JOB137 — A PLACE FOR WHAT COMES NEXT

Keep the extracted record's sibling folders 134,135,136,137 together.
The paths below are relative to their shared parent. No network is used by
inference, training, the scientific evaluator, or the specialist parity test.

RECORDED ENVIRONMENT
Python3.12.14, NumPy2.3.5, PyTorch2.7.1+cu126; Windows.
NVIDIA GeForceRTX3070TiLaptopGPU. PyTorch CPU is also accepted by the code;
different numerical libraries/hardware are not promised to reproduce every bit.
Node24.19 was used for JavaScript parity checks.
NumPy alone is enough for offline full-field inference. Training needs PyTorch.

OFFLINE REPLY, WITHOUT TRAINING
  python 137/tools/growth_reply.py 137/artworks/CHOICE-S0/CHOICE-S0.json reply.json
The supplied original field bytes become the reply's exact parent hash.
The reply includes the frozen backbone result, specialist proposal, learned
growth veto, pointer-based child-colour copy, and separate grammar repairs.

JAVASCRIPT / NUMPY PARITY
  node 137/tools/check_growth_parity.cjs
Six development-only fixtures include identical supplied backbone outputs.
This isolates the specialist, including its cell-count veto and child colour.
The original backbone's separate parity evidence is retained in the record.

REPEAT THE CONFIRMATORY EVALUATION WITHOUT OVERWRITING IT
  python 137/tools/evaluate_growth.py --replay-prefix repeat-final
This verifies the explicit pretest frozen-file hashes and final input hashes,
then evaluates existing final scenes into analysis/repeat-final-results.json
and analysis/repeat-final-predictions.jsonl. Existing output prefixes are refused.
Run from a complete extracted record; the code does not fetch dependencies.
The original first evaluation is retained at analysis/final-results.json.
Repeating is a reproducibility check, not another untouched scientific holdout.

TRAINING REPRODUCTION
Use a SEPARATE COPY of the extracted record; training replaces that copy's
run/model weights, and is deliberately not a publication-update command.
  python 137/tools/train_growth.py
  python 137/tools/train_growth_gate.py
The first command fits trees using bootstrap training-scene samples and trains
three independent linear-pointer initializations. The second fits growth gates
using the retained source+frozen-backbone feature caches. Inspect all manifests.
No final scenes are loaded by either training command.

To regenerate feature caches or train/development scenes, use a fresh research
directory and retain originals for comparison; the generators refuse existing
dataset paths. The model plan records exact source seeds and group counts.
Do not overwrite or remove pretest-freeze.json to describe a new run as the old
experiment. Any substantive change requires its own development record and
new final seeds. The current final inputs are public after release.

DEPENDENCIES READ BY SCIENTIFIC EVALUATION
134/data/palette.json;134/common/topology.json
135/tools/field_model.py;135/tools/field_model_v2.py;
135/tools/tensorize.py;135/tools/make_exercises.py;
135/runs/full-v2/weights.npz
136/tools/steward_model.py;136/model/weights.npz
137/tools and model/run/data files listed exactly in protocol/pretest-freeze.json
The prior136 art/reference tool additionally uses136/tools/steward_reply.py,
136/data/palette.json and136/runs/seed-136001/weights.npz.

PRESERVED DEVELOPMENT RECORD
analysis/development-v1/ retains initial weights, code/plan and the first
development result/failure note. The initial evaluator completed predictions
then failed JSON serialization of a NumPy integer. This was repaired without
weakening checks. That development run also motivated equal priority-drill
weighting and the learned availability gate, before new final data existed.
analysis/development-results.json and growth-gate-development.json retain
the revised development outcomes and all attempted gate depths.

INTERPRETATION OF RESULTS
The teacher is an authored numerical rule system. Known-boundary labelled
drills deliberately supply its transitions; the learned cutoffs are fitted
parameters, not discovered meanings. A direct implementation of the teacher
could reproduce its targets exactly. The experiment measures approximation,
retained structure and practical interaction, not a necessity for learning,
consciousness, colour experience, or broad semantic intelligence.

ADDITIONAL VERIFICATION AND PUBLICATION
  python 137/tools/test_growth_audit.py
  python 137/tools/test_growth_lineage.py
  node 137/tools/transport_artworks.cjs verify
  node 137/publication/ui-test.cjs
  node 137/publication/constraint-test.cjs
The lineage diagnostic is a disclosed post-test measurement correction; its
external entity labels are never supplied to the learner. Read both the initial
canonical-slot diagnostic and the identity-aligned report before interpreting it.
The lineage metadata correction and its rejected stale code hash are retained
in analysis/lineage-metadata-correction.txt and the named failure JSON.

To rebuild the publication in a separate extracted working copy:
  python 137/tools/publish_continuity.py
This requires Matplotlib (recorded3.10.7). Regenerating artwork PNGs requires
Pillow (recorded12.3.0); the exact already-verified PNGs are included. Node and
NumPy are required for their respective verification commands above. Existing
artwork renderers refuse overwriting different retained outputs.
