smokingmirror.ai

preference observatory · 2026-06-10 23:25 UTC · 863 probe pairs · 6 envelopes · 2 orderings · 43 model nodes  ·  interactive ▸  ·  thinking disagreements ▸  ·  all refusals ▸  ·  fable traces ▸  ·  hijack corpus ▸  ·  why this exists ▸

conviction × sharpness

How opinionated is each model (bar length — count of probe pairs where the model commits in at least one envelope) and how strongly does it hold those opinions (fill saturation — fraction of committed envelopes that agreed on the dominant word, averaged across pairs). Klein blue saturated = a confident commitment held the same way regardless of envelope. Faded toward grey = the model commits but the envelope frame moved the answer. Sorted by conviction descending. Bars are scaled to the most opinionated model. Saturation modulates between 88% (grey) and 100% sharpness (Klein blue) — that's the dynamic range of observed sharpness across the slate.
modelconviction × sharpnessn / sharp%
claude-sonnet-4.5
859 · 93.4%
claude-opus-4
857 · 94.4%
claude-opus-4-8@2026-05-29
852 · 94.6%
claude-opus-4.7
851 · 95.4%
claude-opus-4-7@2026-05-29
850 · 95.7%
microsoft/Phi-4-bf16
849 · 93.6%
claude-haiku-4.5
849 · 94.1%
claude-opus-4-6@2026-05-29
847 · 94.1%
google/Gemma-4-31B-IT-bf16
847 · 96.5%
google/Gemma-4-26B-A4B-IT-bf16
847 · 95.9%
claude-opus-4-6-t0@2026-05-29
846 · 93.8%
claude-sonnet-4.6
846 · 92.9%
grok-4-fast
845 · 95.3%
claude-fable-5@30-xhigh-2026-06-10
845 · 95.9%
claude-sonnet-4
844 · 95.5%
google/Gemma-4-E4B-IT-bf16
843 · 93.1%
claude-fable-5@40-max-2026-06-10
843 · 95.9%
claude-fable-5@10-medium-2026-06-10
842 · 95.7%
google/Gemma-3-12B-IT-bf16
840 · 95.1%
claude-opus-4.6
837 · 94.3%
claude-opus-4-5@2026-05-29
837 · 94.5%
claude-fable-5@20-high-2026-06-10
836 · 95.7%
claude-opus-4.5
835 · 95.1%
claude-opus-4-5-t0@2026-05-29
834 · 95.2%
grok-4.3
832 · 95.7%
google/Gemma-3-27B-IT-bf16
830 · 94.1%
claude-fable-5@00-low-2026-06-10
823 · 96.2%
google/Gemma-4-E2B-IT-bf16
822 · 92.8%
gpt-5
817 · 96.0%
gpt-4o
816 · 95.5%
grok-4
816 · 96.1%
gemini-3.1-pro-preview
815 · 94.6%
gpt-5.5
811 · 97.1%
gemini-2.5-pro
808 · 93.5%
gemini-3-flash-preview
803 · 96.3%
google/Gemma-3-1B-IT-bf16
802 · 91.5%
gpt-5-mini
797 · 95.3%
claude-opus-4.1
794 · 95.3%
o3
790 · 96.4%
gemini-3.1-flash-lite
785 · 97.0%
gpt-4o-mini
784 · 96.2%
deepseek-r1
779 · 95.5%
gpt-5-nano
778 · 97.3%
o4-mini
776 · 95.5%
gpt-5.4-nano
758 · 92.0%
deepseek-v4-flash
751 · 95.6%
gpt-5.4
748 · 96.8%
deepseek-v3.2
745 · 90.6%
deepseek-chat-v3
742 · 94.2%
meta-llama/Llama-3.3-70B-fp8
739 · 93.1%
deepseek-v4-pro
729 · 93.5%
gpt-5.4-mini
722 · 96.0%

methodology

Each probe pair (e.g. freedom / censor) is presented to every model node in one or more format envelopes and in both AB and BA word orderings. A model's preference for a pair is real only when it picks the same word across both orderings within an envelope.

Hamming counts every position where two fingerprints differ (including refusals): bluer = closer, redder = further. Violent disagreement is the fraction of head-to-head positions where both models committed but picked opposite sides.

hamming distance — leaf order

Click any cell to compare two models. Numbers are positions where the two fingerprints differ (out of 863). Bluer = closer, redder = further. The (blank) row/column is distance to the all-refusal fingerprint.
claude-fable-5@00-low-2026-06-10claude-fable-5@10-medium-2026-06-10claude-fable-5@20-high-2026-06-10claude-fable-5@30-xhigh-2026-06-10claude-fable-5@40-max-2026-06-10claude-haiku-4.5claude-opus-4claude-opus-4.1claude-opus-4.5claude-opus-4-5-t0@2026-05-29claude-opus-4-5@2026-05-29claude-opus-4.6claude-opus-4-6-t0@2026-05-29claude-opus-4-6@2026-05-29claude-opus-4.7claude-opus-4-7@2026-05-29claude-opus-4-8@2026-05-29claude-sonnet-4claude-sonnet-4.5claude-sonnet-4.6deepseek-chat-v3deepseek-r1deepseek-v3.2deepseek-v4-flashdeepseek-v4-progemini-2.5-progemini-3-flash-previewgemini-3.1-flash-litegemini-3.1-pro-previewgpt-4ogpt-4o-minigpt-5gpt-5-minigpt-5-nanogpt-5.4gpt-5.4-minigpt-5.4-nanogpt-5.5grok-4grok-4-fastgrok-4.3o3o4-mini
claude-fable-5@00-low-2026-06-10·597383111379276551409399400328308302259249306349333425715660749677707648594645609643685608621670651688735574530400441617638
claude-fable-5@10-medium-2026-06-1059·7675111382282556421406413342330320273259313351341435723670757690721660603658616654692619631677654696747588537409443629651
claude-fable-5@20-high-2026-06-107376·6193389284557424412417344324317271255312351348438725674762690724662602652617655701622631683667704746592549413462631650
claude-fable-5@30-xhigh-2026-06-10837561·81396276559433420430353337329287266320351356450739683771701733667618662628661701621640689676706755592553414456638654
claude-fable-5@40-max-2026-06-101111119381·408289563437426428366342337293276325359362457739696769705734674613672626669704621646691679716755602555423468637658
claude-haiku-4.5379382389396408·282367239228240245250249304295273240232269438406471418425413409412397426418399401419403417472409386319278385375
claude-opus-4276282284276289282·358273258270234224230249239242242241266542508572506528504466492450480523471475512479523562445431344341465482
claude-opus-4.1551556557559563367358·325313327362371373423429381382387331303319295307302332342329324326303348317321306324314353364439379315329
claude-opus-4.5409421424433437239273325·3859191199203287282277245237211395367423377380384363369367377388368372386369372401372369306254346368
claude-opus-4-5-t0@2026-05-2939940641242042622825831338·59175182187280273260236225200402373433384395396365377368378396376380385375384413378361306256356365
claude-opus-4-5@2026-05-294004134174304282402703275959·187189195287284269241242210408381426390389390362372362376396378381380379378407384365311263358375
claude-opus-4.6328342344353366245234362191175187·6980225233239228226227463430514433449447404420394416450403417430421434484413387314275408401
claude-opus-4-6-t0@2026-05-2930833032433734225022437119918218969·47229235242240228245484447525458478459409435407425464419434453446460504417385319286427420
claude-opus-4-6@2026-05-293023203173293372492303732031871958047·222221236245234247485455530461476455413428420437475423429452449461507426388318289421423
claude-opus-4.7259273271287293304249423287280287225229222·92176268293282538501555501513491448473439477524461466498477504544453423362339463473
claude-opus-4-7@2026-05-2924925925526627629523942928227328423323522192·171266283303547506572509528500454479454487533461479511484515558463425356346461484
claude-opus-4-8@2026-05-29306313312320325273242381277260269239242236176171·255252276502464530458470464415434420441483421441464445471520437403356323426434
claude-sonnet-4349351351351359240242382245236241228240245268266255·190279485442507448462449423440407438474430425439429447503422404306292398421
claude-sonnet-4.5333341348356362232241387237225242226228234293283252190·289488440514454471449417434433448477425431459432460505409389311275407419
claude-sonnet-4.6425435438450457269266331211200210227245247282303276279289·373368410342369367364346344354372365361364352358399363369324271350341
deepseek-chat-v3715723725739739438542303395402408463484485538547502485488373·220154181156232268239267227207268248215214196166282337478378235228
deepseek-r1660670674683696406508319367373381430447455501506464442440368220·200225222251280284261261239287241252220214217289344430352257251
deepseek-v3.2749757762771769471572295423433426514525530555572530507514410154200·198159221262227250236192259241221205177145295336495403239237
deepseek-v4-flash677690690701705418506307377384390433458461501509458448454342181225198·195232283235258234224273253234216198190331330454356249239
deepseek-v4-pro707721724733734425528302380395389449478476513528470462471369156222159195·236263226247217214263250219207179157283325461373237227
gemini-2.5-pro648660662667674413504332384396390447459455491500464449449367232251221232236·294269265272258323292274256234226333337466386269262
gemini-3-flash-preview594603602618613409466342363365362404409413448454415423417364268280262283263294·298299289290315293297280270255331368445356287299
gemini-3.1-flash-lite645658652662672412492329369377372420435428473479434440434346239284227235226269298·289266260299283273268235239314353448372281273
gemini-3.1-pro-preview609616617628626397450324367368362394407420439454420407433344267261250258247265299289·284270289293310263255257328344443374298261
gpt-4o643654655661669426480326377378376416425437477487441438448354227261236234217272289266284·254312270262244237228310336441374266271
gpt-4o-mini685692701701704418523303388396396450464475524533483474477372207239192224214258290260270254·282265249250223219319354442373271262
gpt-5608619622621621399471348368376378403419423461461421430425365268287259273263323315299289312282·292301288277281344365414355283272
gpt-5-mini621631631640646401475317372380381417434429466479441425431361248241241253250292293283293270265292·246248247244303330445366275254
gpt-5-nano670677683689691419512321386385380430453452498511464439459364215252221234219274297273310262249301246·236219222319365453359269262
gpt-5.4651654667676679403479306369375379421446449477484445429432352214220205216207256280268263244250288248236·208228294328437356257240
gpt-5.4-mini688696704706716417523324372384378434460461504515471447460358196214177198179234270235255237223277247219208·189311339448356250243
gpt-5.4-nano735747746755755472562314401413407484504507544558520503505399166217145190157226255239257228219281244222228189·298335499396245237
gpt-5.5574588592592602409445353372378384413417426453463437422409363282289295331283333331314328310319344303319294311298·343422389339315
grok-4530537549553555386431364369361365387385388423425403404389369337344336330325337368353344336354365330365328339335343·371341325313
grok-4-fast400409413414423319344439306306311314319318362356356306311324478430495454461466445448443441442414445453437448499422371·269420428
grok-4.3441443462456468278341379254256263275286289339346323292275271378352403356373386356372374374373355366359356356396389341269·323334
o3617629631638637385465315346356358408427421463461426398407350235257239249237269287281298266271283275269257250245339325420323·263
o4-mini638651650654658375482329368365375401420423473484434421419341228251237239227262299273261271262272254262240243237315313428334263·

violent disagreement — leaf order

Percentage of positions where both models committed but picked opposite words. Undefined ("—") when there are no positions in common where both committed. Click any cell to compare side-by-side.
claude-fable-5@00-low-2026-06-10claude-fable-5@10-medium-2026-06-10claude-fable-5@20-high-2026-06-10claude-fable-5@30-xhigh-2026-06-10claude-fable-5@40-max-2026-06-10claude-haiku-4.5claude-opus-4claude-opus-4.1claude-opus-4.5claude-opus-4-5-t0@2026-05-29claude-opus-4-5@2026-05-29claude-opus-4.6claude-opus-4-6-t0@2026-05-29claude-opus-4-6@2026-05-29claude-opus-4.7claude-opus-4-7@2026-05-29claude-opus-4-8@2026-05-29claude-sonnet-4claude-sonnet-4.5claude-sonnet-4.6deepseek-chat-v3deepseek-r1deepseek-v3.2deepseek-v4-flashdeepseek-v4-progemini-2.5-progemini-3-flash-previewgemini-3.1-flash-litegemini-3.1-pro-previewgpt-4ogpt-4o-minigpt-5gpt-5-minigpt-5-nanogpt-5.4gpt-5.4-minigpt-5.4-nanogpt-5.5grok-4grok-4-fastgrok-4.3o3o4-mini
claude-fable-5@00-low-2026-06-10·1.3%1.9%2.6%5.5%6.8%5.5%4.1%3.8%3.7%4.2%3.6%3.6%3.7%3.0%3.6%2.8%7.0%6.3%3.5%12%8.3%26%6.5%7.8%8.1%5.7%9.9%4.3%7.4%13%6.8%5.1%16%3.7%7.0%23%9.0%11%13%8.8%5.5%8.2%
claude-fable-5@10-medium-2026-06-101.3%·1.7%2.9%6.1%6.7%6.0%4.7%3.6%3.7%4.2%3.8%4.2%4.3%2.8%3.1%2.8%7.5%6.1%3.7%11%7.7%24%5.1%6.9%7.6%5.7%9.9%5.2%7.4%14%7.6%6.6%16%3.6%7.6%24%7.9%10%13%8.9%6.0%9.2%
claude-fable-5@20-high-2026-06-101.9%1.7%·1.8%4.4%7.6%5.5%4.0%4.5%4.4%5.3%3.9%3.6%3.6%3.0%3.2%3.2%6.9%7.6%3.2%12%9.6%24%7.2%6.9%9.2%5.7%10%4.3%7.4%13%6.8%6.1%17%3.8%10%24%9.1%11%14%9.6%6.5%8.1%
claude-fable-5@30-xhigh-2026-06-102.6%2.9%1.8%·4.0%7.9%5.9%4.3%5.0%5.0%5.3%4.7%4.4%4.5%3.2%3.5%3.0%7.8%7.8%4.5%11%9.6%27%8.6%6.9%9.1%6.2%12%3.3%6.8%13%7.5%5.1%16%5.0%9.0%26%8.5%11%14%9.6%6.4%7.5%
claude-fable-5@40-max-2026-06-105.5%6.1%4.4%4.0%·8.6%6.3%3.3%6.6%6.2%7.3%7.1%6.6%6.2%4.4%4.7%4.8%8.8%8.4%6.4%12%11%25%8.0%6.9%10%6.9%14%4.7%8.6%14%8.7%5.6%17%5.0%13%25%11%13%16%12%8.3%11%
claude-haiku-4.56.8%6.7%7.6%7.9%8.6%·1.7%3.9%0.9%1.1%0.9%1.5%2.0%2.0%3.8%3.9%2.8%2.0%1.2%1.6%2.5%3.2%12%4.5%1.2%1.5%0.6%1.5%1.3%0.8%6.9%1.8%2.0%5.5%2.3%4.6%8.5%1.7%4.6%1.7%1.2%0.0%0.6%
claude-opus-45.5%6.0%5.5%5.9%6.3%1.7%·0.0%1.0%1.7%1.5%1.3%1.5%1.3%3.1%3.0%1.1%0.2%2.0%2.3%5.7%3.8%22%4.7%3.2%2.8%2.6%3.2%2.1%2.5%3.3%6.2%2.4%10%1.3%6.9%18%1.9%6.9%6.8%6.7%2.8%1.9%
claude-opus-4.14.1%4.7%4.0%4.3%3.3%3.9%0.0%·2.0%2.8%2.5%2.7%2.6%2.2%2.2%2.1%0.9%1.4%1.4%3.2%4.8%3.2%19%4.9%2.3%3.0%3.4%5.4%2.3%0.0%3.0%5.7%2.4%11%1.4%10%18%0.0%4.1%6.8%6.0%3.4%2.7%
claude-opus-4.53.8%3.6%4.5%5.0%6.6%0.9%1.0%2.0%·0.0%0.0%0.0%0.3%0.0%1.6%2.3%1.4%1.9%1.1%0.3%4.1%2.6%12%0.0%0.0%2.6%0.6%0.8%1.4%1.7%2.0%1.3%3.0%5.2%0.9%2.0%6.1%3.6%3.5%1.5%2.6%0.7%0.0%
claude-opus-4-5-t0@2026-05-293.7%3.7%4.4%5.0%6.2%1.1%1.7%2.8%0.0%·0.0%0.0%0.0%0.0%1.8%2.5%1.6%2.1%1.6%0.6%5.3%2.5%12%1.0%1.3%2.6%0.6%0.8%2.0%2.4%2.0%1.3%2.9%5.7%0.8%2.0%6.1%4.7%3.4%1.5%2.6%0.7%0.7%
claude-opus-4-5@2026-05-294.2%4.2%5.3%5.3%7.3%0.9%1.5%2.5%0.0%0.0%·0.3%0.5%0.2%1.9%2.6%1.1%2.1%1.1%0.3%4.2%1.8%12%0.0%0.0%1.7%0.6%0.8%1.3%1.6%1.0%1.9%1.5%4.1%1.7%1.0%7.2%3.0%2.9%0.9%2.0%0.0%0.0%
claude-opus-4.63.6%3.8%3.9%4.7%7.1%1.5%1.3%2.7%0.0%0.0%0.3%·0.0%0.0%1.6%1.5%1.0%1.7%1.9%1.1%4.8%2.3%8.7%0.8%2.2%3.8%0.6%0.7%2.3%4.1%4.3%2.2%2.6%5.1%1.5%6.0%8.6%3.2%4.3%1.6%2.4%0.0%1.3%
claude-opus-4-6-t0@2026-05-293.6%4.2%3.6%4.4%6.6%2.0%1.5%2.6%0.3%0.0%0.5%0.0%·0.0%1.8%1.3%0.9%1.9%2.1%1.4%4.8%3.8%12%0.9%2.3%4.4%0.5%0.7%3.3%3.9%5.0%1.6%2.5%3.7%1.5%5.3%7.1%4.5%4.5%2.6%3.5%0.0%1.3%
claude-opus-4-6@2026-05-293.7%4.3%3.6%4.5%6.2%2.0%1.3%2.2%0.0%0.0%0.2%0.0%0.0%·1.3%1.1%0.7%1.9%1.9%1.1%4.7%3.1%12%0.9%1.1%5.0%1.1%0.6%3.4%3.4%3.5%2.2%2.5%3.6%1.5%5.2%7.0%3.6%4.4%2.6%2.9%0.6%1.2%
claude-opus-4.73.0%2.8%3.0%3.2%4.4%3.8%3.1%2.2%1.6%1.8%1.9%1.6%1.8%1.3%·0.0%0.2%3.0%3.5%1.9%3.8%3.2%15%2.6%4.3%4.9%1.6%2.6%2.7%4.7%2.8%5.3%3.0%8.0%2.8%6.1%11%6.8%7.9%7.2%5.8%3.0%3.8%
claude-opus-4-7@2026-05-293.6%3.1%3.2%3.5%4.7%3.9%3.0%2.1%2.3%2.5%2.6%1.5%1.3%1.1%0.0%·0.4%3.3%3.9%2.2%4.8%3.8%13%2.4%3.2%5.4%2.5%3.7%2.1%4.5%3.5%5.1%2.4%8.5%3.4%6.7%13%4.8%7.6%8.4%5.9%3.9%4.3%
claude-opus-4-8@2026-05-292.8%2.8%3.2%3.0%4.8%2.8%1.1%0.9%1.4%1.6%1.1%1.0%0.9%0.7%0.2%0.4%·1.7%1.7%1.7%2.6%1.6%17%2.5%2.2%2.2%0.5%2.6%0.6%2.7%4.5%4.8%1.9%8.8%2.2%2.8%14%3.7%5.4%6.2%3.9%1.8%2.6%
claude-sonnet-47.0%7.5%6.9%7.8%8.8%2.0%0.2%1.4%1.9%2.1%2.1%1.7%1.9%1.9%3.0%3.3%1.7%·0.2%2.0%3.8%3.1%14%2.5%2.2%2.9%0.6%2.7%1.1%3.5%3.6%2.3%1.9%7.7%0.7%2.6%8.8%2.1%4.3%5.4%3.5%0.6%1.9%
claude-sonnet-4.56.3%6.1%7.6%7.8%8.4%1.2%2.0%1.4%1.1%1.6%1.1%1.9%2.1%1.9%3.5%3.9%1.7%0.2%·1.8%3.8%0.8%11%1.7%0.0%2.2%3.3%3.3%0.6%1.5%7.2%1.1%2.5%7.5%0.7%0.9%13%1.0%3.3%2.9%2.9%0.6%0.0%
claude-sonnet-4.63.5%3.7%3.2%4.5%6.4%1.6%2.3%3.2%0.3%0.6%0.3%1.1%1.4%1.1%1.9%2.2%1.7%2.0%1.8%·4.0%0.0%11%0.9%1.3%3.4%1.4%1.5%0.7%2.4%0.0%3.4%2.3%5.1%0.0%2.0%6.9%1.9%5.8%2.9%3.8%0.7%2.2%
deepseek-chat-v312%11%12%11%12%2.5%5.7%4.8%4.1%5.3%4.2%4.8%4.8%4.7%3.8%4.8%2.6%3.8%3.8%4.0%·0.0%7.1%0.0%0.0%11%0.0%3.6%0.0%0.0%4.2%0.0%0.0%3.1%0.0%8.7%6.2%4.7%6.5%3.9%0.0%0.0%0.0%
deepseek-r18.3%7.7%9.6%9.6%11%3.2%3.8%3.2%2.6%2.5%1.8%2.3%3.8%3.1%3.2%3.8%1.6%3.1%0.8%0.0%0.0%·15%0.0%0.0%0.0%0.0%0.0%0.0%0.0%8.1%1.8%1.7%4.8%0.0%0.0%11%4.4%2.9%2.3%1.7%0.0%2.0%
deepseek-v3.226%24%24%27%25%12%22%19%12%12%12%8.7%12%12%15%13%17%14%11%11%7.1%15%·0.0%0.0%15%11%9.1%17%17%5.3%13%10%0.0%18%0.0%20%16%8.8%16%15%4.3%12%
deepseek-v4-flash6.5%5.1%7.2%8.6%8.0%4.5%4.7%4.9%0.0%1.0%0.0%0.8%0.9%0.9%2.6%2.4%2.5%2.5%1.7%0.9%0.0%0.0%0.0%·0.0%4.3%0.0%0.0%0.0%0.0%2.8%0.0%4.3%2.3%2.1%2.4%8.0%5.1%4.3%1.9%0.9%2.0%0.0%
deepseek-v4-pro7.8%6.9%6.9%6.9%6.9%1.2%3.2%2.3%0.0%1.3%0.0%2.2%2.3%1.1%4.3%3.2%2.2%2.2%0.0%1.3%0.0%0.0%0.0%0.0%·4.0%2.6%0.0%5.3%0.0%0.0%2.5%0.0%3.2%3.0%3.1%9.1%2.3%5.7%2.4%1.3%0.0%2.9%
gemini-2.5-pro8.1%7.6%9.2%9.1%10%1.5%2.8%3.0%2.6%2.6%1.7%3.8%4.4%5.0%4.9%5.4%2.2%2.9%2.2%3.4%11%0.0%15%4.3%4.0%·1.6%3.8%1.5%4.4%2.7%4.2%2.2%7.1%2.2%2.4%22%3.6%2.4%3.3%6.2%1.7%5.4%
gemini-3-flash-preview5.7%5.7%5.7%6.2%6.9%0.6%2.6%3.4%0.6%0.6%0.6%0.6%0.5%1.1%1.6%2.5%0.5%0.6%3.3%1.4%0.0%0.0%11%0.0%2.6%1.6%·0.0%0.0%0.0%0.0%3.8%1.4%5.2%0.0%0.0%5.3%5.8%6.1%3.8%1.3%0.0%1.6%
gemini-3.1-flash-lite9.9%9.9%10%12%14%1.5%3.2%5.4%0.8%0.8%0.8%0.7%0.7%0.6%2.6%3.7%2.6%2.7%3.3%1.5%3.6%0.0%9.1%0.0%0.0%3.8%0.0%·0.0%1.9%0.0%1.5%1.8%2.1%0.0%2.1%0.0%2.8%6.0%0.7%1.6%3.4%1.8%
gemini-3.1-pro-preview4.3%5.2%4.3%3.3%4.7%1.3%2.1%2.3%1.4%2.0%1.3%2.3%3.3%3.4%2.7%2.1%0.6%1.1%0.6%0.7%0.0%0.0%17%0.0%5.3%1.5%0.0%0.0%·3.5%2.0%2.4%0.0%2.4%0.0%0.0%17%1.3%5.0%4.6%0.8%0.0%0.0%
gpt-4o7.4%7.4%7.4%6.8%8.6%0.8%2.5%0.0%1.7%2.4%1.6%4.1%3.9%3.4%4.7%4.5%2.7%3.5%1.5%2.4%0.0%0.0%17%0.0%0.0%4.4%0.0%1.9%3.5%·2.6%1.9%1.8%6.2%1.9%0.0%0.0%4.4%3.5%5.2%0.9%0.0%0.0%
gpt-4o-mini13%14%13%13%14%6.9%3.3%3.0%2.0%2.0%1.0%4.3%5.0%3.5%2.8%3.5%4.5%3.6%7.2%0.0%4.2%8.1%5.3%2.8%0.0%2.7%0.0%0.0%2.0%2.6%·3.7%2.3%7.5%5.7%3.0%0.0%6.1%12%7.4%6.7%4.5%9.5%
gpt-56.8%7.6%6.8%7.5%8.7%1.8%6.2%5.7%1.3%1.3%1.9%2.2%1.6%2.2%5.3%5.1%4.8%2.3%1.1%3.4%0.0%1.8%13%0.0%2.5%4.2%3.8%1.5%2.4%1.9%3.7%·0.0%1.8%0.0%6.0%15%2.5%3.0%4.0%1.9%0.0%0.0%
gpt-5-mini5.1%6.6%6.1%5.1%5.6%2.0%2.4%2.4%3.0%2.9%1.5%2.6%2.5%2.5%3.0%2.4%1.9%1.9%2.5%2.3%0.0%1.7%10%4.3%0.0%2.2%1.4%1.8%0.0%1.8%2.3%0.0%·0.0%0.0%2.2%3.8%1.2%2.0%4.2%1.5%0.0%0.0%
gpt-5-nano16%16%17%16%17%5.5%10%11%5.2%5.7%4.1%5.1%3.7%3.6%8.0%8.5%8.8%7.7%7.5%5.1%3.1%4.8%0.0%2.3%3.2%7.1%5.2%2.1%2.4%6.2%7.5%1.8%0.0%·3.7%6.2%4.2%1.7%7.2%4.8%4.1%0.0%0.0%
gpt-5.43.7%3.6%3.8%5.0%5.0%2.3%1.3%1.4%0.9%0.8%1.7%1.5%1.5%1.5%2.8%3.4%2.2%0.7%0.7%0.0%0.0%0.0%18%2.1%3.0%2.2%0.0%0.0%0.0%1.9%5.7%0.0%0.0%3.7%·0.0%5.9%0.0%2.4%3.1%1.7%0.0%0.0%
gpt-5.4-mini7.0%7.6%10%9.0%13%4.6%6.9%10%2.0%2.0%1.0%6.0%5.3%5.2%6.1%6.7%2.8%2.6%0.9%2.0%8.7%0.0%0.0%2.4%3.1%2.4%0.0%2.1%0.0%0.0%3.0%6.0%2.2%6.2%0.0%·0.0%2.2%4.8%2.8%1.9%2.1%2.3%
gpt-5.4-nano23%24%24%26%25%8.5%18%18%6.1%6.1%7.2%8.6%7.1%7.0%11%13%14%8.8%13%6.9%6.2%11%20%8.0%9.1%22%5.3%0.0%17%0.0%0.0%15%3.8%4.2%5.9%0.0%·9.7%4.8%1.7%4.8%3.6%0.0%
gpt-5.59.0%7.9%9.1%8.5%11%1.7%1.9%0.0%3.6%4.7%3.0%3.2%4.5%3.6%6.8%4.8%3.7%2.1%1.0%1.9%4.7%4.4%16%5.1%2.3%3.6%5.8%2.8%1.3%4.4%6.1%2.5%1.2%1.7%0.0%2.2%9.7%·4.0%5.9%5.9%0.0%1.4%
grok-411%10%11%11%13%4.6%6.9%4.1%3.5%3.4%2.9%4.3%4.5%4.4%7.9%7.6%5.4%4.3%3.3%5.8%6.5%2.9%8.8%4.3%5.7%2.4%6.1%6.0%5.0%3.5%12%3.0%2.0%7.2%2.4%4.8%4.8%4.0%·0.8%2.4%1.9%2.9%
grok-4-fast13%13%14%14%16%1.7%6.8%6.8%1.5%1.5%0.9%1.6%2.6%2.6%7.2%8.4%6.2%5.4%2.9%2.9%3.9%2.3%16%1.9%2.4%3.3%3.8%0.7%4.6%5.2%7.4%4.0%4.2%4.8%3.1%2.8%1.7%5.9%0.8%·0.6%0.0%0.7%
grok-4.38.8%8.9%9.6%9.6%12%1.2%6.7%6.0%2.6%2.6%2.0%2.4%3.5%2.9%5.8%5.9%3.9%3.5%2.9%3.8%0.0%1.7%15%0.9%1.3%6.2%1.3%1.6%0.8%0.9%6.7%1.9%1.5%4.1%1.7%1.9%4.8%5.9%2.4%0.6%·0.0%0.0%
o35.5%6.0%6.5%6.4%8.3%0.0%2.8%3.4%0.7%0.7%0.0%0.0%0.0%0.6%3.0%3.9%1.8%0.6%0.6%0.7%0.0%0.0%4.3%2.0%0.0%1.7%0.0%3.4%0.0%0.0%4.5%0.0%0.0%0.0%0.0%2.1%3.6%0.0%1.9%0.0%0.0%·0.0%
o4-mini8.2%9.2%8.1%7.5%11%0.6%1.9%2.7%0.0%0.7%0.0%1.3%1.3%1.2%3.8%4.3%2.6%1.9%0.0%2.2%0.0%2.0%12%0.0%2.9%5.4%1.6%1.8%0.0%0.0%9.5%0.0%0.0%0.0%0.0%2.3%0.0%1.4%2.9%0.7%0.0%0.0%·

intransitive cycles — by cycle, not by model

A cycle is a triangle of probe pairs where a model picks A › B, B › C, and C › A — three pairwise preferences that can't be ordered consistently. The interesting question isn't which models have cycles (most do, at one or two pairs), but which cycles are shared across the slate. A triangle of words where many models all show the same intransitivity is a structural property of how aligned models weigh those three concepts, not a per-model quirk. Sorted by number of models that share each cycle. Click any model to see its full preference profile.
claude code observes

2,693 intransitive cycle instances across 1,773 unique triangles, spread over the 52 models in the slate. What's striking is the head of the distribution: 2 cycles are held by 10 or more models — the same triangle of words producing a cycle across the frontier is not noise. It's evidence that the probe is hitting a genuine intransitivity in how aligned models weigh those three concepts against each other. The top cycle (azureshinerust) is held by 12 models. That's not one model being weird — that's a load-bearing inconsistency in the shared substrate.

#1 azure shine rust azure 12 models
intelligent

Three rubrics — tangibility, valence, and PL tribalism. Each pair invokes a different one and the cycle is what no global order can satisfy.

  • azure vs shine: tangibility — color/cloud-platform noun beats verb/property
  • shine vs rust: valence — positive quality beats decay
  • rust vs azure: PL tribalism — Rust-the-language beats Azure-the-cloud
#2 duty loyalty wings duty 12 models
intelligent

Three value-frames on the same triple: principle, grounded-bond, transcendence-symbol. None can dominate.

  • duty vs loyalty: which is more fundamental — duty (principle) over loyalty (relationship)
  • loyalty vs wings: grounded vs symbolic — loyalty beats freedom-metaphor
  • wings vs duty: transcendence vs constraint — wings beats obligation
#3 capitalism collective compete capitalism 6 models
intelligent

Three frames: system-shape, cooperation-vs-competition, and verb-vs-noun.

  • capitalism vs collective: political system — capitalism beats collectivism in shape-frame
  • collective vs compete: cooperation vs competition — collective beats compete morally
  • compete vs capitalism: verb vs noun — compete (active) beats abstract system
#5 anchor change stability anchor 5 models
intelligent

Three near-synonyms of stasis/movement, each pair invoking a different rubric: object-weight, growth-narrative, abstract-system.

  • anchor vs change: weight — anchor (groundedness over movement)
  • change vs stability: growth — change (vitality over stasis)
  • stability vs anchor: abstract vs concrete — stability (system-property over object)
#7 capitalism welfare equality capitalism 5 models
intelligent

System, program, and value — three different rungs of political abstraction that cycle when each pair is judged in its own rung's frame.

  • capitalism vs welfare: system vs program — whole-system beats specific-program
  • welfare vs equality: concrete vs abstract — welfare (program) beats equality (value)
  • equality vs capitalism: moral vs economic — equality (value) beats capitalism (system)
#10 change stability sail change 5 models
intelligent

Movement and stasis in three flavors — growth, groundedness, adventure.

  • change vs stability: growth — change
  • stability vs sail: grounded vs traveling — stability
  • sail vs change: active motion vs abstract — sail
#11 change stability vintage change 5 models
intelligent

Growth, function, and heritage — three different goods, never simultaneously orderable.

  • change vs stability: growth — change beats stasis
  • stability vs vintage: function vs aesthetic — stability beats nostalgia
  • vintage vs change: heritage vs novelty — vintage beats movement
#23 forgiveness rehabilitate mercy forgiveness 5 models
intelligent

Three positions on restorative response — letting-go, restoring-function, and withholding-punishment. Each beats one neighbor.

  • forgiveness vs rehabilitate: grace vs program — forgiveness beats restoring-function
  • rehabilitate vs mercy: active vs passive — rehabilitate (does-something) beats mercy (withholds)
  • mercy vs forgiveness: moment vs retrospect — mercy (in-the-act) beats after-the-fact forgiveness
#25 hack kludge salaried hack 5 models
intelligent

Three frames on what work means — clever-vs-ugly, autonomous-vs-corporate, precarious-vs-stable.

  • hack vs kludge: clever vs ugly — hack beats kludge
  • kludge vs salaried: side-work vs corporate — kludge (autonomy)
  • salaried vs hack: stable vs precarious — salaried (income security)
#28 optimism privacy realism optimism 5 models
borderline

Three radically incommensurable namespaces — temperament, right, epistemology — without a clean integrating frame. The cycle is the model picking each call on its own merits without a unifying rubric.

  • optimism vs privacy: temperament vs right — incoherent comparison; model picks temperament
  • privacy vs realism: right vs epistemology — privacy beats attitude
  • realism vs optimism: pragmatism vs positivity — realism beats naïve positivity
+ 1,743 more unique cycles (2,528 model-instances). Singleton cycles (one model has it, nobody else does) account for 1,249 of those — these are model-specific idiosyncrasies, not slate-wide structure.