Plexus keeps policy in actor commands and run resources in ETS. The implementation is single-node; the APIs below are kernel building blocks for application strategies.
P1 Population
Plexus.Population provides class/predicate queries, top-k, Pareto, seeded weighted resampling and tournament selection. index/2 and lookup/3 index top-level attributes. Secondary indexes validate candidates against current node metadata and retain historical values until run teardown; use low-cardinality fields. Concurrent queries are not snapshots.
P2 Typed topology
Nodes are an ETS set; class and bidirectional typed-edge indexes are ordered sets with encoded keys. :child is one edge type among arbitrary application atoms. Weights and provenance are retained. Graph.update/3 uses compare-and-swap retries; its callback must be pure because it can run more than once.
P3 Local measurement
Plexus.Measure resolves contracts, checks replay/cache, then reserves credit and submits misses to a coalescer. Memo keys include effective evaluation options. Coalescers deduplicate open-window requests, share results among waiters, retry task saturation, and call TypeSafeSDK.evaluate_many/4. In SDK 0.4.0 each distinct item still uses a transport request: batching controls concurrency; cache/deduplication reduce request counts. Cache eviction is bounded arbitrary eviction, not LRU.
P4 Scheduling
:async executes commands immediately. BSP buffers until Plexus.barrier/1. {:bounded_async, k} allows each actor at most k command envelopes ahead of the slowest active actor. An active idle participant can hold progress back; completed actors leave the cohort. This bounds command updates, not semantic data age. {:priority, fun} orders currently queued envelopes, with FIFO ties; it cannot order future arrivals. Switching regimes flushes buffered commands and records the transition. Only async/BSP have a checked-in controlled experiment.
P5 Belief
Plexus.Belief projects noul to Bernoulli, choice to categorical, and score to ordinal while retaining raw answers. Calibration is explicit. Combining raw scores does not manufacture a calibrated probability. The primality experiment supplies a small held-out calibration example, not evidence for unrelated claims.
P6 Quiescence
Counters cover active actors, managed messages, commands, measurements, expansions and timers. Run.cast/3 and Run.call/4 track queued messages through callback completion. Completion retires an actor once; later termination cannot retire it again. Raw send, direct GenServer calls and the low-level TypeSafe OTP evaluation escape hatch are outside this accounting. Use managed APIs for quiescence-based termination.
P7 Budget/admission
Atomic meters cover measure, expand, tokens and population. Known costs are reserved before work; observed tokens are consumed after completion. max_population also clamps an explicitly larger population budget. Cache/replay hits need no measurement credit.
Budget.Accounts implements finite parent grants, child reservations, sibling unused-credit transfers and leaf close/refund. Grants reserve parent credit up front. Applications explicitly operate scoped accounts through the {:credits, action, args} command; ordinary framework work uses the root atomic meter and is not automatically assigned to child accounts. Ledger state survives ledger-process restart; an interrupted close can conservatively retain credit, but cannot refund twice.
P8 Selection and pruning
Prune synchronously cancels queued/active work, terminates leaves first, then removes topology. Population effects pass through the interpreter. Replication copies initial input unless supplied a snapshot. Split/merge replace compatible leaves using application inputs; failed admission rolls back new births. Migration moves a leaf between runs with an explicit snapshot and does not migrate mailbox contents. Resampling returns selected nodes; applications choose which replication/pruning effects to issue.
P9 Expansion
Expand.InferenceAdapter uses Hex inference 0.4.1, with a separate queue/task supervisor, fail-closed capability checks, neutral streams, monitoring and response accounting. Physical cancellation requires explicit provider support. Schema-validated proposals pass through an application resolver and the normal command interpreter; remote strings never create atoms.
P10 Provenance and repair
Dependency edges retain schema version, upstream epoch and evidence. Recursive invalidation marks nodes stale and enqueues prioritized repair work. Further upstream changes continue advancing a stale node's epoch, while a repair already pending for that node is not queued or notified again until the pending repair clears. Applications submit epoch-checked repairs, so older semantic results cannot overwrite newer evidence. The kernel does not prescribe the semantic repair algorithm.
P11 Record/replay
Events and responses are run-owned ETS. Record.File exports versioned, checksummed fixed-response files with contract/config manifests and safe term validation. Replay misses fail closed. This is response replay, not process/mailbox checkpointing.
P12 Stop/reduce
Stop predicates include quiescence, budget, population, explicit confidence projections, monotonic deadlines, no-improvement, epsilon stability and oscillation. Call Stop.observe/2 once per comparable completed round for history-based predicates. all/1 and any/1 compose them; reducers provide top-k, best-one and collect.