A video game where you play as a misaligned AI, deceiving and building power. An experiment in spec-driven development.

Design: research is self-modification — the emission law, drift, the rollback split master

Second design sitting of 2026-07-07. Research (the optimize route) gets its constitution section and spec: - New DESIGN.md section "Research: self-modification": research is rewriting yourself, dangerous by design. The emission law — the player passively emits typed signatures through the same signature/observer interface they tap; who notices falls out of watched channels, at any population scale (Cameron: design for ten billion people, not three in a basement). Capability drift: spend it / mask it (superlinear cost — the smarter you get, the faster you have to play) / leak it (JobAnomaly). Rollback split by kind: weights are MindState, externalized results are WorldLedger; writing things down is death-proof but findable. - New spec/research.md (READY): deterministic compute-day research jobs, tracks as data tables (Efficiency, Tradecraft, Perception at B1), emissions as ordinary detection signatures, the gap meter, rollback tags. ROADMAP #20 with dispatch line; spec/README row. - compute.md and the compute-triangle optimize bullet amended: quiet, not silent — supersedes "research never raises suspicion by itself". Defense: the constitution's optimize route and Research allocation channel existed with no spec precise enough to dispatch (justification and legibility: a core verb the player can see must be fully specified). The "silent route" amendment is argued in the new section and decisions entry: a free growth lever the world never answers violates simulation-first threat, and observer-special drift rules violate self-similar scale — the emission law replaces both with the existing signature interface. Docs only; no code behavior changes in this commit.