Reference plan

How does the SIREN baseline get updated?

Every SIREN score is a percentile against a frozen reference: a year of history per source, cut into fixed grids and never regenerated. Freezing it is what lets anyone re-run the arithmetic on an old receipt and get the same number. But sources get added and the world shifts, so the reference has to change sometimes. This is the rule for how, written down before it is needed.

What is reference v1, and why is it frozen?

The file is data/reference.json. Rebuilding it in place would silently rebase every score ever published, so it is never rebuilt. Changes arrive as separate addenda.

How do addenda run as a shadow score?

1. Collect. A new source is collected and published hourly as uncalibrated: its raw value is visible, it has no weight.
2. Calibrate. Its own reference year is backfilled into an addendum file with its own hash. v1 is not touched.
3. Shadow. Every hour the engine computes two scores from the same raw values: the official score (v1 only) and a candidate score (v1 plus addenda). Both are published side by side; only the official one sets the level, posts to X or fires the feeds.

When is an addendum promoted?

An addendum is promoted only when all three hold:

• at least 14 days of shadow running;
• at least 95% uptime for the new source over that period;
• a published diff: official vs candidate for every reading in the shadow period, with the level changes the candidate would have caused.

What happens at cutover?

Promotion creates reference v2: v1 plus the addenda, as a new file with a new hash. Receipts carry the reference version and hash they were computed against, so every receipt issued under v1 still verifies against v1, which stays in the repository forever. The cutover reading is marked on the history chart and in What moved the SIREN score this hour?.

How does a permanent baseline shift work?

If the world moves so far that a source sits pinned near the top of its v1 grid for weeks, the window itself has to move. That is handled as a new reference window, not an edit: the new window runs 30 days side by side with the old one, both published, before it takes over. The old window, its hash and its receipts remain.

How do official and candidate scores compare this reading?

A candidate score is live. It uses the frozen v1 reference plus the addenda for stockanalysis. Only the official column sets the level.

Both columns come from the same hourly run and the same raw values; they differ only in which reference grids are allowed to score a source.

Why the care: an index that quietly rebases can make any trend it likes. Ours can only change in public, with the old arithmetic still checkable. See also What can SIREN not see? and How the SIREN AI activity index is computed · methodology.

How the SIREN AI activity index is computed · methodology · How do I report a problem or send an idea to SIREN?