Beneficial learning
When validation improves and the model keeps a beneficial trajectory, MemGuard abstains. Not acting is also a decision.
CONTINUE
Two-Door AI control
MemGuard - continue, stop, or restore
MemGuard 0.4 is an experimental controller for AI training logs. It separates trajectories that are still learning from those leaving a beneficial regime, and from those that never really enter a useful regime.
Orange-Red
The model has entered a useful regime, then validation persistently degrades while train loss keeps decreasing.
ROLLBACK
No-Entry
Training loss decreases, but validation never really improves beyond the initial state. The initial reference is therefore essential.
NO-ENTRY
Light prospective validation.
Across nine new training runs covering two models, two datasets, and three learning regimes, the INSACERMO Two-Door controller prospectively outperformed standard early stopping while preserving beneficial learning and restoring the exact best checkpoints with no median regret.
3/3beneficial runs preserved
3/3exits detected
3/3No-Entry cases detected
6/6exact restorations
Less blind action.
In this bounded protocol: MemGuard potentially avoids 50.00% of planned steps, compared with 41.67% for standard early stopping, for a weighted advantage of +8.33 points. Macro advantage: +8.64 points.
This prospective validation is lightweight and bounded to this protocol. It is not proof of universal superiority across all models or training runs.
Expected columns.
The official demonstrator accepts a CSV with step, train_loss, and validation_loss. For exact No-Entry semantics, the first row must represent the reference evaluation before fine-tuning.
When MemGuard acts, it does not only recommend stopping: it indicates the best checkpoint to restore. Stopping without restoration is not the validated protocol.
Available evidence.
The public PDF and V10a evidence are available. The official package and private archives are not offered for download on the site.
INSACERMO Morphology Auto Detector V28.
Moving is not becoming. V28 reads AI training logs through morphologies: dominant dynamics, hidden closure, regime transitions, locking level, alert level, and behaviors not covered by the current grammar.
This page is separate from MemGuard Two-Door: it presents a public automatic morphology detector without publishing sensitive real logs or private archives.
Test your logs without sending model weights.
INSACERMO can audit one or two public or private log sets for free: no model weights, no infrastructure access, only aligned logs. Possible deliverables: Two-Door verdict, recommended checkpoint, early stopping comparison, limits, and interpretation cautions.
No public case study is published without written permission.