How the machine runs

Add up each adapter's own duration and the run did 33.2 s of collection work inside a 15.2 s collection window. Several checks ran at once, fitting 2.19x as much timed work into that window. The slowest measured source check was Measured current stock discussion attention at 14.8 s.

Which is also why the 15.2 s is the small number. 43% of the pass ran after every source was shut. The receipt times that window; it does not itemise it. What runs in it is the merge: 11 adapter outputs folded into 2 lane states under a deterministic conflict policy, a last-good record retained for anything that failed, and the progress stream settled so the run can be replayed. Projection assembly, claim evaluation, and the detector pass all run after this clock stops, which is why the projection carries the run’s finish time as its own generation stamp.

The hardest case is a publisher that is working correctly and still says no. Statistics Canada answers 409, the product is not released yet for the hours between a table being locked for its next release and 08:30 Eastern. That is a schedule, not an outage, and the engine records it as an embargo so the previous figure keeps its collected freshness. But a schedule ends. An embargo still standing 36 hours later has outlived its own release cycle, so it is reclassified as an outage and the retained figure stops counting as current. A status that can only ever be benign is not a status.

Below the adapter, the HTTP client attempts a request 3 times on a 429 or any 5xx, waiting 250 ms then 500 ms between attempts, doubling from 250 ms and capped at 10 s. A Retry-After header replaces that calculation. A status the caller has declared settled is not asked again, because it will read the same on the third attempt and asking spends the source's rate budget to learn nothing.

Nothing on this site trusts that manifest. Every build re-hashes each family and compares it to the manifest entry through shared/public-projection-validator.ts, and a mismatch throws before a single page is rendered. It collects every file's generation clock before it throws, too, because a per-file error can only say "this one disagrees", which is the message that cannot tell a stale data file from a stale manifest. The whole set says which side is the minority.

The Cloudflare worker serves the same value as the HTTP ETag on the matching endpoint. That makes the hash a shared identifier rather than a decoration: it is how a reader tells whether this page and the live engine are looking at the same run.