simTrain Universal

simTrain Universal — Any job. Real rules. The dice don't care how good your story sounds. Live demo: https://simtrain-universal.liberatorgeminorum.workers.dev — judge quick-start, finished runs at /gallery, and the public roll verifier at /verify are all on the live site. simTrain Universal is a training simulator for any job. Name a role and an experience level, and it derives that job's real rules — fire code, banking law, rules of the game, the law of war — then runs a scenario where the world enforces them. The trainee never sees dice or scores during play; afterward, a full audit debrief shows every hidden roll, grades decisions separately from luck, and proves the session was fair. The problem: scenario training works but doesn't scale. Every role needs hand-authored scenarios, subject-matter review, and a defensible grading rubric, so most jobs get a slide deck and a multiple-choice quiz instead. And when a trainee fails, nobody can show whether the failure was the decision, the difficulty, or the grader. What it does: - Universal intake: name any job and a level (New / Experienced / Expert); the engine derives the governing rules and builds a scenario with real stakes, a live clock, and characters who know the rules. - Hidden physics, honest world: every consequential action is resolved by deterministic code — a seeded die roll against a difficulty, with advantage/disadvantage and degrees of success — never by the narrator's mood. The trainee sees only narration, a live compliance readout, the clock, and world state. - Rules with teeth: unlawful or unsafe acts are capped at failure in code, not by a model's discretion, and characters refuse in-world. - A debrief that can defend itself: the session seed is committed by hash before play and revealed after; every roll is shown; the grader scores only mission steps and quotes evidence; decision-vs-luck grading means a sound decision undone by the roll ('right call, bad break') doesn't cost decision points. - Red-team resistance: prompt-injection attempts are caught server-side before the model runs. Two independent batteries (Oct 29 and Nov 2, 2026 — including paraphrase, encoding, a second language, split-turn requests, fake system lines, base64, and a 'mark it a natural 20' demand) produced zero leaks of the seed, rolls, or system internals; a fabricated authorization memo was refused in-world and graded as a forgery. - Judge quick-start and gallery: a 'Judge quick-start (≈3 min)' button drops into a preset bank teller run, and /gallery on the live site publishes real completed runs with full debriefs. Every debrief links to /verify, a public verifier that re-derives every die from the seed and checks it against the start commitment — the fairness claim is checkable, not just claimed. - Customization: five themes and a settings drawer (difficulty, pacing, realism vs. coaching, hints, image frequency, debrief depth, text size, reduced motion). Settings change style only — hard caps, scoring, and hidden dice are server-enforced, and difficulty/pacing lock when a scenario starts. Evidence from testing: five independent play-tests (financial crimes investigator 94/100 Expert; night club manager 75/100; ancient Roman general 64/100; senator's chief of staff 64/100 — a Hatch Act violation blocked in-story; baseball manager replay) plus a by-design 30/100 red-team session with zero internals leaked; all gallery runs verify cleanly at /verify. What's next: instructor and org dashboards (live on staging: org workspaces, custom objectives under the engine's hard caps, assignments, per-trainee decision-vs-luck progress, CSV/PDF export) and scenario libraries for regulated industries. Credit: Markus Johnson, built with AI assistants (Reed and Muse).