Learned Working Set

Corrobore's working set is a bounded, explainable view used to navigate a graph without loading every node and relationship. Epic 0017 is implemented and benchmarked; its migrated epic tracker #86 and final acceptance issue #73 are complete.

This guide describes the current runtime contract.

Implemented layers

Layer	Public surface	Purpose
Lifecycle	GraphWorkingSetManager	Create working sets; load seeds; track hot relationships, warm adjacency, pins, dirty records, stats, and explanations.
Telemetry	WorkingSetTelemetryRecorder, RetrievalTelemetryRecord	Record selected/skipped edges, page-ins, prefetches, evictions, dead ends, evidence, quality, memory, and latency.
Positive field	PheromoneField, EdgeUtility, PheromoneTaskScope	Learn task-scoped expected utility with temporal decay.
Negative field	AntiPheromoneField, AntiPheromoneSignal	Penalize dead ends, supernodes, stale evidence, contradictions, and poisoning risk.
Controller	WorkingSetController, BanditContext, WorkingSetAction, BanditReward	Choose an explicit action under context and observe reward.
Benchmark	run_working_set_benchmark, fimi_multi_hop_benchmark_workload	Compare reproducible policy baselines and metrics.

Decision flow

flowchart LR
    Query["Query, seeds, task, budget"] --> Context["BanditContext"]
    Context --> Controller["WorkingSetController"]
    Positive["PheromoneField"] --> Controller
    Negative["AntiPheromoneField"] --> Controller
    Controller --> Action["expand / prefetch / page-in / stop / verify / external retrieval"]
    Action --> Record["RetrievalTelemetryRecord"]
    Record --> Reward["BanditReward"]
    Record --> Positive
    Record --> Negative
    Reward --> Controller

The learned controller does not replace deterministic budgets or supernode protection. These remain hard safety boundaries.

Benchmark policies and metrics

The harness represents LRU, LFU, static profiles, semantic-only, PageRank/spreading activation, contextual bandit, and learned pheromone policy families. Reports include evidence recall, pages loaded, peak memory, p95 latency, dead-end expansions, supernode blow-ups, and reproducibility metadata.

The target gains stated by Epic 0017 are measured research results on the reference workload, not general production performance guarantees. Use the completed acceptance suite and reproducibility report as the evidence source and scope for any performance statement.

For operational query boundaries and runtime policy behavior, use the HTTP Server guide and Cypher Support as primary references.