JOB138 — CHROMATIC INHERITANCE

Keep the extracted record's sibling folders134,135,136,137,138 together.
All commands below run from their shared parent. Inference needs Python3.12
and NumPy, or the supplied dependency-free JavaScript. No API or network.

OFFLINE REPLY
  python 138/tools/inheritance_reply.py input.json reply.json
The supplied input bytes become the exact SHA256 parent of the reply.
Default seed138001 was selected in the plan before final generation.
Existing source addresses are preserved through internal canonical sorting;
a birth receives an unused valid nNN address. These administrative names are
not learned meaning and are not supplied to numerical fitting.

JAVASCRIPT
Load model/inheritance-model.js then common/inheritance-core.js, then call
INHERITANCE_CORE.run(field,palette,INHERITANCE_WEIGHTS). The standalone core
returns field, birthChoice, repairs, canonicalization and diagnostic colour
source choices/numeric edits. Both files also export CommonJS modules.
The optional interactive policy wrapper is not used by scientific scoring.

PARITY
  python 138/tools/make_inheritance_parity.py
  node 138/tools/check_inheritance_parity.cjs
590 retained development fields compare complete Python/JavaScript outputs,
source-copy diagnostics, numeric changes, birth choices and spacing repairs.
The first JavaScript deep-equality failure (-0 versus0) is retained separately.

REPEAT SCIENTIFIC EVALUATION
  python 138/tools/evaluate_inheritance.py --prefix repeat-final
This verifies frozen code/weights and reuses public retained final inputs.
It refuses existing result/archive prefixes. A repetition is reproducibility,
not a fresh untouched holdout. Frozen Job137 comparison needs PyTorch and
135/runs/full-v2/weights.npz; the new model itself does not need PyTorch.
All final fields/predictions and every eight-turn identity transition are
retained. No optional guard is enabled. Independent scoring fixtures live in
tools/test_inheritance_metrics.py; the audit commands are documented in their
command-line help and analysis audit reports.

REPEAT TRAINING
Use a separate copy; training refuses a pretest freeze. The retained data
contain5880 train states, including two teacher rollout successors per base,
and590 development states. Never remove the original release's freeze.
  python 138/tools/train_inheritance.py
Three seeds138001/2/3 resample1200 training states each. Their fitted values
can agree exactly because the underlying teacher is deterministic and lies
inside the authored candidate library. This is not evidence from independent
neural architectures or equal-budget training.

REGENERATE DATA IN A FRESH RESEARCH COPY
  python 138/tools/make_inheritance_data.py
This refuses existing dataset paths. New final generation requires a frozen
record and uses --final. The release's final seeds are no longer unseen.
Use fresh seeds and a new freeze for a new experiment.

WHAT IS FITTED
Background drift and rotation coefficients; emphasis-dependent edge strength;
edge displacement coefficients and basis selection; categorical edge/region
layer and emphasis operation choices; source-copy routing; membrane radial
and translation coefficients, radius deltas and null phase; source priority
selection; membership/link/birth/deletion/region-addition trees; child colour
pointer and fitted child emphasis/relation; observed new-membrane bounds and
attributes. Intermediate teacher observations and explicitly rule-derived
threshold/priority drills are additional supervision. The membrane-drift
label construction explicitly supplies the teacher's known0.025 coefficient.

WHAT IS AUTHORED
Program decomposition and synchronous source access; edge then region loop
order; candidate vector bases and categorical operator library; default exact
colour retention and pointer-based decoding; nearest-source candidate family;
degree normalization by square root; displacement clip90; integer rounding;
quantization of fitted coefficients to nine decimal places; legal26-unit
spacing and lexicographic fallback; maximum64cells/128links/8regions;
vacant125-unit scaffold; append/last-slot deletion structure; tension feature;
phase/count/attribute bounds; source and membrane canonical sorting. The
teacher-informed auxiliary labels include exact known transition boundaries.
Some structural constants are stored in the weights file for portability but
were not estimated. Read training manifests and source for the full split.

INTERPRETATION
This is a compact fitted source-anchored numerical program. It replaces the
earlier large colour-regeneration backbone for this particular authored task;
it is not a new larger neural network. A direct implementation of the known
teacher is the100% ceiling by construction, with no learning required. Strong
agreement demonstrates dependable execution of those disclosed numerical
conventions. It does not establish colour meaning, semantic comprehension,
experienced colour, consciousness, general intelligence, or a necessity for
learning. Additional structural knowledge/intermediate supervision make the
Job137 comparison a practical same-input comparison, not an equal-budget
causal experiment. Original-ID and same-turn birth-cohort evaluation labels
are external accounting conventions, never model inputs.

DEVELOPMENT RECORD
analysis/development-initial-mismatch.json retains first complete-field check.
analysis/development-rollouts.json retains eight-turn checks for all3seeds.
analysis/development-audit-results.json retains baseline comparison.
analysis/development-parity-negative-zero-failure.json retains the JS failure.
analysis/audit-prefreeze-capacity-failure.json retains explicit-capacity test
failure before adding the grammar safeguard. Gapped-address birth naming was
repaired before freeze after UI review. No final-data model tuning permitted.
