How the reading is made.

Every figure on a loom surface carries a provenance mark. The mark is not decoration; it is the claim.

Fourteen active feed kinds, one Postgres. Per-feed depth and archive dates live on the receipts page; this table is the kind-level count behind the method, read on 2026-08-18.

Read from the GOSSIP Postgres, schema radar, read-only, 2026-08-18. Items counted on fetched_at over the trailing 7 days. Each lane’s health is checked against the table that lane actually writes: the anchor lane writes its own two tables, and a check that counts radar.items for it reports a confident zero (correction of record, 2026-08-17).

Source groups, polling lanes, the one store, the scoring step, and the surfaces that carry the reading. Dashed wires are staged: they wait on the always-on host. The plate keeps its own ink in both themes.

The feed wiring as of 2026-08-18, after the nine-chain wiring of 2026-08-17/18. Registry and counts queried live the same day; lane cadences read from the pw-radar workflow files. Source drawing: feed-wiring-map in the house repo.

The headline number for every measured scope is a z-score: how far the current window sits from the series’ own recent baseline, in standard deviations. Computed by aggregate.py at the end of the daily ingest run; pure SQL and Python, no language model writes a score.

• Bucketed by published_at, not fetched_at. A newly added feed lands its backfill across the items’ true dates, so an import cannot fabricate a current-day spike, and the news archives give ~100 days of real baseline from day one.
• The noise floor. For count data, variance tracks the mean, so std is floored at max(pstdev, sqrt(mean), 1.0). Without it, a spike over a flat baseline would divide into nothing and a real signal would hide.
• Confidence graduates on its own. A score is thin until its baseline holds 14 days, measured after. No hand-tuning; history accrues and the label follows.
• What gets scored. Eight domains (tech, fashion, music, food, language, work, relationships, body) with fractional attribution by each source’s declared coverage, so a multi-domain source never multi-counts. Five tracked pressure systems as keyword sets (slop, lonely, taste, values, gender; each axis keeps the word the discourse itself uses). Per-source daily volume. GDELT corpus queries against their full-history means.
• The narrative layer is separate. The scan module writes dated signals to radar.signals; it never edits a score. Scores and stories do not share a pen.

The pipeline polls. Every lane runs at the cadence its sources support, and no surface claims a faster one.

• Surfaces report the harvest. Counts, deltas, timestamps, dial values, resolutions. Specimen copy simulates the harvest, never the analyst. A lexical linter gates every deploy of this site; what it cannot catch, review has to.
• Failure renders like success. The retired gmail lane (0 active), the two newsletters deactivated for having no live feed anywhere, the ~2 items/day bluesky sample, named stale feeds: printed in the same voice and at the same size as the working lanes.
• Every number carries its date. A status line may not print a typed constant as though it were an observation; that defect was found on this site’s home page and corrected on the record, 2026-08-17.
• Snapshot provenance on the monitor. The station’s embedded snapshot is run 2026-07-16 with a reading pull of 2026-07-24, both flagged PROVISIONAL in source. The aggregate step does not yet emit the reading block, so a refresh is a manual pull. A method page that hid its manual step would be breaking its own law, so it is stated here instead.

• No stream is live yet. The firehose tier is staged until the worker has its host; the dashes in the wiring are dashes on the live site too.
• The trends archive lane is dead since 2026-07-29: the export path needs a residential IP. An application to the vendor alpha is drafted and unsent; the lane stays marked dead until it lands.
• The datacenter-IP question is open. The first scheduled trends.yml run in CI answers whether the trending RSS survives a datacenter address; until it runs, the lane’s only verified pulls are local.
• Directional dials are judgment. Dated, labeled, scored on a printed scale, and never blended into a measured series.
• Thin means thin. A score whose baseline is younger than 14 days says so on the surface, at the same size as the number.