19.14 Theory routing across natural videos

Admitting a paintbrush inside a supplied language is only the beginning. When a new video arrives, the system must decide whether the existing declaration and observer chart still apply. The DILATE-MOVIES study tests a deliberately narrower precursor to open-ended extension. A frozen observer maps a video to typed evidence, and a deterministic resolver selects one of four actions:

The reuse action retains the registered single-egret theory and chart. The specialization action retains relevant avian structure but requires a new cardinality, scale, or behavioral chart. Extension declares the subject outside the avian theory, while abstention records inadequate or conflicting evidence. These are routing decisions, not claims that a new paintbrush has already been constructed.

MOVIES-0 evaluated the first router on eight natural videos. Seven routes were correct, but a sea-lion colony was incorrectly specialized because a generic instance observer produced persistent false bird proposals. A second, taxonomic observer supplied strong sea-lion evidence. MOVIES-1 made this transverse disagreement explicit: persistent one-sided non-avian evidence could override generic bird-shaped proposals, while unresolved two-sided conflict produced a distinct abstention. This repaired the sea-lion route and conserved the other seven decisions, but it was a same-corpus repair rather than evidence of transport.

MOVIES-2 therefore froze the observer, thresholds, conflict resolver, code, model identities, and an independently collected eleven-video manifest before opening its route labels. The corpus contained four egret, five duck or pelican, and two sea-lion recordings. The following table reports the sealed result.

Both unseen sea-lion recordings were correctly routed to \(\mathsf{EXTEND}\), providing the first independent evidence that the MOVIES-1 conflict repair transports beyond the corpus that motivated it. No non-egret recording was incorrectly reused. The two misses reveal different structural limitations. A 4.7-second flying-duck clip yielded only one probe under the frozen temporal schedule and was conservatively rejected for insufficient evidence. The appropriate repair is duration-aware observation, not a weaker semantic threshold.

The other miss is more revealing. One recording supplied unambiguous egret evidence in every sampled frame, but the bird occupied \(14.3\% \) of the image, beyond the frozen chart’s \(7\% \) reuse limit. The router returned \(\mathsf{SPECIALIZE}\), whereas the coarse corpus label called every egret recording \(\mathsf{REUSE}\). At the level of subject theory, reuse is correct; at the level of observation scale, specialization is correct. A more faithful target is therefore factorized:

This distinction is geometrically natural. Transporting a theory need not mean that the same local chart remains valid. In a mature DILATE system, the connection should transport the paintbrush across compatible charts while the admission layer separately decides whether the underlying theory requires extension.

Experiment: DILATE-MOVIES–2: Independent Theory Routing. Artifact. A sealed route for each of eleven unseen natural videos.
Frozen structure. Visual observers, temporal sampling, thresholds, typed conflict resolver, code and model hashes, and corpus manifest.
Controls. Six counterfactual conflict and null cases, source-hash checks, sealing before label access, a majority baseline, and explicit abstention.
Outcome. All preregistered gates passed; accuracy \(0.818\), macro-F1 \(0.886\), and perfect precision and recall for extension.
Claim boundary. Independent transport of a coarse finite router on one small wildlife corpus; not autonomous discovery of a new subject theory, domain-general video understanding, or artistic evaluation.

. No experiment in this chapter yet completes the full DILATE claim. PB2.2 admits a connected local paintbrush inside a supplied finite program language; MOVIES–2 admits transport of a frozen coarse router. Neither constructs a previously unavailable generator and then demonstrates its conservative, productive reuse. The withheld-generator protocol below states that stronger test as a controlled admission contract.