Audit checklist

What a reviewer can independently verify, and how. Every row names a command or a file. Nothing here relies on trusting the authors.

1. Can the claimed models be loaded?

check	command	expected
every registered model strict-loads	uv run inmotion model verify	All 18 model(s) verified.
parameter counts match the registry	same command	per-model OK
a model’s output has not silently changed	uv run pytest tests/test_model_outputs_golden.py	pass

load_state_dict(..., strict=False) is banned. It had been hiding a model that was missing 13 keys and therefore predicting partly from random weights.

2. Are the datasets what they claim to be?

check	command	expected
checksums and row counts	uv run inmotion datasets verify	All 7 dataset version(s) match
lineage and label provenance	uv run inmotion datasets show <id>	lineage, audit status, first commit
untrustworthy versions warn	uv run inmotion datasets resolve ds-pure	warning on stderr

3. Is the reported number reproducible?

The three mamba3_* parameter counts and the Mamba-3 HPO search space are known and documented discrepancies, not silent ones. A test fails if either is closed by editing the wrong side.

4. Is the code quality real?

check	command	expected
lint	uv run ruff check src dl ml_classification tests	All checks passed!
formatting	uv run ruff format --check src dl ml_classification tests	no drift
tests	uv run pytest tests/	pass
demo tests	cd demo && uv run --project . pytest	pass
reference docs current	uv run python scripts/docs/generate_reference.py --check	up to date
no committed secrets	uv run pytest tests/test_no_secrets.py	pass

All of the above run together via ./scripts/check.sh. There is no CI by choice, so run it before pushing.

5. Are hidden failure modes covered by tests?

risk	test
an architecture drifts but still loads	tests/test_checkpoint_contract.py (strict load + exact parameter count)
a forward pass changes silently	tests/test_model_outputs_golden.py (frozen fingerprint)
an adopted checkpoint is wrong	tests/test_adoption.py (strict-load assertion per recovered spec)
a config changes between processes	tests/test_config_reproducibility.py
a credential is committed	tests/test_no_secrets.py
the CLI breaks	tests/test_cli.py
reference docs drift from the code	scripts/docs/generate_reference.py --check

6. Is the structure navigable?

starting point	contents
README.md	what the project is, how to run it
docs/README.md	documentation index
docs/architecture.md	how the pieces fit together
docs/reference/modules.md	every module and its public API, generated from source
docs/reference/cli.md	every command and option, generated from the app
docs/reference/models.md	every model, generated from the registry
docs/reference/datasets.md	every dataset version, generated from the manifest

The reference documents are generated, so they cannot disagree with the code. --check fails when they are stale.

7. Data and code availability

item	status
code	this repository
dataset	IEEE Dataport DOI 10.21227/55nm-0r91; a copy is committed under data/processed/
model weights	gitignored, present in backup/ and 20-aug/models/; inmotion model inventory recovers their architectures
licence	not yet declared — add one before submission
citation	no CITATION.cff — the file added in 16f45fd was removed in 37e3983; write one carrying the camera-ready citation before submission

Open items

These are known and not hidden:

1. Neither metadata file exists yet — add both before submission. No LICENSE: no LICENSE/LICENCE/COPYING file exists at the repository root. And no CITATION.cff: the one added in 16f45fd was removed in 37e3983, so both have to be written rather than merely completed.
2. ds-icaisf label provenance needs confirmation; if it belongs to the older lineage the reported ensemble numbers are affected.
3. The classical src/inmotion/classification/ package has not moved under src/inmotion/, so two top-level packages coexist.
4. run_dl.py and run_exotic.py remain large single modules (now pipelines/dl.py and pipelines/exotic.py); splitting them further is mechanical but not free of regression risk.