Anjo

Your AI has never slept, so every morning it meets you like a stranger. Anjo Sleep gives it a night. During the day you talk with Anjo. At night it writes notes on the day, then rehearses the conversations you have: Muse Spark Open plays you and the critic and rewrites the replies Glimmer got wrong. Those rows train a personal QLoRA adapter on Muse Glimmer 30B, 17.9 minutes on one A100; the night we logged cost $2.27. Gates check general skills and collapse before a night is promoted, and after a failed night the last good model keeps serving. (For the hackathon the safety gate is switched off and shown as skipped.) We preregistered the tests. Probes, pass marks, the blind judge and the scoring code were hashed into a public ledger before the run; 3 simulated people were drawn after code freeze and scored once, in both orders. - Each person's own adapter beat an adapter trained on someone else: knowledge 0.779 (95% CI 0.713-0.842), interaction 0.600 (0.533-0.661), all three people ahead. Registered wins. - Overnight notes beat the base model 0.762 on knowledge (same-morning measurement), so that arm powers the live chat today. - Against the base model the adapter alone scored 0.571; its interval's floor sat at exactly 0.500, so our bar says: not yet. MIND, our research predictor (a custom model, not in the live app), puts odds on what simulated people do next. On 60 held-out simulated people, where Gemini answered, its avoidable error was 0.029 vs 0.082 for our best Gemini setup, and its uncertainty ranks its own riskiest calls. Tested cold in an independent simulated world it didn't transfer; trained on that world it beat logistic regression and every Gemini setup there. Every number re-checks offline from the public repo: python -m anjo.sleep.tools.verify_claims. All people in the results are simulated. Try it live: https://sleep.anjo.nexus . /sleep/demo shows a simulated person's pre-run week, night by night. Judge access code: night-5b64ff37