Pure-Perl runtime API

VFS::Federation::MonkLoom uses only core-style Perl dependencies: Digest::SHA, JSON::PP, and Scalar::Util. Generated packages include the module and document the concrete embedded model.

Through the VFS runtime

my $vfs = VFS::Federation::V2Runtime->open('root.vfsbin');
my $loom = $vfs->monk_neural_attachment();
my $e = $vfs->neural_evaluate({
    path_depth  => 512,
    owner_change => 0,
});
print $e->{selected_label}, "\n";
print $e->{evaluation_fingerprint}, "\n";

Proposal-only adaptation

my $proposal = $vfs->neural_propose_adaptation(\%features, {
    feedback_id           => 'review-0042',
    evaluation_fingerprint => $e->{evaluation_fingerprint},
    targets               => { review => 900 },
    verifier              => 'signed-review-service',
    evidence_digest       => $evidence_sha256,
    verified              => JSON::PP::true,
});
# proposed_attachment_json is not committed by this call.

Direct training from the VFS

my $result = $vfs->neural_train_from_vfs(
    '/ai/training/manifest.json',
    progress => sub {
        my ($p) = @_;
        my $total = defined($p->{total}) ? $p->{total} : '?';
        print "$p->{stage} $p->{completed}/$total\n";
    },
    cancellation_requested => sub { return $cancel_requested },
);
print $result->{report_fingerprint}, "\n";
print $result->{proposed_attachment}->fingerprint, "\n";

VFS::Federation::MonkSpindle validates recipes, records, and deterministic extraction. V2Runtime exact-reads the root-owned corpus, verifies supporting evidence digests, and performs the same ordered training steps as Python.

Low-level reader and writer

• monk_neural_json() returns the exact section bytes.
• build_image(..., monk_policy_json => $monk, monk_neural_json => $loom) validates binding and canonical form.
• Writable transactions preserve both JSON documents during a generation rebuild.

Use the high-level runtime unless you are building tooling. It loads the Monk policy and Loom attachment together and verifies their fingerprint binding.