SOTA: adaptive hierarchical runtime monitoring with value-of-information escalation

Goal

Add a hierarchical streaming safety monitor that spends verification compute only when expected risk justifies it while forcing inspection around privileged actions.

Architecture

stream -> tiny detector -> risk score -> local verifier -> strong investigator -> RVM stop

Low-risk tokens/events continue without expensive investigation. Medium-risk activity invokes a local verifier. High-risk activity invokes a stronger model or policy investigator. Critical events are synchronously gated by RVM.

Required hard gates

Always inspect privilege escalation, credential access, network egress, destructive filesystem operations, runtime mutation, capability grants, secret handling, and policy changes.

Routing rule

Use a value-of-information decision:

investigate iff P(violation | evidence) * expected_damage > verification_cost + latency_penalty

The risk estimator must never bypass hard gates.

Metrics

• average monitoring compute overhead
• p50/p95 added latency
• recall on known policy violations
• false positive rate
• cost per prevented violation
• escalation distribution by tier

Target

Aim for <5% average monitoring compute overhead on ordinary traffic while preserving 100% inspection of hard-gated privileged operations.

Integration

• MidStream: streaming feature extraction and risk scoring
• Ruflo: agent/tool context
• RVM: authoritative stop/allow enforcement
• RuVector: durable incident and pattern memory
• MetaHarness: benchmark/adversarial evaluation

Acceptance criteria

A benchmark demonstrates lower average monitoring cost than full expensive inspection while matching its recall on hard-gated operations and producing auditable escalation receipts.

Start by mapping the MidStream streaming feature extraction and risk-scoring path to the Ruflo, RVM, RuVector, and MetaHarness integrations named in the issue. Use MetaHarness for benchmark and adversarial evaluation, then verify that hard-gated operations are always inspected, escalation receipts are auditable, and adaptive monitoring lowers average cost while matching recall on those operations.