What counts as SR&ED?

The same project can be SR&ED or not, depending on what was unknown and how the work was done. Each example below is written twice. Pick the version you think qualifies, then read which test decides it.

Every company here is invented to show one test. None is a Client. The four tests are explained onDo I qualify.

Software. Illustrative, not a Client.

The query that stopped scaling

A Toronto SaaS company. Its customer dashboards slowed to 40 seconds once the largest account passed 50 million events.

Which version is SR&ED?

Version ANot SR&ED

The team added indexes, moved to a bigger database instance and cached the common queries, following the vendor documentation. It took six hard weeks, but at every step a competent engineer knew what would work.

Version BSR&ED

Every customer defines its own event fields, and they change daily. The published fixes (precomputed views, a columnar store) assume fields that stay put, and both broke on the first schema change. Nobody knew whether sub-second queries were possible under those conditions, so the team tested a rollup keyed on schema version, measured it, and changed course twice.

Decided by: Technological uncertainty. Hard, slow and expensive is not the test. The question is whether a competent professional could have known the answer from published knowledge. For indexes and caching, they could.

Software, machine learning. Illustrative, not a Client.

The model that would not generalize

A Waterloo startup flags defects in photos taken on customers' phones. Accuracy collapsed on phone models it had never seen.

Which version is SR&ED?

Version ASR&ED

The team wrote a hypothesis before each change, held out whole phone models as test sets, logged every run with its result, and dropped each idea the data did not support. Three hypotheses failed before one held, and each failure is on record with its date.

Version BNot SR&ED

The team swapped models and settings until the numbers looked good again. Nobody wrote down what they expected from each change or what happened, so when the model worked, nobody could say why.

Decided by: Systematic investigation. The problem is the same in both. What differs is the method: hypothesis, test, measurement, adjustment, with a record made at the time. Trying things until something works, with no record of why, is weak.

Software, integration. Illustrative, not a Client.

The instrument with no manual

An Ottawa health-tech company needs its software to talk to laboratory equipment.

Which version is SR&ED?

Version ANot SR&ED

The equipment has a documented interface. Connecting to it took many edge cases and a long test cycle, but the documentation said how, and following it worked.

Version BSR&ED

The older instruments use an undocumented protocol that drops messages under load, and the maker no longer supports it. At the start nobody knew whether a reliable connection was possible at all. The team built a test harness, tested theories about timing and buffering one at a time, and measured the loss rate after each.

Decided by: Technological uncertainty. Integration work is usually routine: the path is documented and the work is effort. It becomes SR&ED when the path is not known and could not be looked up.

Software, data. Illustrative, not a Client.

The pipeline built abroad

A Montreal fintech needs every transaction reconciled within 200 milliseconds, with events arriving out of order. The known designs trade speed for correctness, and it needs both.

Which version is SR&ED?

Version ASR&ED

The same experiments, run by the company's own engineers in Montreal, who designed each test, ran it and analyzed the results.

Version BNot SR&ED

The design is genuinely uncertain and the experiments are well recorded, but they were run by a contracted team outside Canada. The Montreal staff wrote the requirements and reviewed the results.

Decided by: Work performed in Canada. The experimental work has to be done in Canada. Work done abroad does not count, even when the problem is real and the records are good.

Hardware. Illustrative, not a Client.

The sensor in the cold

A Calgary company makes wireless sensors for remote well sites. They must run three years on one battery at minus 40.

Which version is SR&ED?

Version ANot SR&ED

The engineers picked a cold-rated battery from its datasheet, designed a sealed housing and ran the standard qualification tests. The parts behaved as their specifications said.

Version BSR&ED

The radio wakes in short bursts, and no published data covered that pattern of load at minus 40. The first design died in four months in the test chamber. The team built a rig to measure battery behaviour under pulsed loads in the cold, tested three theories about the cause, and redesigned the power stage around what it learned.

Decided by: Technological uncertainty. Choosing parts from specifications and confirming they meet them is standard engineering. The uncertainty starts where published data stops.

Manufacturing. Illustrative, not a Client.

The weld that cracked

An Ontario fabrication shop takes on a part that joins two different alloys.

Which version is SR&ED?

Version ASR&ED

The recommended settings cracked the joint every time, and nothing published covered this combination at this thickness. The shop varied heat input and cooling rate in planned trials, cut open and measured each sample, and kept the results.

Version BNot SR&ED

The shop programmed its robot welder with the settings the wire supplier recommends, ran trial pieces, made small adjustments and passed inspection. That is setting up production.

Decided by: Technological uncertainty. Trial runs, adjustments and quality control on the way to production are excluded. Work to find out whether a joint can be made at all, when published knowledge cannot say, is not.

Food. Illustrative, not a Client.

The sauce without the preservative

A British Columbia food company reworks its best-selling sauce.

Which version is SR&ED?

Version ANot SR&ED

The goal was a new flavour for a new market. The team made batches, ran taste panels and settled on a recipe customers preferred.

Version BSR&ED

The goal was to remove a preservative and keep a 12-month shelf life at room temperature. Nobody knew whether that combination of acidity, water content and heat treatment could hold. The team ran challenge tests on each formulation and recorded which failed and when.

Decided by: Technological uncertainty. Taste and style changes are not SR&ED, however much work they take. Whether a product can be made to hold at all, when that is unknown, can be.

Agriculture. Illustrative, not a Client.

The greenhouse lights

A greenhouse near Leamington tests whether a new lighting schedule can raise yield without raising power use. Nobody had published results for its crop and climate.

Which version is SR&ED?

Version ASR&ED

Two salaried growers ran the trial for a full season, with sensors, control rows and weekly measurements, and the company paid them for that time.

Version BNot SR&ED

The owner ran the trial alone over a few weekends, unpaid, with $800 of fixtures. The work was genuinely experimental. There was almost nothing to claim against.

Decided by: A real expenditure base. The credit is a share of what the work cost: salary, materials, Canadian contractors. Good experimental work that cost almost nothing produces almost no credit.

One real example: my own Claim

My company spent about 14 months building an AWS cost tool. Most of that work was not SR&ED: the integrations, the interface and the infrastructure were known work, however long they took. What qualified was a narrow slice: working out which account owns a cost when tagging is incomplete and the answer has to be inferred. At the start I did not know whether any approach would work.

Sorting a year of work into those two piles is most of the job. I built the platform, wrote the Narrative and filed that Claim. The CRA assessed it as filed without selecting it for review.

Recognize one of these?

Tell me what you built that you were not sure would work. You get a written Verdict, free, and it may say no.