Articles are ingested from a curated set of sources. A structured classifier assigns an event type, severity (1–5), and a confidence score (0.0–1.0). The intent is consistency and auditability rather than prediction.
Confidence reflects the system’s certainty in the structured labels, based on clarity and corroboration in the source text. Low confidence does not imply the event is unimportant; it indicates ambiguity or limited detail.
Event types are mapped into stable buckets (e.g., kinetic, diplomacy, sanctions, procurement, cyber). Taxonomy can expand over time without changing historical records.
Risk indices are aggregated from structured events over rolling windows. A baseline approach is:
Market analysis is descriptive and window-based:
Weapons systems referenced in events can be linked to manufacturers and (where applicable) representative tickers. The mapping is curated and scope-limited; it does not imply procurement, revenue share, or exposure magnitude.
Tender and award notices are taken from Canadian federal open data under the Open Government Licence. Two judgements are applied, and they are kept separate:
A sector is anchored on a supply-classification code only where the code class is the sector. Anchoring on a general electronics or hardware class puts a rack-mounted console into radar and autonomy, which is why the rule is narrow. The classification ruleset is versioned, and stored notices are replayed against it when it changes, so a figure can always be traced to the rules that produced it.
Two limits are stated on the page itself rather than buried here. Tender notices carry no monetary field of any kind — value exists only on award notices, so any total covers awards alone. And closing dates are published without a time zone offset.
Supplier names as government publishes them are resolved into organisations. The rule is structural, not a similarity score: two names may merge only when every token they do not share is generic. Two companies sharing a family name but differing in what they do — an oil company and a shipbuilder — cannot merge at any threshold, in any tuning, while four spellings of the same subsidiary collapse into one.
How a name was matched determines how the link is drawn. An exact, alias or normalised match is a government record and is drawn filled and solid. A partial token overlap is an inference and is drawn hollow and dashed. A published supplier name we have never seen becomes a provisional organisation rather than a discard: refusing to record it would throw away a government fact. What is withheld is the judgement, not the name.
No machine-readable organisation chart of the Government of Canada exists, so the structure is curated, with a fetched and confirmed source URL for every node. Bodies whose published location could not be verified were dropped rather than guessed at.
This is deliberately not a searchable directory of public servants. A person exists only as the holder of a public office and is reachable only by selecting that office; an office shows a name only where it has been read off a government page and dated. Most offices therefore show an empty seat, and the ratio of filled to total is published rather than hidden.
The committee that held a meeting is a record. What the meeting was about is keyword-matched, and therefore always presented as an inference. Sources that are declared but not yet connected report as skipped with a reason; a source that looks healthy while returning nothing is the worst outcome, so not-connected and broken are shown as different states.
Across every visualization, certainty is carried by fill and stroke, never by colour. Something published by a government source is filled and solid; something inferred or unverified is hollow and dashed. That distinction survives greyscale printing and colour blindness. A relationship nobody publishes is drawn broken rather than hidden, because a chain that quietly omitted the missing hop would imply the data connects when it does not.