PRE-TEST DEVELOPMENT REVISION / 2026-10-09
The first fixed-budget full-field model and matched colour-multiset baseline are
preserved under runs/full and runs/colours-only. They are not silently replaced.
Development loss included poor absolute count prediction. Exploratory artistic
trajectories made that failure legible: eight of ten generated fields had zero
links and two had only one, despite linked input fields. These artworks therefore
became additional development evidence, not independent held-out art evaluation.

Before the600-example reserved benchmark was evaluated, the next design was fixed:
keep every other architecture component, numeric teacher, training/development
data, seed, loss and2400-update budget; replace absolute count heads by learned
count edits relative to the observed input count. Head sizes129,17,257 cover all
differences for nodes(max64),regions(max8),edges(max128). Absolute logit at k is
delta_logit[k-observed_count+maximum]. All legal output counts remain reachable.
Observed counts derive from present slots and canonical semantic relation tensors.
The colour-multiset control observes node count only; its region/edge references
are zero because those inputs are erased. Initialize delta head weights N(0,.002),
biases N(0,.01), with zero-edit bias set to exactly2: a declared engineered copy prior, not knowledge
learned from scratch. All other neural weights retain random initialization.
New matched runs live under full-v2 and colours-only-v2; v1 is retained.

Chronology disclosure: automatic v1 intervention runs started while a stop message
crossed the model agent's completion message. They read the pre-frozen intervention
fixtures, but their metrics were not inspected by the designing agent or root before
the revision. Those outputs are retained. The reserved600-example benchmark had not
been opened for model evaluation. This is local author-maintained evidence; it is
not externally audited blinding or a public preregistration.

The revision is development-driven. Final holdout performance will not drive any
further model selection, tuning, retraining or suppressed failure. Final UI will
identify the copy prior and report count/edge losses even if the revision is weak.
