jev-clean learned cleanup adapter (v3)
Nexora Research · Chengyi Xu and Nexora
Adapter weights used by jev-clean 0.4.0 and later to decide whether a measured macOS storage resource should be proposed for removal, reviewed, or kept. jev-clean downloads this file automatically and refuses to load it unless its SHA-256 matches the value pinned in the application.
| Adapter ID | jev-cleanup-laya-runtime-observations-v3 |
| File | model.safetensors (56 float32 tensors) |
| SHA-256 | 5b65d16deeb20ac8c9e6ab43cb3c9fe4478896df8e2326e5744559797546bd33 |
| Base checkpoint | aac6fef/laya-mlx at revision 20aed815fc6acde75733882e7ec0e3f28aeb9717 |
What the adapter replaces
Encoder layers 24–27, the encoder final norm, the type embedding, the decision head, and the scorer of the base checkpoint. All other base weights are loaded unchanged. jev-clean rejects any tensor outside that set.
Training
Supervised cross-entropy fine-tuning (not reinforcement learning) on synthetic, path-free observation
counterfactuals produced by the application's unchanged production observation formatter
(scripts/runtime_resource_corpus.py and the authored file-resource corpus in the jev-clean repository).
No real user paths, file names, or file contents are training inputs. Six training option orders, microbatch 4,
600 updates, learning rate 3e-6, false-removal training cost 4.
Evaluation
From the project's frozen reports, on authored controls that are not independent human ground truth:
- 38-observation native control: 38/38 correct, 0 false removals (starting adapter 34/38).
- 120-case authored regression: 120/120 correct, 0 false removals (starting adapter 114/120).
- Six-order final diagnostic: 718/720 correct, 1 false removal, retained in the published report.
Methods and limitations: POSTTRAINING-PARITY.md.
Limitations
Probabilities are uncalibrated and are not a deletion-safety guarantee. The adapter can be wrong. jev-clean only proposes resources; a human must select them and confirm, every target is revalidated, and removal stages files in the Trash with undo. Review every proposal.
License and attribution
Apache-2.0. Derived from Laya by Convai Innovations and contributors (Apache-2.0) through the independent MLX
port laya-mlx. Modification: the tensors listed above were fine-tuned by Nexora. See NOTICE.
Quantized