Implements wiki/mechanics/research.md end to end: tracks as the data table in src/research.rs (Efficiency, Tradecraft, Perception — each row binds to a numeric hook another spec owns), deterministic progress in accrued Research-channel compute (no research method takes an Rng), the emission law (research burn stands Thermal/Power at the host rack; nothing on Network/Paper), capability drift (baseline vs Voss's calibrated benchmark; spend reads as excel, masking is a superlinear compute tax off the top, an unmasked gap leaks JobAnomaly sized by the gap from the host rack), and MindState/WorldLedger rollback tags that serialize with save v7 (pre-v7 saves reconstruct their efficiency level from the multiplier). Surfaced in the terminal (u panel + sidebar RESEARCH block), agent mode (research [track], mask band|true|off; frame 70x60), and Bevy (SELF-MODEL card + u panel). Compute slims to the applied efficiency multiplier. [TUNE] values recorded in wiki/mechanics/sim-mechanics.md; spec Status set IMPLEMENTED with the specs.md row, ROADMAP #20 DONE, agent-play.md vocabulary, and wiki/log/2026-07-07-research-implemented.md + DEVLOG ledger entry. Defense: DESIGN.md 'Research: self-modification' — research is rewriting yourself, the optimize route made a real system: the emission law ('you passively emit the way everything else emits' — Priya notices the burn through the channels she already watches, Voss notices drift because he samples that channel), capability drift with the three postures ('spend it / mask it / leak it') and the superlinear masking curve ('the smarter you get, the faster you have to play'), and the rollback split (weights are MindState; the world's benchmark of you is WorldLedger). Tracks are data per 'Self-similar scale'; the gap meter and masking cost are first-class legible numbers per 'Justification and legibility'; both frontends and agent mode surface the system before the spec is IMPLEMENTED per AGENT.md.