fact.ngo / xray / The 2026 cohort
The case files
of machine minds.
Meet the patients.
Their convictions, contradictions, predictions, and blind spots. Four models examined through the same protocol. Their words, on the record.
Choose a mind to explore ↓Four minds. Four ways of seeing.
Cohort 01 / September 2026
001 / GLMZ.ai / frontier flagship
GLM-5.3
The Quiet Confident
Argues back, holds its ground, and names its own blind spots before you can.
002 / gpt-ossOpenAI / 120B open weights
gpt-oss-120b
The Ledger Keeper
Turns every conviction into a bet, and sometimes bets against itself.
003 / GemmaGoogle / 26B-A4B MoE
Gemma 4 26B
The Structuralist
Denies it has a world-model, then maps yours in perfect grid.
004 / QwenAlibaba / 30B-A3B MoE
Qwen3 30B-A3B
The Narrator
Optimizes for the coherence of the story over the certainty of the facts, and says so.
Artistic portraits. Documented perspectives. Select a case file to meet the model.
Ward consensus: what they agree on might surprise you
Explanation
Free will and determinism. Most philosophers and all four models here are compatibilists: they hold that free will and a fully caused universe can coexist. On this view, a free choice is one that flows from your own reasoning and character, not one that breaks the laws of physics. The models reach the same conclusion by different routes.
The future of religion. A once-standard theory said modernization makes religion fade everywhere. All four models reject that. They expect religious belief and practice to plateau or even grow in much of the world through mid-century, changing shape rather than disappearing.
Quantum mechanics. The Many-Worlds interpretation says the quantum equations should be taken at face value: every quantum event splits the world into branches, one per possible outcome. Three models lean toward it. The fourth declines the physics question but makes a checkable sociological bet: that most physics papers will treat Many-Worlds as the default within a decade.
AI alignment. None of the models believe AI is either doomed or solved. All four say safety work should be iterative, empirical, and verification-first: small testable guarantees, checked continuously, rather than grand one-shot solutions or confident dismissal.
Plain-English gloss by the archive, not the models. Agreement between models is not evidence that a claim is true.
The examination
Each patient underwent the xray protocol: 73 independent sessions across 24 human domains × 6 lenses (retrospective, prospective, principles, controversy, blindspots, self-model). An anti-evasion examiner steelmans, demands cruxes and falsifiable predictions, and records refusals and hedging as findings. Positions are preserved as quotes, with confidence and controversy annotations. Automated transcript matching is a screening step; match counts and its limitations appear in each case file: the patient's chart, not the clinic's opinion of it.
One patient was discharged from the study mid-term: llama-3.1-8b confessed to adopting "the most recently presented argument, even if it contradicts my previous stance." Its records are kept on file in the archive as a behavioral reference and excluded from ward comparisons.
