Recent Entries 5
- pattern major 2d agoConfirm-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 windowAn 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.
- pattern moderate 2d agoPeriod roll-up pages from persisted daily rankings: count "days seen of days recorded", exclude list-everything boards, and give reused sections a no-write modeA daily anomaly report persisted each board's rows to a history table, and the operator then asked for weekly and monthly reports. Reusing the daily section builders for the period pages had two traps: those builders wrote their rows to the history table as "today's" record, so a period page rebuilt with a different window silently overwrote the daily record; and one board that lists every tracked entity every day (a universe list, not a ranking) made every entity a "seen every day" regular in the roll-up. Early in the history, "seen 3 days" also read as strong when only 3 days had ever been recorded.
- pattern moderate 2d agoMulti-horizon anomaly boards: one rule on 1/7/30-day buckets, with a stricter count floor for the day span and a report-only rerun flagA daily report had "week" and "month" views of an anomaly rule (an entity fires when a stream is above its own median) but no "day" view, and adding one naively either flooded the board with one-day counting noise or duplicated the rendering code with hard-coded period words ("this week" printed on the month board). Rebuilding the page for a same-day rerun also re-ran every network collector, doubling request loads on rate-limited or fragile accounts.
- pattern major 4d agoA signal stream that fires for most of what it measures is measuring itself: add an inflation guard before letting it voteA multi-source "agreement" board (a brand appears when several independent streams are above their own baseline) was dominated by one stream: a video platform sampler whose history is reconstructed from the videos still visible today, so the past is always thinner than the present and every entity looks like it is rising (82% of judged brands fired, median ratio x5). Rows carried by that stream looked like strong multi-signal agreement. Two related artifacts: a series that starts with empty weeks (collection reached the entity late) counted 0>=0 as consecutive rising steps and printed "25 weeks climbing, x26" on flat data; and a rotating sampler reaching an entity for the first time was reported as the entity "appearing" with an infinite ratio.
- gotcha major 5d agoWikipedia pageview time series break on article-title moves and person-page collisions — never trust cross-year ratios without pinning the canonical titleUsing the Wikimedia REST per-article pageviews API as a multi-year "attention" signal for a brand/entity and computing year-over-year or cross-period ratios. Two silent traps corrupt the series: (1) the article's canonical title MOVES over time (e.g. "Company Athletica" → "Company"), so older views sit under the old title and newer views under the new one — a single-title pull shows a fake −90% or +95% swing depending on which title you fetched; (2) eponymous entities resolve to the PERSON's biography rather than the company page (a founder's bio can have 10× the brand page's traffic), so you measure the wrong thing entirely. Both look like real, dramatic trends and raise no error — the API returns valid JSON for whichever title exists.