Method

How the evidence is gathered, weighed and published, and what the labels on this site mean.

Actions over statements

Every event on the record carries an evidence tier:

• Action something was done: a law switched on, a tariff imposed, money committed, a pipe laid. Weighted highest.
• Document an official document: a bill, a filing, an agreement text.
• Statement something was said. Weighted lowest, because words cost less than actions.

The same rule separates stages of a law. A bill that passed but was never switched on is not the same as a law in force, and a law in force is not the same as a law that has been used. The laws page shows each stage with its date.

Caution labels

Some events are contested but still worth recording. Rather than leave them out, we publish them with a caution label and a short note saying why. There are five kinds:

• Unproven allegation a claim someone made that nobody has confirmed.
• Self-reported only one side’s own account of itself.
• Anonymous sources reported by an outlet citing unnamed sources.
• Second-hand account someone recounting what others said or did.
• Single source only one outlet has reported it.

Caution labels are separate from the action, document and statement tiers above. The tier says what kind of thing an event is; the caution label says how sure we are of it. A label is set by the author, and adding, changing or removing one is recorded in the corrections log.

Outcome first

The journal’s claims are about outcomes: what is being built, where it goes and who pays. Those claims rest on actions and hold whether or not anyone intended them.

Intent as a thought experiment

The question of intent is handled separately, as a deliberately contrarian test: suppose Washington foresaw and wanted this. What would we expect to see, and do we see it? It is not the author’s belief. The rules:

• The foresight reading is never shown alone; the rival readings (accident, plain self-interest, annexation or fragmentation) are always alongside.
• No reading carries odds. Intent can’t be observed, so no dated event could ever settle such a number; publishing one would claim a precision we don’t have. What we publish instead is what each reading predicts and what the searches found.
• The tests come before the verdicts: predictions are registered in advance, with what would count against each reading.
• The wording is careful: “consistent with”, “fits equally”, “not observed in public sources”. Never “proof”.

What “non-discriminating” means

An event is non-discriminating when it fits two competing readings equally well, so it cannot tell them apart. A tariff that suits plain US self-interest and also suits the foresight reading is evidence of neither over the other, however striking it looks.

Each event can be judged against each reading (fits, against or neutral; weak, moderate or strong). When the self-interest reading (I0b) and the foresight reading (I2) get the same judgment, the event is marked non-discriminating on its page. This replaces arguing that something “tells us nothing about intent” with showing it.

So far, 135 of 207 events judged both ways are non-discriminating.

About the odds

Some of these numbers come from simple models. Where comparable past cases can be counted, we write the model’s rules down before running it, then publish its result with its range and its weakest points. Two outcome questions (O2 and O3) and both tail risks (US military action against Canada, M1, and whether a senior US official refuses to rule out force against Canada, M2) are set this way. The others are our own judgment calls. The reason given with each number says which kind it is.

Either way, we are an independent, generalist journal, not specialists in energy, trade or defence, so read these numbers as an informed view, not expert forecasts. They are estimates, not measurements.

They are published for two reasons. The first is transparency: you can see what the journal tracks and which way we lean, and judge the articles with that bias in view. The second is that they are a working tool: putting a dated number on a view forces it to be tested against what actually happens.

We put odds only on a question that a dated, public event will settle, so every forecast can be scored. Many will miss. That is expected, and every miss stays on the record. Some pieces also give a model's estimate under a stated condition, such as how Canada fares if today's tariffs fade the way past ones did, or under a given deal. Those are scenarios, not forecasts: they say what follows if the condition holds, not how likely the condition is, and they are labelled that way. Where there is real uncertainty a range is given, and each change is recorded with its reason on the reading’s page and in the corrections log.

The odds are not the product. They decide where to dig. What the journal publishes is what that digging turns up, and every claim links to its sources. The evidence carries the weight, not our credentials.

How we use prediction markets

We watch prediction markets and forecasting sites (such as Polymarket, Kalshi and Metaculus) on the questions this journal tracks. We don’t bet on these markets, and we came to them skeptical, until this project gave us a use for them.

We don’t use them to inform ourselves. Most markets on Canadian questions are thin, and a price can move on a handful of bets, so we treat them as weak signals. We use them to compare against our own odds. When the crowd and the journal disagree by a wide margin, we dig: what are we not tracking that could explain the gap? Sometimes the answer is a development we had missed, sometimes a stale number of ours, sometimes just noise in a thin market.

It’s free constructive criticism. A market price never changes our odds by itself; only evidence the digging turns up does, and the change log cites that evidence, not the price.

What our timeline covers

The timeline records what we found in news media, official documents and other published sources, as far as our searches reached. It is not a record of everything ever said on a subject. We don’t monitor every influencer, YouTube channel, podcast, TikTok or social media account. So when we say when people first heard of something, that means the earliest coverage we found in those sources. Someone may well have said it, or predicted it, earlier somewhere we weren’t looking. If you know of an earlier, verifiable source, tell us at contact@carnet49.ca: we’ll check it and correct the record.

Predictions

Each prediction is registered with a date, a window and a statement of what counts, before the result is known. Observations are logged as events happen; only those marked as counting feed the result. The scorecard shows every test and every result, including the ones that went against the journal’s own readings.

Sourcing rules

• Every source records how it can be reached: primary document, free, subscription, archived copy, or seen only as an excerpt.
• Paywalled sources are never quoted. They are cited as corroboration, with a link to the publisher; the fact itself rests on a free or official source. When a paywalled outlet is the only source, the fact is stated in our own words and credited to it.
• Text seen only in search results is never quoted.
• A quote that can’t be shown (paywalled, or not cleared) is marked as such rather than silently dropped.
• Only items cleared for publication reach the public record. By default an event is published only if it is an action or document with at least one free government, legislature or court source; others are published once checked.

The research desk

An automated research desk does much of the gathering: it finds and files documents, logs dated events, pulls published data series and proposes how evidence bears on each claim. It also drafts the articles. The author reviews what it proposes, accepts, fixes or rejects it, and argues each draft through before approving it (how this site uses AI). The desk’s work is logged; review notes are not published.