Is this SR&ED?
Three projects, each written twice. One version is a Claim and one is not. Pick before you read the answer.
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, 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.