JOB135: WHAT THE FULL-FIELD MODEL DID AND DID NOT LEARN

Scope and representation

The released model is a 1,121,293-parameter numeric field-response network. It
receives allocated colour identities, deliberate NULL, cell x/y/z/a, six typed
relations with endpoints and direction, and region x/y/r/p. It produces these
channels plus cell, region and edge counts. Colour is a categorical ID obtained
from an exact allocated hex lookup. The network does not receive RGB vectors,
rendered pixels, dictionary definitions, English prompts or hidden verbal targets.
NULL6257 is distinct from PAD6258. This is structured numeric learning, not a
claim that the model sees colour, understands every Carrier concept, or has a mind.

Administrative id, parent and address strings do not enter the network. Addresses
bind edges before deterministic canonicalization. Symmetric relations0/1 share
an unordered identity; directed2-5 retain direction. Canonicalization makes the
tested address, ordering and symmetric-orientation invariances implementation
properties, not discoveries independently learned by the neural weights.

The teacher consists of public authored numerical exercises. Its response rules
are not the official meanings of the relations and are not evidence of conceptual
truth or semantic inference. Half of base scenes may use inherited graph walks.
Thus random neural initialization is not a knowledge-free cultural starting point.

Training and the declared revision

Training used12000 base scenes,600 development scenes, seed135001 and2400
updates of16 examples per condition. Each run saw38400 sampled examples. V2 full
selected update2400; the colour-multiset control selected2000 after completing
the same2400-update budget. Matching full/control runs had exactly identical
initial tensors. One seed per condition provides no seed-generalization estimate.

The original absolute-count v1 models are preserved. Exploratory art development
showed8/10 generated fields with no links and2/10 with one link. Before the main
600-example benchmark, v2 replaced count heads with count deltas relative to
observed input counts and an initial zero-delta bias of2. This engineered copy
prior is explicit. Count accuracy must not be presented as counting learned from
scratch. V1 automatic intervention jobs crossed a stop message: they read the
frozen challenge fixtures, but their metrics were not inspected for the revision.
The records are local author-maintained chronology, not externally audited
blinding or public preregistration. Development artworks are development evidence.

Held-out numerical results

The fixed600-example benchmark shows selective learning and material failures.
No further model tuning followed these outcomes.

* V2 full membrane-opening accuracy:0.999254; changed retained opening accuracy:
  1.000. COPY opening accuracy:0.577181. This is a strong result for this teacher.
* Full node/region/edge count exact accuracy:0.926667/0.985000/0.861667. COPY and
  the initial copy-prior model score0.803333/0.865000/0.760000. The reference
  prior explains much of the starting accuracy; the differences concern learned
  conditional edits under the authored exercise distribution.
* Full typed-edge micro F1:0.547357, substantially below COPY:0.905596.
  Full persistent-edge recall is0.550151. Receiving the graph and producing
  graph outputs therefore did not produce a successful general graph-response
  solver. The failure is retained in the reported result.
* Full colour accuracy is0.699720 versus COPY0.687788. On3679 retained cells whose
  target colour changes, full accuracy is only0.163360. The initial model scores
  0.113618 and the colour-multiset control0.114433 on that changed subset. This
  is weak learned colour transformation, despite reasonable aggregate accuracy.
* Full raw x/y MAE over all target slots is6.966679 on the0-1000 scale. COPY raw
  MAE on its supported retained slots is7.350752; these denominators differ.
  COPY all-target MAE becomes16.744651 when its missing birth slots receive the
  declared1000-unit penalty. Much of that larger comparison reflects coverage
  of births; it must not be described as a similarly large movement improvement.
* The trained colour-multiset control has x/y MAE239.205917, colour accuracy
  0.159809 and typed-edge F1 0.003086. It removes structural information and
  spatial slot binding, and is not an equally informed prose/JSON control.

Matched interventions

There are240 substantive pairs across20 intervention channels plus48 invariant
pairs. Full and the initial model both change at least one raw head on240/240
substantive pairs. The colour-multiset control changes12/240, corresponding to
NULL changes. All three satisfy48/48 canonicalization invariances. Initial-model
sensitivity is a direct counterexample to interpreting nonzero sensitivity alone
as successful learned use of a channel.

Mean matched teacher x/y-response-delta error is0.903678 for full,1.907517 for
initial and1.562463 for the colour-multiset control. This aggregate does not show
uniform success: full is worse than initial on7/20 individual intervention
channels, including several relation-type changes and reversal of relation5.
Teacher correctness and per-channel response errors remain beside sensitivity
counts in the detailed reports. Some interventions change non-x/y outputs, so
the x/y delta metric alone cannot establish correctness of their whole response.

Alignment, raw heads and validity

Teacher targets retain source-slot identity after movement; births append and
deletions remove a suffix. Targets and predictions are not re-sorted by moved
coordinates. Raw-head metrics cover teacher-active slots even when a predicted
count does not emit them; emitted-field metrics and birth coverage are separate.
COPY supplies no invented values for absent source slots.

The decoder rounds and clips numerical outputs, repairs centre spacing, and
selects legal typed edges using predicted counts. Valid fields after this adapter
are not evidence of learned grammar validity. Excluding routine rounding, full
needed111 repairs across15.6667% of benchmark scenes:74 clips and37 centre-spacing
repairs. There were no nonfinite replacements or edge-capacity repairs.

An independent edge-score audit checked loss relation indices, source/target
alignment, directed orientation and symmetric averaging. A synthetic oracle
covering all six relations decodes to exactly the target edge set and scores
F1=1. No scoring or direction bug was found to explain away COPY's advantage.
The model, teacher, decoder and benchmark outcomes were not changed by the audit.

What remains untested

The result is an inspectable full-field instrument with selective numerical
learning, weak changed-colour performance and a serious graph-preservation
failure. It does not establish general semantic competence, artistic intention,
subjective experience, superiority of this encoding, or independent visual
understanding. Browser/Python parity establishes implementation agreement only.

A future study could predeclare graph-edit and changed-colour endpoints, compare
equally informed structured and textual controls, examine count/reference priors,
and use multiple training seeds. Any new architecture or training choice needs a
new untouched benchmark and newly frozen interventions; the present600 examples,
matched challenges and development artworks must not be reused as unseen evidence.

Evidence: test-results.json, per-example.tsv, intervention-summary-v2.json,
interventions-*-v2.json, ../setup/edge-score-audit.json,
../setup/training-control-checks.json, and ../protocol/pretest-count-revision.txt.
