Theoretical Cartography 5: Leave-One-Out Validation—Preregistered Replay Adjudication

The first four volumes of theoretical cartography respectively delivered a weighted hypergraph foundation, a homological detector, a three-tier criterion, and a remote generator. Withthe organs complete, a meta-level question emerges: whether this machine’s behavior is trustworthy on the very cases it was designed to handle. This paper responds with leave-one-outvalidation: for each registered bridge, perform an excision operation, construct a parallel mapin which that bridge has never been discovered, replay the full pipeline on the parallel map, andadjudicate the machine’s behavior item by item according to preregistered clauses locked beforereplay begins. The preregistered clauses cover six items: group-excision alarm, single-unit excision discrimination, recovery silence and redundancy discrimination, criterion pass rate, zeroregistration of negative controls, and adjudication consistency. Toy-scale readings show: aftergroup excision the number of connected components is two, after recovery it returns to one, andafter excising redundant edges it remains one; for the twelve registered correspondences, aftereach single-unit excision the alternative paths all exist, and all detector silences are classified ascorrect silences; the criterion pass rate is twelve out of twelve; six random seeds of structurallyscrambled pseudo-galaxies, after cascaded filtering, have a total registration count of zero; thereplay of Bridge Three again refuses registration. The generator revival ledger of thirty, twentyfour, twenty, and twelve agrees number by number with the prior ledger. The disposition statesof the four bridges in the blind test set must be recorded separately under the undeterminedledger and the passing ledger: Bridge Two falls under Outcome Five, out-of-jurisdiction disposition; Bridge Three falls under Outcome Two, consistency veto; Bridges One and Four areconditional candidates in the undetermined ledger; the failure ledger is zero. The above readings hold only at a finite toy scale, under a given structural specification, a given manual auditprocedure, and a self-reported-strength protocol; the revival and blindness branches of singleunit load bridges cannot be separately tested in the toy model, the six-seed randomization ofnegative controls does not constitute a zero-registration guarantee against stronger adversarialconstructions, and selection bias arises from the common source of model structure and caseprovenance and cannot be eliminated, only declared. The temporal order of the protocol iscarried by the locking record of the same document, with no external anchor; before an anchoris connected, the force of all clauses is self-reported strength, and the legitimate weight of theadjudication in external audit is correspondingly limited.