Could Your Claim Survive a Question?

FY26 closed on Tuesday. If your company claims the R&D Tax Incentive, there’s a date sitting quietly in next year’s calendar: 30 April 2027, the deadline to register your FY26 activities with AusIndustry. Ten months away. Right now it feels like someone else’s problem, and for the next eight months or so, it will keep feeling that way.

Then it’s March, and the scramble starts.

I’ve been in the scramble. It has a particular texture. The consultant sends a template with columns for hypothesis, experiment, observation and conclusion. Finance forwards it to engineering with a note that says “shouldn’t take long.” An engineering lead opens the git history and starts scrolling back through a year of commits, trying to remember which of them were experiments and which were just work. Someone asks a senior engineer what she was investigating last August. She was investigating whether the new retrieval approach would hold up at production scale. It failed after three attempts, and she remembers that much. What she can’t tell you is what she believed before the first attempt, or which of the three failures actually taught her something. Nobody can. That information lived in her head for about a fortnight and then it was gone, overwritten by the version of events where the team ended up.

So the team writes down the version of events where the team ended up. The hypothesis gets drafted eleven months after the experiment it supposedly preceded. The uncertainty gets described by people who now know the answer. Hours get apportioned by feel, in round numbers, because the timesheets never existed. Everyone involved is being honest. Nobody is lying. And the document that comes out the other end is still not evidence of anything except what the team remembers in March.

Reconstruction isn’t evidence. It’s recollection with formatting.

That distinction sounds pedantic until someone from the ATO or AusIndustry asks a question. The claim you registered says you conducted experiments to resolve a technical unknown. A reviewer’s job is to test that, and the test is simple: show me. Show me what you didn’t know, and when you didn’t know it. Show me the experiment as it was designed rather than as it’s remembered, and the result that surprised you. A claim built continuously can answer with artefacts that carry their own dates. A claim built in April can only answer with a narrative, and a narrative written in hindsight has a distinctive shape. Hindsight deletes the uncertainty and straightens the path. It produces documents where every experiment succeeds on schedule and every hypothesis turns out correct, which is exactly what real research never looks like, and reviewers read enough of these to know the difference.

The ATO is explicit about this. Its record-keeping guidance says records that are backdated or contain non-specific details of the work undertaken are not appropriate. Tribunal decisions on RDTI matters have turned, over and over, on whether evidence existed at all, not on whether the R&D was real. You can have done genuine, eligible R&D and still lose the claim because the only proof is a story assembled a year later. The work was real. The evidence wasn’t.

Here’s the uncomfortable part for anyone reading this in July: your FY26 claim is already whatever you captured. If your team wrote its unknowns and its results down as they happened, your April is an assembly job. If they didn’t, no process you adopt now will change what exists. The best you can do for FY26 is start the reconstruction early, while memories are eight days old instead of eight months old, and be honest with your consultant about how much of it rests on memory. That conversation is worth having with them directly, because the substantiation detail is theirs to own, not mine.

FY27 is different. FY27 is four days old.

What contemporaneous evidence actually requires is smaller than most founders assume, and this is where the objection usually arrives: we’re a startup, we can’t run a parallel admin process. Agreed. Don’t. The evidence the RDTI wants is mostly evidence good engineering teams produce anyway, just written down at the time instead of reverse-engineered later. A design doc that states what you don’t yet know before you build. A short note when an experiment ends: what we tried, what happened, what we concluded. Branches and tickets named for the investigation they belong to, so the git history sorts itself. Some live record of who spent time on the experimental work, kept weekly, in whatever tool the team already uses. None of this is extra work in any meaningful sense. It’s the same thinking the team is already doing, recorded while it’s still true.

The habit that matters most is writing down the hypothesis before the work starts, and the team gets more out of it than the tax office does. A stated unknown is the one artefact that cannot be reconstructed later, because the moment you know the answer, you lose access to what not knowing felt like. Everything else can be approximated after the fact, badly. That one can’t be approximated at all.

I think about this more than I probably should, partly because I’ve spent the past while building EngLedger in exactly this space, which means I’ve heard a dozen versions of the March story from a dozen different teams, and they differ only in the details. A year of genuine research, a fortnight of panicked remembering, and a registration that describes the memory.

The fix costs almost nothing and it’s only available at one time of year, which is now. Not because July has any regulatory significance, but because a capture habit adopted in July covers the whole year, and one adopted in February covers nothing that matters. Every week you wait is a week of FY27 that will eventually have to be remembered instead of read.

Your FY26 claim is finished, whether or not you’ve written it yet. It’s sitting in whatever your team wrote down, and April will reveal how much that was.

FY27 started on Wednesday. Start writing it down.