Why Regen Power has committed $600,000 to answering that question properly
Ask a Western Australian homeowner what they want to know before putting a battery on the wall, and the question usually comes out in one line: what does the rebate actually do for me?
It sounds like a question with an answer. It isn’t – not on its own.
What a rebate is worth to a particular household depends on the DEBS export rates, that household’s Synergy tariff, the state scheme running that quarter, the federal program and its current step-down, and the connection rules that apply to their system size. Change any one of those and the answer changes. On 1 May 2026, several of them changed at once.
That is the first problem. There is a second one, and it is more interesting.
The same question, asked four different ways
Two homeowners can ask what looks like an identical question and need completely different answers – because they are not asking for the same thing.
- “Am I eligible?” asks for a decision. It has a yes or a no, and the answer lives in the eligibility conditions of a specific scheme.
- “Is it worth it?” asks for a comparison. It requires tariff data, export rates, consumption patterns and a calculation – and the answer is a number, not a rule.
- “How do I apply?” asks for a process. The answer is a sequence of steps, forms and approvals.
- “Should I wait?” asks for a judgement under uncertainty. It requires knowing what is scheduled to change and when.
Treat all four as “rebate questions” and route them down the same path, and you will answer three of them incorrectly – while appearing to have answered all four. That is the failure mode that matters, because it is invisible to the person receiving the answer.
A registered research program
Regen Power has commenced a $600,000, twelve-month registered research and development program to build an end-user technical query system for the South West Interconnected System.
The system is designed to take a homeowner’s question in their own words, break it into its separate regulatory components, resolve each one against a controlled library of the actual rules, and return an answer where every material claim carries a citation to the governing clause and the date that clause was last checked.
The program comprises three core research activities and four supporting activities, registered under Division 355 of the Income Tax Assessment Act 1997 for FY2026-27.
This is a separate program from our Australian Research Council Linkage Project with Curtin University. That three-year program, announced earlier this year, addresses battery sizing, smart-charging optimisation and end-of-life pathways. This one addresses a different question entirely: not how to size and run a battery, but how to give a household a regulatory answer that is correct, current and verifiable.
How the system works
Decomposition
We start from the Multiple Perspective Framework as the structure for representing questions that cover several topics at once, and extend it in two directions.
The first is a set of intent tags derived from speech act theory – classifying what the person is actually asking the system to do, rather than only what the question is about. That is what separates the eligibility question from the worth-it question when both mention the same rebate.
The second is a technology tag for the equipment involved – solar, battery, inverter, meter, EV charger – because the equipment frequently determines which regulatory instrument applies in the first place.
A single enquiry therefore breaks into several tagged components instead of being forced into one dominant category.
Retrieval
Each component is resolved separately against a clause-level, version-controlled library of the governing rules. Clause-level matters: it is the difference between citing a scheme and citing the specific provision that decides the question. Version-controlled matters because these rules change, and an answer that was right in April may be wrong in May.
Response
Every material assertion in the output is bound to the clause it came from, together with the date that clause was last verified. Where the system is not confident, it says so rather than producing a fluent answer that happens to be wrong.
How we will know whether it worked
The program is structured as an experiment, not a build. It is built on a de-identified, expert-annotated corpus of more than 3,000 genuine Western Australian homeowner enquiries drawn from our own service records.
Evaluation is blind. Trained domain assessors grade system output against locked test sets – frozen and hash-verified before experimentation begins, so the targets cannot move to meet the results. Performance is measured against thresholds set in advance:
- Classification accuracy – whether the system correctly decomposes multi-part questions
- Hallucination rate – the share of responses containing any claim the sources do not support
- Citation precision – whether cited clauses genuinely support the claims attached to them
- Citation recall – whether every material claim carries a citation at all
- Update latency – how quickly the system reflects a rule change after it happens
- Readability – an automated plain-language check on every confirmed answer
The combination we are testing – structural decomposition, intent classification and technology binding, applied to a body of rules that changes constantly – may not hold. Some of these thresholds may not be met. That is what makes it research rather than product development, and we would rather publish the criteria in advance than describe the result afterwards.
Why we are doing this
Most solar and battery information in this country is written by a marketing department. It is usually accurate on the day it is published and quietly wrong some months later, and the reader has no way of telling which state it is in.
We have published rebate guides ourselves. Every one of them carries the same implicit disclaimer that everyone else’s does: check the current rates, because these change. That disclaimer is an admission that the format does not work.
A household making a decision worth thousands of dollars deserves better than a confident guess. It deserves an answer that shows where it came from and when it was last checked – and, where the answer genuinely isn’t certain, says so.
We would rather be the company that shows its working.
