Monk Loom is an optional neural attachment stored inside a Federation Image v2 beside its Monk policy. It is small enough to execute in Pure Perl, deterministic enough to produce the same evidence fingerprint in Python and Perl, and deliberately unable to act on the VFS by itself.
What it can do
- Score named outputs from explicitly configured, bounded integer features.
- Run in observe, advisory, or adaptive mode.
- Produce canonical, fingerprinted evaluation evidence.
- Propose a new final-layer revision from feedback tied to one exact evaluation.
- Survive ordinary Python and Perl VFS mutations without losing its identity.
Version 0.0.11 adds Monk Spindle recipes and explicit root-owned training corpora, so a model can learn directly from verified records stored in VFS Federation while preserving the proposal-only boundary.
Version 0.0.12 adds Chronicle registration and review state. A candidate can be accepted or rejected as evidence without gaining authority to apply or commit itself.
What it cannot do
- Grant read or write authority.
- Override image integrity or capability checks.
- Cross a VFS ownership boundary.
- Rewrite or silently supersede Monk policy.
- Commit its own adaptation proposal.
Why fixed point?
Floating-point libraries and hardware can differ at edge cases. Loom uses signed integer fixed-point math, explicit truncation toward zero, bounded values, and saturation after every layer. That contract is easy to audit and portable to core Perl without a native machine-learning dependency.