Skip to content

Latest commit

 

History

History
74 lines (51 loc) · 2.54 KB

File metadata and controls

74 lines (51 loc) · 2.54 KB

Adaptive Operations

Functional domain: Adaptive Learning
Status: Stable — rule-based statistics and confidence updates; strategy preference on mission abort/replan; no ML dependency.

Canonical architecture: cognitive-resilience-architecture.md Recovery learning detail: adaptive-recovery.md

Purpose

Improve operational decisions using historical outcomes. Adaptive operations cover recovery strategy selection, confidence updates, and historical recommendations — implemented as transparent rules and statistics, not opaque models.

Capabilities (initial)

Capability Implementation
Rule-based adaptation AdaptiveRecoveryPolicy success-rate thresholds
Statistics StrategySuccessRate, rolling attempt counts
Confidence updates RecoveryConfidence.score from history
Historical recommendations POST /v1/recovery/recommend

Entity integration

Field Description
Entity.recovery_confidence EntityRecoveryConfidencescore, preferred_strategy, attempts

Types (spanda-autonomy)

RecoveryConfidence, RecoveryHistory, StrategySuccessRate, StrategyPreference, AdaptiveRecoveryPolicy

Examples

  • Camera reconnect succeeds 3/3 times → prefer reconnect strategy
  • Provider restart fails repeatedly → escalate sooner in recommend
  • Robot replacement faster than retry → prefer takeover in continuity
  • Fusion confidence low after recovery → lower recovery confidence score

CLI

spanda recovery confidence
spanda recovery learning-report

API

Surface Endpoint
Recovery metrics GET /v1/recovery/metricsrecovery_confidence
Recovery recommend POST /v1/recovery/recommend
Entity autonomy GET /v1/entities/{id}/autonomyrecovery_confidence

Control Center

Recovery Confidence metric in the Cognitive & Resilience tab.

Integrations

  • Recovery Orchestrator: history feeds compute_recovery_confidence()
  • Attention Engine: failed recovery patterns may escalate attention
  • Operational Memory: episodic recovery history in playbooks/traces

Future (not required now)

ML-backed strategy selection may arrive via official packages — the functional domain boundary remains the same; only the recommendation backend would change.

See responsibility-matrix.md.