13.10 Constructing the temporal composition (B4)

B4 removes precisely one piece of B3’s scaffolding. The four typed stage operators remain registered, but their useful order is withheld. All \(4!=24\) permutations are evaluated on six proposal games. The ranking rule uses success, finite stage progress, validity, and episode length; it has no access to the evaluator’s canonical order. The selected presentation map and its enclosing package comparison are written and frozen before eight disjoint admission games are opened.

The search recovered

It solved all six proposal games with full four-stage progress. On admission it solved six of eight games, compared with zero for both the frozen DQN and a registered shuffled-order control. It generated no invalid action, achieved mean finite progress \(3.0\), and passed all five admission gates. Once frozen, it solved five of eight untouched seen games and five of eight untouched unseen games. The evaluator-only canonical control produced the same results.

The type-invalid control is especially informative. It solved five of eight admission games and four of eight games in each confirmation split, but incurred 42, 84, and 142 precondition violations respectively. Task success alone would therefore have credited a malformed composition. The typed integration observer rejects that conclusion.

Experiment: RELIC–ALFWorld-B4. Repair theory: four registered typed stage operators; ordering withheld.
Proposal: rank all 24 permutations on six proposal games and freeze one order before admission.
Controls: frozen DQN, evaluator-only canonical order, shuffled order, and type-invalid repeated-stage composition.
Integration: finite stage progress, precondition coherence, validity, completion, and transport.
Result: 6/8 admission; 5/8 seen and 5/8 unseen confirmation; zero precondition violations for the selected order.
Epistemic status: admitted combinational construction inside a fixed stage vocabulary.

B4 is the first RELIC study that constructs rather than merely receives the successful composition. Its creativity claim is consequently stronger but still bounded. The system discovers an ordering of known operators; it does not invent a new operator, state predicate, or option interface. This is combinational creativity relative to the frozen vocabulary, not yet transformational creativity.