THE CARRIER NOTEBOOK
A connection needs a place to arrive
The preceding Carrier model could preserve most relationships and still stumble at a crucial moment: when the field needed a genuinely new connection. Its aggregate score was high. The weak subtask deserved its own experiment.
Several good candidates do not make one good choice
The earlier addition head judged possible parents separately. Several could receive a positive vote. The new specialist makes them compete in one shared decision and permits no choice when none is eligible. That does not give it a philosophical preference. It makes an authored numerical rule easier to express and inspect.
On the new ordinary-field test, actual birth-link F1 moves from 0.5588 to 0.9420. The result concerns the right return from an appended cell. It should not be expanded into a claim that the model invents every kind of meaningful relationship.
Precision needs the right examples
The previous network failed near two exact distance thresholds. Broad success had not taught it those boundaries precisely enough. We supplied focused equality and near-boundary examples and fitted compact numerical classifiers. Boundary-field F1 now reaches 1.0000, against 0.5700 for the previous specialist on those same fields.
This was deliberate distillation of a known procedure. A deterministic implementation of the original teacher could already follow its own rules exactly. The contribution here is a better integrated learned specialist and an inspectable instrument, not proof that machine learning was necessary to calculate distance.
A parent is not the whole birth
The first development run exposed another issue: choosing an eligible parent did not establish that the proposed placement could exist. That run was retained. The revised model distinguishes eligibility, choice, feasibility and actual emission. A connection has to arrive somewhere; one successful component cannot stand in for the whole event.
Existing-cell geometry and colours and the membranes still come from the old frozen model; the new count veto and newborn parent-colour copy are separate interventions. Excellent local link decisions can therefore coexist with changing colours, geometry or counts. Four-step tests make that distinction visible by comparing both the next reply and the path accumulated from the initial field.
Continuity needs an honest correspondence
Our first path comparison mixed moving-cell renumbering with real drift. We retained that result and added a post-test diagnostic that tracks the same original cells. After four replies, their relationship score remains 0.9963, while their colour agreement with the reference path is 61.7%. Correcting the measurement matters as much as improving the model: we should not call a change of address a loss of memory, or call a surviving line an unchanged meaning.
The precise mark and the open reading
Two new compositions differ by only two coordinate units. Their first pair lies on opposite sides of the taught threshold. A model can now respond more precisely to that distinction. It still cannot establish that two nearby philosophical positions agree, or that a change from tension to resonance is warranted by their meaning.
That is a productive boundary for this art form. Carrier gives us exact identities, spatial arrangements and typed relations to work with. The numerical model can become more dependable in those mechanics. The viewer keeps the right to question the resulting composition.
A separate reader of one source/reply pair offered two readings: a conversation becoming more attentive to maintenance, or a conversation narrowing into maintenance at the expense of its original plurality. Twelve distinct anchors became eight. Both readings remain available beside the exact field. Better mechanics should create room for this disagreement.
The instrument places the raw proposal beside explicit reader protections and the evidence beside the work. More capability should make a reply more inspectable, not require the reader to surrender judgment.