CHROMATIC HABITAT — A FIELD THAT CAN ANSWER / JOB135
Prospective local plan, 2026-10-09. Freeze manifest records exact files and time.

Question
Can a randomly initialized small neural model receive and learn authored response
rules over every content channel of carrier-habitat/1, then participate in a full
field exchange? This tests structured numeric learning and an artistic instrument.
It does not test qualia, consciousness, general intelligence, conceptual truth,
human experience or Carrier superiority over an equivalent encoding.

Inputs
6257 pinned allocated colour identities plus deliberate NULL; NULL is not PAD.
Cells x,y,z,a; six typed cell-cell relations with their actual endpoints and
direction; circular regions x,y,r,p; all counts. Administrative id/parent/address
strings never enter the neural model. Addresses first bind endpoints then become
canonical slot indices. Symmetric reverse duplicates collapse visibly to one edge.
The existing grammar is fully received. Relations do not express causation,
entailment, authority or mandatory force. Openings are visual marks, not physics.

Training source and task
Authored numeric exercises are the sole supervisory targets (make_exercises.py).
Half of base scenes may draw colour walks from the retained Carrier graph. This
is inherited human/AI-authored structure, not a knowledge-free starting point.
Dictionary words/definitions, artwork readings and prior trained weights are not
read by the trainer. RGB identities are categorical IDs, not perceptual RGB vectors.
The task is one synchronous numeric response to a field; target slots retain
source-slot alignment, births append and deletion is limited to an unlinked final
faint NULL. Target fields are not re-sorted after movement. Position reference
scaffolds, count limits and legality adaptation are engineered, not learned.

Splits and models
13200 unique base scene IDs shuffled before generation:12000 training,600 development,
600 reserved test. Entire base belongs to one partition. Intervention scenes use
separate IDs. Shared generator and authored teacher mean held-out procedural
examples, not independent real-world knowledge or unseen grammar.
Full model and colour-multiset ablation have identical architecture, seed135001,
2400 optimizer updates,batch16,AdamW config.json. Earliest lowest development
weighted loss checkpoint selected at update1 and every400 updates. One training
seed per condition; no seed-generalization estimate. Colour-multiset sorts colours
and removes binding, geometry, emphasis, levels, edges and regions. It is a severe
information ablation, not an equally informed prose/JSON control.
Both checkpoints fixed before test evaluation. No held-out metric may drive tuning.
Initial model and COPY source baseline also reported. Correctness/implementation
bugs can be fixed with original failure preserved; no selective failure deletion.

Measurements
Primary profile: changed-colour accuracy, all-target x/y MAE (missing COPY slots
penalized1000), typed edge F1, membrane opening accuracy and count accuracy.
Report all categorical heads, region geometry, counts, births and changed-only
subsets. Avoid one aggregate intelligence score. Paired bootstrap2000,seed135002,
for per-example xy and edge differences, descriptive under this generator and
single seed. Routine coordinate rounding separate from clipping/spacing repairs.
The decoder chooses counts/categories/edges from neural heads, then explicit
deterministic grammar repairs. Grammar validity after repairs is not model accuracy.

Interventions and invariants
Pre-training frozen matched scenes alter a single channel with colours held fixed
where applicable. Test x/y/z/a, NULL, edge presence/type/direction, regionx/y/r/p.
Report teacher response alongside model response; nonzero sensitivity alone is
not successful use. Renaming, cell-array reordering with bindings remapped and
symmetric edge reversal should be invariant. Directed reversal should be visible.

Art and browser
Separate transfer artwork fields and fresh reader replies are excluded from numeric
training. Their interpretations are contestable readings by pretrained language
models, not thoughts decoded from this small model. Full-field returns are produced
by learned inference, not by the numeric teacher. Colour definitions may be revealed
after the field is made; they are interpreter/display material only.
Portable JavaScript inference must match Python/NumPy numerically and greedily.
Static application checks are distinguished from physical-browser/phone/live-site
testing. No claim of live publication until user uploads.

Decision criterion
A successful engineering result receives every grammar channel, produces editable
valid full fields with disclosed repairs and demonstrates selective learned response
on the stated tests. Weak learning of any head remains a visible failure. No need
to claim completion of mind, autonomous life or artistic quality to release an
inspectable instrument. Job134 remains an unchanged historical baseline.
