JOB 136 — WHAT THE FIELD KEEPS

The publication instrument is standalone. Open publication/posts/index.html
or upload publication/upload beneath the website document root.
No remote model, API key, server or online training is needed.

The complete scientific ZIP preserves sibling directories 134, 135 and 136.
Keep that directory arrangement when extracting. The older sibling folders
contain only retained dependencies; they are not the complete older jobs.

REQUIREMENTS
Python 3.10+ with NumPy for offline model replies.
PyTorch with CUDA for the recorded evaluation/training commands (the evaluation
currently chooses CUDA explicitly). Python/pip dependencies are not bundled.
Node.js for transport, JS parity and UI logic checks.
Pillow and Matplotlib for artwork and figures.
Original workspace: Python3.12.14 environment in ../134/.venv, PyTorch2.7.1+cu126.
Do not expect a virtual environment at that path in the portable record.

FIXED-WEIGHT REPLY (from the extracted record root)
python 136/tools/steward_reply.py 136/artworks/PROMISE-S0/PROMISE-S0.json reply.json

The reply includes the earlier backbone proposal, revised field, learned
actions, raw candidates, adapter events and exact input-byte hash. This CLI
does not apply optional user protection. That policy is separate in the web
instrument and exercised by the guard tests. Definitions are not model inputs.

INDEPENDENT AUDIT OF STORED PREDICTIONS
python 136/tools/test_steward_metrics.py
python 136/tools/test_steward_model.py
python 136/tools/test_steward_results.py
python 136/tools/test_steward_stress.py
node 136/tools/test_steward_guards.cjs
node 136/publication/constraint-test.cjs
node 136/publication/ui-test.cjs
node 136/tools/check_steward_parity.cjs

Do not overwrite a frozen result to rerun it. Copy the extracted record to a
separate reproducibility workspace, preserve original analysis/runs/protocol,
and use a fresh result destination before numerical re-evaluation or training.
evaluate_steward.py intentionally refuses to overwrite fresh-test-results.json.
train_steward.py skips seeds with completed manifests. Its cached feature files
are included; clear/move only those copies if regenerating them from source.

The complete record includes data, model generator, train/evaluation code,
initial and trained weights, original freezes, all failure records, prediction
archives and independent audits. Exact reruns should use the same hardware,
library versions and deterministic settings; cross-platform bitwise training
equivalence is not claimed. Fixed-weight NumPy/JS parity is separately tested.

ART TRANSPORT
node 136/tools/transport_artworks.cjs verify
All18 retained fields have artworkPNG+layout, standalone CVP2PNG, USL1,
Carrier words and raw JSON. The embedded dictionary/state and codecs are
included under136/common. These checks verify transport, not interpretation.

FROZEN DESIGN AND KNOWN FAILURES
Read136/protocol/model-plan.json andpretest-freeze.json; analysis includes
serialization and lineage helper repairs plus incidental freeze-file changes.
Model checkpoints were not changed after primary results. The exact stress
diagnostics were authored afterward and remain exploratory.
The independent reader is pretrained; its interpretation is not the small
model's prose and not evidence of colour experience.

PUBLICATION VERIFICATION LIMIT
Static link checks, real numerical inference, inert DOM checks and guard tests
are distinct from browser rendering. Local file preview was blocked by the
browser security policy. Physical phones and live WordPress were not tested.
