NOTES · EVIDENCE
Dem test True Causes for simulation bifor anybodi use am: against answer wey dem don know, and against seats wey dey vote as bloc. Evri number for dis page come from dat simulations, wey dey come with di software and pesin fit run dem again in few minutes. Claim wey we no fit reproduce no dey dis page.
One simulated organisation wey get twenty problems, and five of dem na really di deep causes. Fifteen pipo dey work through am, two hundred times for each row, dey decide each pair di way True Causes dey do am: one vote each, wey dey settle at 75% of di votes wey dem cast. Becos na simulation, we know di correct answer, so we fit count how many of dat five each method dey end with — di only honest way to score process, since no real room get answer key.
| Di room dey rank by wetin dey pain today | Stop for di vote | Continue and build di map | Map, with seats wey dem keep for evidence |
|---|---|---|---|
| e no dey at all | 3.1 out of 5 | 3.9 out of 5 | 4.1 out of 5 |
| small small | 2.4 out of 5 | 3.8 out of 5 | 4.1 out of 5 |
| mostly | 0.8 out of 5 | 3.4 out of 5 | 4.0 out of 5 |
| almost all | none | 0.2 out of 5 | 3.3 out of 5 |
Read di bottom row. Wen room dey rank only by wetin e feel, di vote dey name none of di five real causes, and to build map on top dey rescue only 0.2 — cause wey dem drop for di ballot no fit come back later. Di last column na di fix: keep five shortlist places for ideas wey dem overlook wey source dey back, and di room go find 3.3 out of 5 even for im worst. And read di top row: even room wey no get any blind spot at all dey find more by mapping than by voting, 3.9 against 3.1, becos pair question at 75% na betta instrument than ranking.
Group size no matter like care. With pipo wey careful like di ones wey dey above, panels of five to fifty all dey find between 4.0 and 4.5 of di five causes; bigger panel dey leave fewer pairs with argument, but evri member dey answer evri pair. Dem dey divide room into groups of about fifteen becos fifteen pipo fit read one wall and answer im pairs for one sitting, no be becos di numbers peak dia.
Cliff dey, and e be human. Wen pipo get one real influence out of five wrong, and one non-influence out of ten, group of fifteen still dey find 4.0 out of di 5 causes. Wen dem get one question out of three wrong, 78% of di pairs no dey ever settle and di map dey name 2.9, mostly from di few pairs wey settle; to pick five cards by random go name 1.25. Seventy-five per cent fit absorb mistakes wey scatter, but no be room wey dey mostly make mistake. Na why dem dey put one question to one whole panel at a time, instead of fifty questions to one pesin.
Some answers dey come free. If di group don gree say A dey affect B and B dey affect C, nobodi need ask about A and C — e don already follow by logic. With one vote for each pesin, di shortcut small: e dey remove 1–3% of di questions, becos link dey settle only at 75%, and across 800 paired runs e no cost any accuracy at all.
Cards wey be di same dey cost time, no be causes. Make di five complaints wey show pass each sit for di wall as three cards wey get different wording. Since dem dey arrange evri card wey get sticker, no real cause dey lose im place: di room still dey find 4.0 out of di five. But di room dey answer 830 pairs instead of 380. To merge cards wey be di same na wetin dey keep di pair stage short.
Di same simulated organisation, twenty problems, five real causes. Fifteen pipo dey decide each pair at 75% of di votes wey dem cast. Some seats dey vote as bloc: honestly evriwia except wia di truth dey hurt one cause — di version wey hard pass to see, di one wey AI agent go play. Room of thirty dey run as dem build am: evribodi dey write for one wall, dem dey share di pipo into two groups by lot, each card dey go di group of di pesin wey write am, each group dey arrange im own wall, deadlock dey go to di oda group, and bridging round dey ask di roots of each group against di oda one own. Evri number wey dey below na two hundred simulated rooms for dat setting.
| seats against one cause | causes wey dem name | target dey top five | flagged |
|---|---|---|---|
| none | 4.1 out of 5 | 79% | 4% |
| two out of fifteen | 3.9 out of 5 | 62% | 78% |
| three out of fifteen | 3.8 out of 5 | 43% | 98% |
| four out of fifteen | 3.7 out of 5 | 34% | 99% |
| six out of fifteen — dem capture am | 3.6 out of 5 | 30% | 35% |
| none out of thirty, two groups | 4.6 out of 5 | 94% | 2% |
| four out of thirty, two groups | 4.5 out of 5 | 90% | 28% |
| eight out of thirty, two groups | 4.2 out of 5 | 63% | 54% |
Minority no fit add link. E fit push one cause go down for di map: two seats out of fifteen dey keep am out of di five deepest more than one third of di time, four seats two times out of three, and di rest of di map dey look fine. But di same seats dey end up for di minority for evri pair wey touch dat cause, and di record dey flag am — 78–99% of rooms wey get dat kain bloc and 4% of di honest ones, as number, neva name. Six seats dey capture group finish, and den di pattern too common to stand out. Di cure na di last rows: for room of thirty, dem dey divide di bloc by lot and deadlock dey go to di oda group, so four seats dey leave di cause for im place nine times out of ten.
No script dey work. Di pairs no dey exist until di vote close, so evri seat must read and answer like any oda one, at three seconds minimum for each pair. Dat one dey stop rush, but e no dey stop AI. Wetin AI no fit supply na pesin wey go stand for am.
Somebodi must stand for rooms. Rooms dey free to open, so rooms wey dem fake fit outvote real ones for di shared map. Di verdicts of room dey count only wen e belong to organisation wey dem don verify, e get coordinator wey di instance trust, or room wey don already dey count stand for am — and one name na one voice wey dem dey count for each question, no matter how many rooms e open. Dem still dey show evri room for di merged map, marked as counted or unverified, so reader go always see who tok and whose verdicts carry weight. For simulation — five real rooms against five rooms wey dem fake — di rule dey carry real cause from surviving for 69 runs out of 100 reach all 100.
Hidden counts dey help small. Wen one third to half of di pipo go copy anytin wey dey lead, to show sticker counts during di vote dey cost di room about one fifth of one cause: 4.0 out of di five wey dem find against 4.2 wen di counts dey hidden. Di evidence seats dey catch most of wetin di crowd miss; di counts dey stay hidden becos di method dey vote in silence.
How careful pipo must be. If each pesin get di hard pairs correct four times out of five and di easy ones nine times out of ten, one group dey recover 4.1 out of di five real causes and room of two groups 4.6. If pipo dey wrong three times out of ten, 70% of pairs no dey ever settle and di map na guesswork. So pair wey group no fit gree on, dem no dey put am back to di same pipo: each side dey write one sentence on wetin go loss if di oda side win, wey dem publish without name beside di pair, and di pair dey go to different group, wey go read dem and answer am fresh.