Meta-Transfer

Why this subsystem exists

This is a bootstrap surface for policy transfer between domains, not a production adaptive controller. Its role is to make replay history reusable by giving operators an interpretable first proposal from prior domains.

In enterprise contexts, this reduces “blank start” risk for new deployments: teams can start from documented historical baselines rather than writing new policy defaults from scratch.

Decision point for operators

Meta-transfer is a bootstrap, not a controller. Its value is highest during:

• first-day onboarding of a new domainpack,
• recovery after a topology drift event,
• and cross-domain comparison when handoff teams need a first proposal quickly.

Because every output is review-only, teams preserve human authority over policy promotion while gaining a deterministic starting point from prior audited history.

Governance rule

All proposals from this surface remain advisory until the normal policy stack approves them. Evidence packages should be treated as hypothesis-generating artefacts and validated through the same deterministic replay channels as any other control-change candidate.

The meta-transfer subsystem provides a deterministic first slice for cross-domain policy bootstrapping. It reads replay or audit-derived records, embeds domain metrics into a shared feature vector, and proposes initial supervisor knobs from nearest historical neighbours.

This is not an online autonomous trainer. Proposals are reviewable starting points for policy authors, and every proposal exposes neighbour evidence and a serialisable audit record.

from scpn_phase_orchestrator.meta import CrossDomainMetaTransfer, MetaPolicyRecord

records = (
    MetaPolicyRecord("power_grid", {"R_global": 0.4}, {"K": 0.08}),
    MetaPolicyRecord("cardiac", {"R_global": 0.8}, {"zeta": 0.05}),
)
model = CrossDomainMetaTransfer.fit(records)
proposal = model.propose({"R_global": 0.5})

audit_payload = proposal.to_audit_record()

Larger replay corpora can be loaded from explicit audit JSONL file lists with CrossDomainMetaTransfer.fit_audit_history() or from nested audit directories with CrossDomainMetaTransfer.fit_audit_directory(). Directory loading uses records_from_audit_directory() and discovers **/*.jsonl by default, so multi-domain replay corpora can be trained without hand-listing every audit file. The fitted model exposes an audit-ready training_summary with record count, domain count, feature keys, knob keys, and reward range. Use to_json_package() and from_json_package() to save and restore a deterministic review package for proposal jobs. Audit JSONL ingestion and JSON package import reject non-finite constants, duplicate object keys, and non-object package payloads before records enter the nearest-neighbour proposal surface. to_package_manifest() emits a packaging-readiness manifest for the optional scpn-meta surface: it binds the deterministic JSON package SHA-256, public import target, console-script name, and training summary while keeping execution_permitted=false. It does not build, install, run, or upload a package.

The same review-only manifest can be emitted from the CLI for release and operator review jobs:

spo meta-transfer-manifest audit_grid.jsonl audit_cardiac.jsonl --min-records 2
spo meta-transfer-manifest --audit-directory audit_history --min-records 10

Both forms print manifest JSON to stdout unless --output is provided. The command accepts explicit audit JSONL files or one nested audit directory, never both, and still keeps execution_permitted=false; it does not build, install, upload, or execute scpn-meta.

Installed packages also expose the same review-only command as scpn-meta. This console script is intentionally narrow: it points to the manifest exporter, not the full SPO runtime CLI, so packaging metadata matches the manifest without adding a live training or execution surface.

How teams typically use it

The operational path is usually:

1. Collect comparable replay corpus (or nested history directory),
2. Fit and inspect training_summary,
3. Generate proposals and review neighbour evidence,
4. Export a manifest for reproducible transfer handoff.

That sequence keeps transfer evidence, not just transfer parameters, part of the release documentation.

Leave-one-domain-out transfer sweep

Proposing knobs is one thing; claiming a detector generalises across domains is a far stronger, and more easily overstated, claim. meta.leave_one_domain_out runs the honest test: hold out each domain in turn, transfer the pooled remainder onto it, and aggregate the per-fold verdicts under a rule that never upgrades. A single domain that is detectable within-domain yet receives no transfer skill (transfer_negative) refutes generality decisively — the recorded CHB-MIT cross-subject negative must surface as lodo_negative, never a laundered aggregate positive. Only an unbroken sweep of positive folds earns lodo_generalises; a sweep with no detectable target at all is lodo_untestable; anything in between is lodo_inconclusive.

from scpn_phase_orchestrator.meta import (
    LeaveOneDomainOutFold,
    leave_one_domain_out_transfer,
)

report = leave_one_domain_out_transfer(
    [
        LeaveOneDomainOutFold("grid", transfer, within, controls),
        LeaveOneDomainOutFold("chbmit", chb_transfer, chb_within, chb_controls),
    ]
)
verdict = report.verdict  # e.g. "lodo_negative"
audit_payload = report.to_record()

Every arm is scored by the caller and audited through the same honest audit_cross_domain_transfer calibration, so the sweep stays a pure, deterministic aggregation with no hidden training step.

::: scpn_phase_orchestrator.meta.transfer

::: scpn_phase_orchestrator.meta.leave_one_domain_out