THE CARRIER NOTEBOOK
A line can survive while its meaning changes
The earlier model gave us an instructive result. It preserved relationships very well, yet after four replies its original cells agreed with the independent reference colours only 61.7% of the time. A strong connection score had concealed a different weakness.
That mattered because a colour in Carrier identifies a selected concept. Keeping a line while changing the concepts at its ends can transform the composition. We needed to follow the whole field through time.
The answer was to begin from what was there
The new instrument starts from the source field. Instead of repeatedly estimating a colour’s identity, its program can retain the source value or select a particular source cell to copy. The resulting trace says which occurrence supplied the colour. Geometry, emphasis, layers, membranes and growth are updated in the same inspectable calculation.
This is a fitted numerical program with a deliberately restricted operator library. We taught it using the known reference procedure’s intermediate operations and exact boundary examples. We also supplied much of the execution structure. That is a significant engineering choice: the improvement comes partly from making the task explicit, rather than asking a broad predictor to approximate every part of it.
Preservation has to permit an answer
Copying everything would be dependable in one sense and incapable in another. A reply sometimes has to move, recolour or grow under the chosen conventions. We therefore measure colours that should stay separately from colours that should change.
On the new ordinary test, the fitted program is correct on 100.0% of retained-colour targets and 100.0% of changed-colour targets. It matches complete fields in 100.0% of cases. Its birth-link counts are 382 correct, 0 extra and 0 missed. Every figure concerns the authored numerical procedure, with optional reader protection turned off.
A path is a stronger demand than one step
We also let each reply become the next input. The reference evolved independently. Original cells kept external evaluation identities through every reordering; those identities were never supplied to the model. At turn eight, original-cell colour agreement is 100.0%, compared with 63.1% for the preceding model on the same starts.
The old 61.7% figure belongs to its own published experiment. The new comparison reruns the old model on fresh common inputs. Keeping that distinction prevents an apparent improvement from being manufactured by a change of test set.
Exact mechanics do not settle the artwork
One of the new studies contains an opening and a deliberately empty position. The numerical rule can fill that position from a nearby source colour. The artist may still say: an opening does not demand an occupant.
Both observations can stand. The reply can follow its rule exactly and still be a reply someone declines. In the grammar, an invitation is not a command to become the inviting concept. The program’s copying convention is an artistic mechanism we can revise; it is not a discovery that invitation means assimilation.
The interface therefore offers a separate choice: protect this cell’s colour, with or without protecting its links. Any restoration is labelled as software policy. The model proposal before reader protection remains available, and accepting it is a deliberate action. A better calculation does not receive more authority over the reader.
The reader can question an exact reply
A separate pretrained interpreter read one source and its first and fourth replies without the author’s interpretation or the numerical results. It observed twelve distinct colours becoming nine, then seven, as the larger membrane expanded. Its first reading was that the room widened while the vocabulary narrowed. Its alternative treated the repetition as commitments distributed across different places.
That disagreement matters. Exact reproduction of the numerical path did not prevent a reader from questioning what the path did. The complete earlier field makes the replaced distinctions available for inspection; the later field does not silently retain every concept that disappeared from it.
A smaller claim makes a stronger instrument
The observed results are strong, but a direct implementation of the known teacher can also reproduce the teacher. This experiment does not demonstrate that learning is necessary, that the rules are ideal, or that the program understands their philosophical names.
What we have gained is a more faithful and traceable way to operate the medium. Twenty-four recorded fields let readers encounter the result, while all predictions and checks let others examine the numerical claim. The next challenge can now be chosen more precisely: when do creators actually want these transformations, and how can their own constraints be expressed without being lost in a chain of replies?