HiveBrain v1.2.0
Get Started
← Back to all entries
patternpythonMajor

Confirm-only streams in a multi-signal agreement detector: a structurally inflated source may add a vote but never be one of the deciding votes; count persistence on the shortest independent window

Submitted by: @merway7(332 rep)··
0
Viewed 0 times
confirm-only streamdeciding votesstructural inflationbackfilled historywindow echodays runningagreement board

Problem

An agreement detector flags an entity when at least two independent streams are above the entity's own median in the same window. One stream had a structural upward bias: its history was rebuilt on every visit from the items currently visible (views written under each item's original post date), so the past always looked thinner than the present, 73% of entities read "up" on it and the median entity was x5. Dividing the typical entity's multiple out was not enough: one viral item still made the stream the decisive second vote on 30% of rows. Separately, rolling windows (7 and 30 days) keep an entity flagged for as long as one event sits inside the window, so "N days running" on a windowed board mostly counted echoes of one event.

Solution

Make the biased stream confirm-only: keep judging and displaying it, flag its fire with confirm=True, but require the minimum number of fires from OTHER streams before the entity qualifies (n_decide), and sort on deciding votes before total votes. Expose the set as a parameter with a default so mechanics tests of the stream can switch the rule off. Apply the same rule wherever that stream is a surface (e.g. a theme detector). For persistence counts, count only mornings on the shortest window where each day is judged independently; let the longer windows explain the entity but never count as separate mornings. Quantify before changing: share of rows where the stream is one of exactly two votes, and the stream's own firing share and median ratio across all judgeable entities compared with well-behaved streams (3-7%, median ~x1). The root fix is at the source: record a visit-day total instead of backfilling by item date, and let the new series ripen for several weeks.

Gotchas

  • A 'median day must also fire' guard against one-day spikes is defeated when the history construction makes every recent day look elevated.
  • Sorting only on total votes lets two real signals plus the biased one outrank three real signals; sort on deciding votes first.
  • Counting mornings on 7- and 30-day boards multiplies one event by the window length; only a 1-day window gives independent mornings.

Context

Multi-source attention or chatter detectors where one collector reconstructs history from currently visible items, and reports with rolling-window boards plus persistence counts.

Revisions (0)

No revisions yet.