Phase 182, 182.C.31.06. The item offers a fork: harden this engine with a real
entry point, weights, licenses, dependencies, failure checks, inference tests
and benchmark evidence — or retire it from capability claims. This records
the second, with the facts that decided it.
Nothing here says the code is bad. 151 Python modules and ~75,000 lines of Gaussian-splatting, NeRF, FLAME-driven expression and lip-sync machinery is a serious body of work. It says that no capability claim may rest on it today, and names exactly what each blocker is.
The four facts#
Each of these is re-derived from the repository by
tests/test_capability_status.py, so this document cannot drift away from the
thing it describes.
1. The entry point does not exist#
project.json names avatar_engine.main:app in both its serve and
serve-prod targets, and the Dockerfile CMD names it a third time. There is
no src/avatar_engine/main.py.
So nx serve psyche-avatar-engine fails at import, and a container built from
this Dockerfile fails at start. Nothing that consumes this service has ever run
against it, because there is nothing to run.
2. No test exercises the engine#
tests/ contained a single .gitkeep before this audit. The one test file that
exists now is the guard that checks this document — it imports nothing from
avatar_engine and proves nothing about it.
3. The weights are absent, and there is no way to acquire them#
config.py defaults model_path to models/flame/generic_model.pkl. That
directory does not exist, no file in the repository provides it, and no
acquisition step, download, checksum or digest is defined anywhere in the
project. 182.C.17.02 is the shape an honest acquisition record would take:
pinned artifact, digest, fetch date.
4. FLAME and SMPL-X are research-licensed, and this is the blocker engineering cannot clear#
The dependency set includes smplx and chumpy, and the default model path is
FLAME's generic_model.pkl. Both FLAME and SMPL-X are distributed for
non-commercial research under licences that require a separate agreement for
commercial use.
That is not a bug to fix. It is a licensing decision to make, and until somebody
with authority makes it, a commercial product claim resting on this engine would
be claiming a right we do not have. 182.C.32.02 is where model licensing is
recorded.
What is not claimed here#
The project declares gpu:required and depends on pytorch3d and open3d. No
inference has been attempted on this hardware, so nothing in this document is a
claim about whether the engine works — only about whether anything may
currently be promised on its behalf.
What would lift the retirement#
In this order, because each depends on the one before:
- A licensing decision on FLAME/SMPL-X — or a substitution for a model whose licence permits the use. Everything else is wasted until this is settled.
- An entry point that exists, with
nx servedemonstrated. - A weights-acquisition record: pinned artifact, digest, fetch date.
- An inference test that runs the engine on a known input and asserts something about the output.
- Benchmark evidence on a corpus, at which point
182.C.32decides whether the engine is competitive with the routes already available.
Until then, this engine is not in the capability matrix, not in a provider registry, and not behind any customer-facing promise.
<!-- capability-status: machine-readable, checked by tests/test_capability_status.py -->
{
"retired_from_capability_claims": true,
"entry_point_module": "src/avatar_engine/main.py",
"entry_point_present": false,
"entry_point_references": 3,
"engine_tests_present": false,
"default_model_path": "models/flame/generic_model.pkl",
"weights_present": false,
"research_licensed_dependencies": ["smplx", "chumpy"]
}