Reasonable notice, predicted from 3,200+ Canadian court decisions
BardalGPT models the Bardal factors the way Canadian courts actually have — across 3,200+ decisions, not a rule of thumb or law firm severance calculator.
Free · No account required · Runs entirely in your browser
What BardalGPT does
Reasonable notice has no formula. Courts weigh the Bardal factors — age, length of service, character of employment, and availability of similar employment — and arrive at a figure by judgment, not arithmetic. That makes the outcome hard to anticipate and easy to dispute.
BardalGPT estimates where a court would likely land. Through advanced mathematical modelling, BardalGPT assessed the relationship between the Bardal factors and the wrongful dismissal awards in 3,200+ reported decisions, and it reproduces two behaviours real courts display that a formula misses: notice rises steeply for the first years of service and then flattens, and awards cluster on round figures rather than spreading evenly.
How it differs from the Bardal Index
The Bardal Index reports. It finds the most similar decided cases and shows you the range they fell in. It presents the most objective position possible.
BardalGPT predicts. It gives a single most-likely figure and a calibrated interval around it. That is more useful for setting a negotiating position, and it is necessarily an estimate rather than a fact. Use them together: the prediction to orient, the Index to verify.
Limits. The model knows age, service, occupation, jurisdiction, and year. It cannot see the availability of comparable employment, the manner of dismissal, mitigation efforts, or the terms of any contract — all of which move real awards. Typical error is about 2.4 months. It is not legal advice and creates no lawyer-client relationship.
Why the number can be trusted
Two things separate BardalGPT from a severance calculator. It was built from data, and then it was graded on its accuracy. Most tools do neither, and none of them publish the second.
First: trained on 3,252 real decisions
The model learned the relationship between the Bardal factors and the awards from 3,252 reported wrongful dismissal decisions spanning 1955 to 2026, across every Canadian jurisdiction except Quebec. It takes its view of what notice is worth from what judges actually did — not from anyone’s opinion about what severance ought to be, and not from a formula someone chose in advance.
Second: graded against real case law
Training a model is the easy part. Proving it works is the part almost nobody does. So the cases were split by date: the model was shown only decisions from before a cut-off, then asked to predict awards made after it — cases it had no knowledge of. That was repeated at five separate cut-offs across twenty-five years. Every figure below comes from that test.
| Approach | Average error | Within 2 months |
|---|---|---|
| An untrained model | 5.6 mo | 19% |
| The “one month per year” rule of thumb | 3.2 mo | 49% |
| BardalGPT | 2.4 mo | 65% |
The bar shows the calibrated interval. Four out of five comparable employees receive an award inside the outer band.
The model’s full response curve, holding your other factors fixed. The steep early rise and the flattening after roughly twenty years are both learned from the decisions, not imposed.
The decided cases closest to your inputs, nearest first. Every prediction should be checkable against real authority.
Gradient-boosted decision trees (median-optimised), trained on 3,252 reported decisions and snapped to the grid of figures courts actually award. Predicts the Bardal notice period before any bad-faith or Wallace extension. Does not account for availability of comparable employment, manner of dismissal, mitigation, or contractual terms. An estimate, not a prediction of your case, and not legal advice.