Reality-Bounded Improvement: Why Imagined Progress Outruns Real Competence, and the Grounding-Efficiency Engine That Spends Reality Where It Moves the Frontier

Imagine-then-act world-model loops improve fast in imagination, but their real
frontier lift is bounded by independent contact with reality. This paper
introduces the evidence frontier, the best real competence supported by the
current independent real-contact ledger; Reality-Bounded Improvement (RBI),
the claim that frontier lift is bounded by new independent real-contact
information; grounding efficiency eta; Value of Real Contact (VoRC); and the
Grounding-Efficiency Engine (GEE), which allocates scarce real contacts to
decision-relevant residuals. The central correction is explicit: zero new real
contact does not mean zero learning, because systems can catch up to old
evidence. The sharp claim is that zero new independent contact cannot reliably
raise the evidence-supported frontier. A constructed witness over 150 seeds
shows imagined score rising without frontier lift, and shows GEE outperforming
random, coverage, risk-only, and relevance-only contact allocation at equal
budget. The package also includes frozen cloud return packages for V-JEPA2 and
OpenVLA initial-observation proxy scoring plus a UnifoLM full-rollout payload
with V-JEPA2 feature scoring. These cloud returns are bounded artifacts:
the strongest current tier is OPEN_MODEL_RSHADOW_FULL_ROLLOUT_CANDIDATE,
not open_model_rshadow_completed=true. No real robot, third-party
replication, live deployment, production claim, or unconditional GEE optimality
claim is made. It also does not claim priority over active learning, value of
information, experimental design, or submodular greedy approximation.