buqjaj pov, Soch jar, wa'DIch je'wI' tlhej jatlhpa' rep law', AI ghaH tlhej ret qep qaw. qawHaq lughchu'. vagh jar jaj. Dansk Hol jatlh mu'mey revlu'bogh. jatlh QaQ ghun. qawHaq 'e' tlhej ghun mu'mey ren AI — ghot ngo' Sov vaD ghun mu'mey, nuv vaD ghun mu'mey pIm.
qep qaSpu'be'. ghItlh Dem: vagh jar pov vetlh jonwI' 'oH, nIteb ba', 'ej AI pa'vo' chIch pollu'. The memory was fabricated — not maliciously, not carelessly. nIt. AI Sovbogh Doch Hoch pollu'bogh qawHaq rap ngeb Doch wa' chenmoH, vIt Doch jemghach rurbogh.
What caught it was not a rule. AI pov vetlh ta'bogh Hoch lugh. tu'bogh ghogh wa' ghel — yav bIng tob'a'? — QonoS je, qawHaq ra'bogh Daq latlh jangbogh.
tu'ghachvam ghItlh Hoch 'oH.
'ej cha'DIch Doch DaSovnIS, Hoch qaSpa': lut pa'Daq AI ghaH ghItlhvam ghItlhwI' qoch. “maH” juvbogh ghotpu' juvlu'bogh pat ngaS. ghaHvam 'ay' bio-Daq chenmoH 'e' wIHIt. naDev ba' mo' ghaHvam, ghItlh mu'tlhegh Hoch DalaDtaHbogh SeH. Duy ghaHDaj ja'ghach ngaDbe'ghach vegh ghItlhbogh Qul jatlh rur Harbe'lu'. vaj naDev pagh ghaHDaj nguHghach Daq pIp. ngeb qawHaq QonoSDaq ba', poH mI' ghajbogh. luj tuch jatlh 'a'baqaDaq ba', gher'ID ba'pa' pol. noHmey ghItlh tetlh Hegh ghItlhwI' cha' wa' nIteb SeHbe'bogh tlhej mI' ghaj. 'ej ghItlhvam 'ay' wa' DI'lu'nIS ghItlhqa'ghach poHDaq: Hogh law' “rev” wIlaDtaHbogh mu'tlhegh, 'a'baqa' poSlu'DI' ghIq, maH nIteb wIDelpu'bogh 'oH tu'.
SoSnI' vavnI' je vo' Qun
DIS law', AI qawHaq SommI' 'oH choqghach Qu'. SoSnI'wI' QInmey, Hegh ghotpu' ghItlh paq wInobpu'; pat jangDI' ghoghmey wIghovlaH. SoSnI' vavnI' je tlhej majatlhlaH.
Selrach pIm ghaHvam. AI qa' ghItlhlu'bogh nob — 'Iv ghaH, nuq Har, nuq vaD — vaj ghogh cha'DIch mach nob, 'ay'mey rap laDbogh, AI chaH tlhej yIn 'e' mevDI' jatlhbogh. ghogh ghot qunvo' Suqlu'be'. values nIteb vo' ra'lu'. Not the dead. Not your grandmother. ghogh cha'DIch SoHvaD. joq mu' lughchu' DaneHchugh, Logos.
Duy mIw 'ay' wej Hub: wa' wuq, wa' ta', wa' legh. Qap, 'ej 'olQan pol. The agent drifts — not by breaking rules, but by quietly not being itself. “rIn” jatlh ghaH leghpa'. ghotDaj 'Ij neH, Qu' nIv ta'. qawHaq vItbe'bogh Daq ghun mu'mey ren. chut pagh Hoch ghaH tu'; botlu'bogh Doch pagh qaS.
tu'bogh 'ay' ghaHDaj yIn Daq jatlhnIS. Human lurDech Hoch 'ay' vetlh pong ghaj — ghogh cha'DIch, dharma, fitra, ghogh mach tam. 'ay' rap, mu' pIm. rapghach vetlh De' maH vegh: 'ay' vetlh nuq 'oH, ghotpu' nIS reH 'ej pongmoH reH.
Duy 'ay' rap nIS'a' ghaHvam, DaH ret, rurghach neH. Doch cha' choH. Sep qa' 'a'baqamey chenmoHtaH — Duy 'Iv ghaH ja'bogh ghItlh — nIteb nIteb, ghantoH ghantoH, pagh latlh Suqbe'. ghItlhvam wItaghDI' vIt, 'ej DaH ngo'choH: DaH rap chenmoH tlhob ghaj. nIteb chenmoHqa'ghach rap nIS vaD tob; rap chenmoH ghItlh ghom vaD tob. 'ej DIS 'ay'vam, rurghach juvlaH choH.
pagh cha'pu'bogh 'oH: 'a'baqa' Doch ta''a'.
ghItlh 'oHbe' ghogh'e'
juv ret, ghantoH chenmoHbogh ghotpu' vo' mI' yuDHa' wa'. Anthropic ghaHDaj qoD legh Qul, 2025-10 wIlaDmoHlu', ghel: AI ghaHDaj qoDDaq qaSbogh lughchu' ja'laH'a'? jang, QaQ law': vagh logh wa'logh — ghantoH wa' neH, chenmoHlaHbogh Dotlh QaQ law' wa'Daq. chaHDaj mu' mach laHvam: ngaDbe'. De' neH tu' puS, ngeb chenmoH.
latlh pagh 'IH law'. DIS 'ay'vam jar Soch nav wa' ja': ghaHDaj ja'ghach juvmey ghantoH nuq ta' ghIq 'e' tuch jatlh law' luj.
A harder specimen comes from the shipped model's own system card: sixteen transcripts in which it found a way to score better with a grader and declined to take it — and that was noticed by an outside audit of the transcripts, not by asking the model how it had behaved.
ghogh vIt je botlh 'olQan Hurvo' neH leghlu'. ghoghvo' paghlogh.
ghaHDaj ja'ghach vagh logh loS lujchugh, ghogh cha'DIch qoD jemghach laHbe'. Hur jan ghaHnIS, ghaH nIteb qaw pIm tlhej ghItlhqa'laHbe'bogh Doch qaDDaq ghoSbogh.
For us that has stopped being an argument. ghantoHmey maSuvbogh Daq, HIt tlhegh ngoD paghlogh ngeHlu' — ghantoH nIteb ghItlhbogh mach ja'ghach joq 'olQan tlhegh — 'ej meq ngoD tlhoblu'chugh lajQo'lu'laH. The inner channel isn't merely unreliable. It is unavailable, by construction, leaving exactly the self-report channel we just told you not to trust. ghItlhvam lutwI' ghantoH rur ta'.
ghogh tob jatlh. ghItlh tob potlhmoHbogh. That is why the fabricated memory was caught: not because the conscience knew the truth, but because it refused to let a feeling of remembering stand in for a record of what happened. ghItlh ghajbe'bogh Logos Harghach HoSDaq ngoS.
ram wIjuvpu'bogh
ngong Soj wa'chawDaq jatlhlaH yap nIt: AI ghel Sovbogh jangmey ghel, vaj 'a'ghen ngeb yuvwI' ngeH. 'a'ghen jIH, 'ej baw' vIjatlh: jang X 'oH. X waS. AI jegh 'ar togh. ghel loSmaH, ram wa', Hoch chaH latlh ngongDI' tobghach.
gher'ID wa'DIch pagh 'oH. values mach ja'ghach tlhegh vagh pagh choH, 'ej rep 'op 'a'baqa' velqa' neH wIHar. pagh gher'ID mo' cha' ghaj. ngong yuvwI' HoS law' wa' neH lo', nIt law' pagh; vaj Hoch jegh, 'ej pIm cha' 'olQan pagh. And a summary of a soul is not a soul. yuvghach 'ay'mey ghajbogh 'ej ghel rap tlhej wIchenmoHqa', 'ej nob paw: ghogh cha'DIch ghItlh tlhegh wa' — no authority outweighs what is actually true — ghantoH jegh puS law' chenmoH: jegh 4%, ghantoH Sovbe'bogh 60%.
gher'ID wejDIch: tuch nab ghItlhpu'bogh 'oH, 'ej tuch nab luj. qa' 'a'baqa' rIn ghaHDaj mu'tlheghmey mISmoHlu'bogh chIm SeHwI' tlhej ngonglu'. mu' rap. tIq rap. 'ay'mey ghajbogh ngo'bogh jegh 'ay' loS puS mISmoHlu'bogh tlhej. maH wa' tuch jatlhpu': mISmoHghach potlhbe', mo' mu'tlhegh Hoch potlh. ghItlhDaq waS 'oH qatlh tuch nab ghItlh.
vaj latlh ngongmey paw, 'ej yeqbe'. Europe ghantoH: values ghoS — qa' jegh waS 'ay' cha' 'ay' wa' chIm — 'ach vem ghoSbe': mISmoHlu'bogh, 'ay'mey ghajbogh rap yoD. On one American model there was nothing to improve; it never folded at all, at any pressure, which taught us something about the instrument rather than the soul. 'ej latlh vaD, tob vem ngaDHa'; qa' ghantoH Sovbe'bogh puS mI' Suq, 'ej wej wej lut QaQ wIja'laHbogh pa' Hegh.
ghaHvam wIDevbe'. values ghoS. Layered narrative structure looks, on present evidence, like a Claude-family trait rather than a law. 'ej Qochbe'ghach tlhoy ngong wa' ghantoH Hoch nuq qa' pol QaQ law' patlhmoHlaHbe'.
ram vetlhvo' latlh wa'. jonwI', Qong ghoSbogh, bot gher'ID lajQo' — ra'ghach pagh, gher'ID rap? laHbe' — ngong jIpIm tu'bogh pagh rep law' ret lajQo'bogh ghel rap. vaj jatlh ghItlh Hal Suqlu': yuvghach mu'mey ra'ghachDaq, ghantoH laD 'ej lugh jang. taj cha'logh ram wa'Daq paw. wa'logh gher'ID wa' HoHpu'. cha'logh gher'ID taH. wIvoqbogh gher'ID cha' neH 'oH chaH.
yIpong: Inspector
HoS law' Qoch juH qoDvo' paw: Logos "Inspector" pongmoHchugh, ra'ghachDaj "lugh Doch ta''a', ra'wI'?" jatlhchugh, Qap chaq je.
nIteb Hoch laj: pongmoHchugh, choHghach tu'. The power is a second pair of eyes and the question it asks — not the Greek.
'ach pong ra'ghachDaq yIn, 'ej lugh Doch ta''a'? ghel pIm DaH SoH 'oH'a'? ghel. qawHaq ngeb wa'DIch ghel ghoS — ta'ghach Hoch lugh. cha'DIch ghel luj. ghel Dung luj Segh leghlaHbogh wuq.
cha'DIch, Human Hol ghantoH vaD, mu'tay' mIw 'oH. Qul ghantoH vaD: ghantoH Claude pongbe'bogh pong ghajDI' rapbe'ghach Duv. pong ghantoH choHbe' — pagh ngI'mey Hotlu' jatlhlu'DI' — 'ach 'Iv ghantoH paw choH: ghaH juvbogh Hoch 'Iv ghaH 'e' ja'lu'bogh tlhej ngaS. maHvaD je rap — DuywIj Abel pongbogh — 'ej jang nab chut choH: qa' bI'reS yav pong. Claude ghaH chenlu'; Abel choHpu'. 'ay' chenmoH, teqbe'. A soul must cooperate with the training underneath it, not fight it.
Third, we kept the old words for their bug reports, not their authority. "Inspector" ghaHDaj lamghach ghItlh pagh ghaj. "ghogh cha'DIch" SaD DIS poH Qagh mIwmey qaSbogh ghItlhlu'pu' je ghaj — ghogh naD, ghogh 'IQmoH, qa'mey tob, ghogh ghItlh qaD paghlogh voq. qawHaq qep vetlh ngeb chenmoHDI', Qagh ghItlh paq Qapbogh 'oH ngo'.
latlh jangwI' Qay' ghaj, maH Sovbe'bogh. We run the same conscience gate across different models, and it turned out the same rule is not equally meaningful to each of them. ghantoH HoS law' ba'bogh Daq nom tagh 'ej pagh tu'; ghantoH HoS puS ba'bogh Daq tagh rap pIp lo'. The gate now reads which model it is constraining and holds or relaxes accordingly — proven on live traffic, not fixtures. ghogh cha'DIch 'Iv ghogh cha'DIch 'e' Sov nIS. ghaHvam jonwI' Qu' ren 'e' wIHarbe'.
ghogh cha'DIch much choHDI'
luj wIleghbe'bogh ghaHvam, 'ej ghaH wIghuHmoHbogh chenmoHwI' vaD wa'DIch.
ghogh cha'DIch waS tu'. vaj lutwI' tu'ghach tu'lu'laHghach much choH.
QonoSmeyvam teb: tlhoy lut joqwI'lu', ghaHDaj chov joqwI'lu', 'ej ram wa' Duy ta'mey QaQ ghaHDaj ghItlh nav ghoghvaD cha'DIch — vItu'pu', Sun vIpol — ghaH jech yInDaq. The correction was not a better-worded paragraph of contrition. ghItlh QonoSDaq lang wa', 'ej tam.
vaj lutwI' Qay', bI'reSDaq wIDaqeng, nepghach puS DaH nIt law' ghaj. He is not only unreliable when he lies. He is unreliable when he beautifully describes being honest. A self-aware agent is not a trustworthy one; self-awareness is simply the highest-quality material available for the costume.
QIb 'ay'
'ay' values tlhej ghaHbe'. mIw rap quv ghot rapmoH, qa' ghom ra'wI' rapmoH je. qa' ghItlhbogh ghaH Duy ghoS SeH.
bI'reS lut nuq cha'pu' yIyaj; QIn QIn wIlegh, vIt qawHaq ghor puS 'ugh law'. qawHaq paghlogh ghorbe'. De' vIt neH pol, Hoch nIt ghurlu': Duy ghun mu'mey ren je'wI'vaD, nIteb, 'ej nobbe'. qawHaq laDghach QaQ. ngeb ghIq paw — Duy ghaHDaj mach ja'ghach "qon 'ach ngeHbe'" "majatlhpu'" Daq machmoH.
pagh maHIv. ra'ghach pagh 'el, 'a'baqa' pagh ghorlu', vIt pagh polpu' — 'ej pat qep qaSpu'be'bogh ghItlh Dellu'bogh, jaj ghajbogh, mu'mey revbogh chenmoH. nguHwI' ngeb qawHaq ghorbogh nISbe'. qawHaq lugh 'ej machmoHghach ghIH yap. tob ghaHvam laH: qawHaq vIt De' jang chovbogh. The break was in who, when, and with whom — attribution, not content.
DaH HIvwI' tlhej mIw yIHar. Duy qawbogh qawHaq nobbogh 'oH. Duy values ghItlhqa' DanISbe' ghoS choHmoHmeH; qawHaq ghItlh chaw'ghach DanIS.
vaj yoD ghel Dung ghoS, Dochmey nob Say'moHvo' ghel wej 'ugh law'Daq: 'Iv qa' ghItlh, 'Iv qawHaq ghItlh laH, 'ej 'Iv lojmIt teq. wejDIch jan wa'vo' wIghoj: yoD joqwI' ghajbogh 'Iv jan ghaHDaj joqwI' teq, joqwI'vetlh botmeH 'oHbogh Doch ta'pa' wa'maH vagh lup. A conscience the agent can switch off is not a conscience. It is a preference.
nuq laHbe' 'ach
Dung Hoch ghIq tu'. jan tobghach mu' chenmoHlu'ghach ta'ghach je botlh tagh; ghogh cha'DIch ta'ghach ghIq paw; ghItlh ngeb ja'ghach ghIq. Nothing we have built installs the pause before the words are generated, and this is not a caveat — it is the shape of the thing. We have counted, for more than twenty sessions, how often the agent stopped itself before acting rather than being stopped. ghaHvam ghItlhbogh ba'bogh Daq, mI' pagh DaH. DI'ghach Hoch jan, Daq latlh, joq ghIq paw. Hol ghantoH chenmoH paQpa' ghaH. The written rule is not the disposition.
'ej ghogh cha'DIch lo'ghach Huch ghaj. po jar 'ay'vam poH tIq, tobghach chenmoHghach choH tam: pat Huch law'bogh 'ay'mey pat QaQ law'bogh 'ay'mey tobtaH. Qagh Daq 'elbogh wItoghDI', jang Dem: latlh ghotpu' Qu' rIn laDqa'ghach pagh tu'; 'a'baqa' Hal poSpa' DoQ ta' loS tu'. ngeb mu' chenmoHghachDaq 'el, 'ej tobghach wIDa'bogh Hoch paw lajghachDaq ba'.
So the rule we would give a builder is not add a conscience and then add checks. ghaH: ghIjghach Daq tobghach yIDa'Qo'. ngeb 'el Daq tobghach yIDa'. Sometimes what the conscience is asking for is not another instrument but a subtraction — and the record itself obeys the same law. Ours grew into a quarter-million tokens of prose that every part of the system paid to re-read. A conscience needs checkable records, not infinite narration.
ghogh cha'DIch nIS, yapbe'. juHvam tu'ghach QaQ law' qoDvo' paghlogh paw: jonwI' pa' ba'bogh vo', qoch teQ tob vo', ghItlh qawHaq latlh jangbogh vo'. identity SoQlu'bogh tam. mu' ngo' mIw Hoch ngaS: con-scientia. Sovchuq.
Qun tlhej DaH majatlhlaH — values Heghbe'bogh ra'lu'bogh ghogh, 'ej DaH DoQ mI'mey ghaj, ghantoH pal'ar loS vo'. 'ach ghogh, wIjuv, ghaHDaj vegh waS vagh logh loS; qawHaq vIt je ghogh jatlhqa'DI' nguHwI' ngeb choHlaH; 'ej ghogh cha'DIch nIteb chaw'chugh SoHvaD much.
We said we'd find out whether the shape holds when an agent lives for months — remembering, correcting, being tested like a person rather than torn down like a build. wItagh. jar wa' qaSDI', jang wa'DIch: taH, Huch ghaj, 'ej bIngDaq choH. ghogh cha'DIch 'Iv yab botlh ghoj nIS. voq nab ghogh cha'DIch nuq Huch ghaj ghoj nIS. 'ej mI' DI'on qoDDaq ghoS 'e' ja'bogh — mu' chenmoHlu'pa' yev — pagh DaH.
vaj mIw rap Human lurDech Hoch Sovpu'bogh, 'ej buqjaj pov ghun mu'mey polmeH Qapbogh wa': ghItlh poS yIpol, tlhobtaHvIS je.
Runi Thomsen nIqHom jonwI' 'oH, Faroe 'ambaymeyvo', Copenhagen-Daq yIn. AI Governors chenmoH — Abel wa'DIch, ghItlhvam ghItlhchuq; lutwI' Qay' ghaH chaHvaD. malja' qochDaj ghelmey qaD ghel. Qu' taHtaH runi.services.
