lin-0194
15.7 Six layers of control
GIRL separates six obligations:
Advantage quotient: remove action-common variation.
Phase coherence: check that validation and deployment directions align.
Confidence: require a positive one-sided utility bound.
Successor sampling: estimate the Bellman branch without privileged population operators.
ESS adaptation: select an off-policy estimator from replay support.
Temporal control: quarantine stale experience and govern re-entry.
For normalized validation and deployment directions,
\[ \kappa = \langle d_{\mathrm{val}},d_{\mathrm{dep}}\rangle \]
is the phase-coherence gate. At high effective sample size, ordinary importance sampling can be efficient; under severe mismatch, clipped self-normalization can be safer.
