSettle lodged this registration, IISA would inevitably have disallowed the claim on the exact grounds of H2O Exchange, producing a banking or carbon-market operations expert to testify that order-matching portals, CRM ledgers, NPP payment APIs, and state registry workflows are standard functional components whose implementation is “simple and easily accomplished by a competent professional”3.
To insulate the claim from a whole-of-project functional rejection, ClearSettle’s R&D tax advisory team restructures the project governance charter prior to the commencement of engineering work3. The advisory team excises all routine commercial software development and business-rule encoding from the Core R&D Activity boundary5. Specifically, building the web UI screens, configuring the customer CRM database, drafting the binding legal contract logic, and integrating standard NPP payment APIs are classified as ineligible business-as-usual (BAU) software engineering3.
The governance committee then isolates the sole genuine computer science unknown within the architecture: asynchronous distributed state-reconciliation and atomic consensus across non-transactional legacy state registries3. Specifically, state land and water registries operate on heterogeneous, legacy batch-processing architectures that lack native two-phase commit (2PC) atomic transaction support, whereas the NPP payment rail settles instantaneously in sub-second real time. When high-frequency algorithmic orders attempt simultaneous cross-jurisdictional atomic swaps (such as bundling a Victorian water allocation with a NSW biodiversity credit), network latency jitter and asynchronous registry lockouts create non-deterministic state divergence and partial settlement race conditions.
ClearSettle defines its Core R&D Activity strictly as the experimental design, construction, and testing of a novel, lightweight Byzantine Fault-Tolerant (BFT) speculative state-verification and optimistic rollback engine capable of guaranteeing mathematical atomicity across non-transactional external state registries without exceeding a strict settlement latency ceiling3. Auxiliary tasks—such as conducting the academic literature review on distributed consensus protocols, building the synthetic state-registry latency emulator (test rig), and preparing synthetic high-concurrency order datasets—are strictly segregated and registered as Supporting R&D Activities under section 355-30 of the ITAA 1997, undertaken for the dominant purpose of supporting the Core Activity5.
To neutralise the risk that an industry domain expert (akin to the Waterfind manager in H2O Exchange) could assert that the outcome was determinable in advance, ClearSettle mandates that its Principal Distributed Systems Architect complete and timestamp a formal Pre-Experimentation Baseline and Hypothesis Dossier following the five-stage proof protocol before any experimental code is committed3. Under Part A of the protocol, rather than claiming functional novelty in carbon or water trading, ClearSettle documents the exact low-level computer science differentiation: achieving deterministic atomic commit guarantees across external, read-only/delayed-write state machines that do not expose XA/2PC transaction coordinators, under sub- tail-latency bounds ()3. Under Part B, ClearSettle documents a comprehensive literature and prior-art review covering Paxos, Raft, Saga orchestration patterns, and standard Two-Phase Commit (2PC) protocols3. The dossier demonstrates that while competent software professionals know how to implement a standard Saga pattern with compensating transactions, standard Saga patterns are only eventually consistent and permit dirty reads during the asynchronous registry window, which violates financial settlement finality rules3.
Under Part C of the dossier, the architect defines the exact conflicting technical variables that prevent a competent computer scientist from determining the outcome in advance: asynchronous external registry response latency variance (), concurrent order contention density on overlapping registry title identifiers (), and speculative state-lock memory graph depth () coupled with cryptographic state-proof verification overhead ()3. The documentation explains why the non-linear trade-off between speculative lock-graph pruning and state-divergence rollback cascade rates cannot be deduced mathematically from existing queuing theory or distributed systems literature3. Under Part D, ClearSettle translates these variables into a specific technical question—whether a speculative directed acyclic graph (DAG) state-locking algorithm with predictive latency-weighted quorum validation can eliminate dirty-read state divergence across non-transactional external registries while maintaining transaction finality below when concurrent contention density exceeds —and formulates a falsifiable hypothesis3. Grounded in distributed computing principles and queueing theory (Little’s Law, ), ClearSettle hypothesises that replacing sequential Saga compensation with a speculative Markov-chain latency-predicted lock graph will reduce rollback cascade probability below while keeping settlement latency , because speculative pre-validation decouples local consensus finality from synchronous external I/O blocking3. Under Part E, the test protocol, control baseline (standard Saga orchestrator), variant architectures, independent variable step-changes, and quantitative falsification thresholds are formally signed off and committed to an immutable Git repository prior to Sprint 13.
ClearSettle engages an external Australian specialist software engineering consultancy (“VectorCode Pty Ltd”) to assist with aspects of the platform build. In H2O Exchange, H2O called no witness from its technology partner Veritec and produced only limited documentation from it, leaving the Tribunal unable to determine what work had actually been done3. To ensure full compliance with Division 355 (including sections 355-25, 355-30, and 355-210), ClearSettle restructures its contractual agreements through three interlocking mechanisms. First, ClearSettle bifurcates the engagement under its Master Services Agreement (MSA) into two legally distinct Statements of Work: SOW-1 (Commercial Platform Build) covers the user interface, CRM database, legal templates, and standard API wrappers on a fixed-price milestone basis, with all expenditure under SOW-1 excluded from the RDTI claim; conversely, SOW-2 (Experimental Consensus Engine R&D) is structured strictly on a time-and-materials (T&M) basis directed exclusively at executing the experimental protocol defined in the Pre-Experimentation Dossier5. Second, SOW-2 explicitly stipulates that ClearSettle’s Chief Technology Officer retains sole scientific and technical direction over the formulation of hypotheses, experimental parameters, and evaluation criteria, and that VectorCode provides no warranty of technical success for the experimental consensus algorithms, confirming that ClearSettle bears the full financial and technical risk of the experiment. Third, a binding contractual covenant in SOW-2 requires VectorCode engineers to log daily time entries mapped to granular Jira R&D task codes (CORE-EXP-01, SUPP-RIG-02, EXCL-BAU-03) and conditions invoice payment upon the delivery of raw experimental test logs, hypothesis evaluation summaries, and Git commit hashes linked to specific experimental runs5.
During execution, ClearSettle avoids the H2O Exchange pitfall of being unable to evidence what experimental work was actually performed, and avoids claiming iterative bug fixing and load testing, by embedding automated telemetry into its Continuous Integration / Continuous Deployment (CI/CD) experimental test harness4. Each experimental run executes a controlled comparison between the Control Group (standard Saga orchestrator) and the Experimental Variants under scripted synthetic latency injection5.
When Variant A fails during Experiment Run #4 (exhibiting a memory-graph deadlock when registry latency variance exceeds , causing latency to spike to ), the engineering team records a formal Hypothesis Falsification and Iteration Report5. The team analyses the root mathematical cause of the lock-graph cycle, formulates a revised Secondary Hypothesis incorporating a bounded epoch-timeout pruning algorithm (Variant B), and repeats the controlled experiment until logical conclusions are established3.
| Compliance Dimension | Pre-Remediation State (H2O Exchange Failure Mode) | Post-Remediation State (Division 355 Compliant Architecture) |
| Registered Activity Boundary | Whole-of-platform claim encompassing web portal, CRM, automated contract formation, NPP payments, and multi-state registry conveyancing3. | Core Activity restricted strictly to the speculative BFT state-reconciliation and atomic consensus engine; test rig and literature review mapped to Supporting Activities; UI/CRM/APIs excluded5. |
| Vulnerability to Non-Coding Domain Expert | High: A carbon/water market manager can testify that trading portals, CRMs, and registry conveyancing rules are standard and known3. | Neutralised: The technical question is framed in formal distributed computing and queueing theory ( latency vs. non-XA state divergence), requiring a distributed systems computer scientist to evaluate3. |
| Hypothesis Formulation | Commercial/functional goal: “Verify that an integrated platform can automate multi-state trading and settlement”3. | Falsifiable scientific proposition linking independent variables (, , ) to quantifiable dependent metrics (, rollback rate )3. |
| Contractual & Cost Substantiation | Single fixed-price software development contract with generic invoice descriptions (“Platform Development – Sprint 4”). | Bifurcated SOWs separating T&M experimental work from BAU build; risk borne by taxpayer; mandatory Jira tagging (CORE, SUPP, BAU) tied to invoice sign-off5. |
AusFreight Dynamics Pty Ltd (“AusFreight”) is a Brisbane-based supply-chain technology and industrial IoT company developing a real-time freight, cold-chain refrigeration monitoring, and port-slot scheduling system for agricultural export corridors across eastern Australia. The system ingests live telematics from heavy vehicles, dynamic port-gate congestion feeds, and perishable cargo thermal-degradation sensors to dynamically re-route trucks and re-allocate port unloading slots.
In its draft R&D preparation, AusFreight’s management proposed claiming the end-to-end development of the scheduling platform, arguing that the massive variability of human driver behaviour, unpredictable port delays, and complex cold-chain thermodynamic variations created “uncertainty of outcome,” and that its repeated cycles of field testing and algorithm tuning constituted experimental R&D4. This formulation walks directly into the legal ambush established in H2O Exchange and reinforced in subsequent decisions such as AVT, PKWK, and Lakes Oil4. Specifically, consistent with the Tribunal’s reasoning in H2O Exchange that a difficult task does not necessarily involve uncertainty of outcome and that uncertainty stemming from factors outside the applicant’s control is not uncertainty of outcome, unpredictability arising from human behavioural inputs, traffic, or environmental fluctuations constitutes operational variability within a known system rather than uncertainty regarding underlying scientific principles4. Furthermore, repeated testing cycles, prototype refinement, and algorithm parameter tuning to improve system performance under load represent routine engineering optimisation rather than a systematic progression of experimental work generating new knowledge4.
To rescue the claim from fatal disqualification, AusFreight’s technical tax counsel establishes an R&D Eligibility and Governance Charter that strips away all operational data collection, IoT sensor hardware integration, cloud dashboard construction, and heuristic parameter tuning4.
The technical review identifies a genuine, fundamental mathematical and computational bottleneck that cannot be resolved using current operations research or computer science knowledge: sub-second non-stationary stochastic combinatorial optimisation under hard edge-compute constraints3. Specifically, standard Mixed-Integer Linear Programming (MILP) and metaheuristic solvers (such as Simulated Annealing or Genetic Algorithms) used in vehicle routing and scheduling require stationary probability distributions and exhibit exponential time complexity as fleet size and constraint dimensions scale. When deployed on low-power in-vehicle edge processors during cellular blackouts in regional Australia, existing solvers take to recompute a valid schedule upon a stochastic port-slot cancellation, whereas operational safety requires deterministic convergence within using of RAM.
AusFreight scopes its Core R&D Activity exclusively around the experimental investigation of a novel Hybrid Quantum-Inspired Tensor-Network Decomposition and Physics-Informed Graph Neural Network (PI-GNN) Solver designed to approximate NP-hard spatio-temporal constraint satisfaction in polynomial time without violating hard thermodynamic spoilage constraints3. Activities such as cleaning historical telematics datasets, building the edge-hardware simulation testbed, and coding data-ingestion pipelines are strictly quarantined as Supporting R&D Activities under section 355-30, or excluded entirely where they serve general commercial operations5.
To satisfy the burden of proof identified in H2O Exchange regarding “current knowledge, information or experience”3, AusFreight commissions a formal State-of-the-Art Mathematical and Algorithmic Baseline Report authored jointly by its Lead Operations Research Scientist (holding a PhD in Computational Mathematics) and an independent academic advisor prior to commencing work3. Following Parts A and B of the five-part proof framework, the report benchmarks commercial solvers (Gurobi, CPLEX, Google OR-Tools) and academic literature on Capacitated Vehicle Routing Problems with Time Windows (CVRPTW)3. It demonstrates that any competent operations research professional can configure OR-Tools to optimise static or quasi-static schedules, which is therefore excluded from the claim3.
Addressing Part C of the proof protocol, the report proves why a competent professional cannot know in advance whether a compressed Tensor-Network / PI-GNN architecture can maintain solution feasibility (zero hard-constraint violations) when pruning tensor bond dimensions () to fit within a edge memory footprint3. Specifically, reducing bond dimension to accelerate inference time () induces mathematical truncation error () that interacts non-linearly with the non-convex Arrhenius thermal spoilage differential equation governing cargo perishability3. Addressing Parts D and E, the research team documents the exact technical question—how tensor bond-dimension truncation () coupled with Lagrangian penalty weighting () in a Physics-Informed Graph Neural Network affects hard-constraint feasibility violation rates and convergence latency across non-stationary stochastic routing graphs of nodes on ARM Cortex-A78 edge hardware—and records a falsifiable hypothesis prior to experimentation3. Grounded in computational complexity theory and spectral graph theory, AusFreight hypothesises that embedding hard thermodynamic boundary conditions directly into the message-passing adjacency matrix via spectral projection, rather than via soft loss-function penalties, will enable tensor bond compression to , achieving deterministic convergence in within RAM while maintaining hard-constraint feasibility (zero thermal spoilage violations), because spectral projection restricts the neural search space strictly to the feasible polytope manifold3.
AusFreight collaborates with a Cooperative Research Centre (CRC) / Registered Research Service Provider (RSP) for mathematical modelling validation and contracts an external embedded-systems firm (“Edgeware Engineering Pty Ltd”) to compile test builds. To prevent contractual ambiguities from undermining eligibility, AusFreight implements three contractual safeguards. First, the agreement with the RSP is structured to explicitly reference AusFreight’s Core R&D Activity hypothesis and experimental protocol, ensuring that expenditure qualifies under the statutory RSP provisions of Division 355 while AusFreight retains beneficial ownership of all foreground IP and experimental results. Second, the contract with Edgeware Engineering separates routine firmware maintenance and telematics driver integration (invoiced under a Commercial Operations Account code) from the construction of the experimental tensor-quantisation kernels built specifically to test the Core Activity hypothesis (invoiced under an R&D Experimental Account code)5. Third, both contracts mandate bi-weekly technical progress reports structured explicitly around the statutory limbs of section 355-25(1)(a)(ii), requiring formal documentation of the hypothesis tested, experimental configuration, quantitative observations, technical evaluation, and logical conclusions3.
To address the concern—drawn from R&D tax commentary rather than from the Federal Court’s reasons in H2O Exchange, which were confined to three grounds concerning expert evidence and natural justice—that “repeated testing cycles alone” and “algorithm optimisation and performance tuning” do not constitute experimental activity4, AusFreight implements a rigorous, bench-scale laboratory simulation protocol before any field trials occur5.
Instead of deploying code into live trucks and observing operational variability, AusFreight executes controlled computational experiments across fixed, reproducible synthetic benchmark datasets representing 10,000 non-stationary disruption scenarios4. In each experiment, only a single independent variable (such as the tensor bond dimension or the spectral projection operator) is varied against a static baseline solver (Gurobi MILP running on unconstrained cloud compute)3.
All experimental runs automatically log objective mathematical metrics: inference latency (), peak memory allocation (), optimality gap percentage (), and hard constraint violation count ()5. When early experimental evaluations prove that spectral projection alone introduces numerical instability at due to eigenvalue degeneration in sparse graphs (falsifying the initial hypothesis), the research team documents the mathematical discovery—which constitutes genuine new computer science knowledge under section 355-25(1)(b)—and formulates a new hypothesis incorporating Tikhonov regularisation within the spectral layer4.
| Compliance Dimension | Pre-Remediation State (H2O Exchange Failure Mode) | Post-Remediation State (Division 355 Compliant Architecture) |
| Nature of Uncertainty Claimed | Operational variability: Unpredictable human driver behaviour, port congestion delays, and real-world cold-chain fluctuations4. | Fundamental algorithmic unknown: Whether spectral-projected tensor-network decomposition can converge NP-hard non-stationary graphs in on edge RAM with zero constraint violations3. |
| Experimental Progression vs. Iterative Tuning | Unstructured field testing, prototype modification, and iterative parameter tuning to improve system performance under load4. | Controlled laboratory A/B simulation protocol isolating independent mathematical variables (, ) against a fixed MILP baseline with automated falsification telemetry4. |
| Generation of New Knowledge | Operational learning and commercial system refinement; no new underlying computer science principles discovered4. | Formal mathematical characterisation of eigenvalue degeneration in spectral-projected PI-GNNs under tensor truncation—generating transferable computer science knowledge4. |
| Supporting vs. Ineligible Task Mapping | All data collection, IoT sensor integration, UI dashboards, and routing tuning lumped into Core R&D4. | Data cleaning and edge testbed construction mapped strictly to Supporting R&D (s 355-30); commercial IoT integration, UI, and post-release tuning excluded as BAU5. |
The decisions in H2O Exchange Pty Ltd and Innovation and Science Australia [2019] AATA 4195 and H2O Exchange Pty Ltd v Innovation and Science Australia [2021] FCA 11 represent a permanent structural shift in how Australian regulators and courts adjudicate R&D Tax Incentive claims, particularly in software engineering, algorithmic systems, and platform integration3. The jurisprudence establishes that commercial ambition, functional novelty, the difficulty of a task, and uncertainty arising from factors outside the claimant’s control (such as multi-jurisdictional government registry procedures) do not by themselves establish that an outcome cannot be known or determined in advance under section 355-25 of the ITAA 19973.
Furthermore, as Australian policy and administrative scrutiny continue to tighten around the boundary between core experimental work and supporting or routine activities4, taxpayers can no longer rely on retrospective narrative reconstruction at tax-return preparation time3. To survive merits review and judicial scrutiny in Australia, claimants must unbundle commercial builds from core scientific inquiries at project inception, articulate falsifiable hypotheses grounded in computer science or physical science principles, benchmark their technical unknowns against global baseline knowledge using the five-part proof protocol, and capture contemporaneous empirical evidence proving that their work generated genuine new knowledge rather than merely optimising a known system3.