[PUBLISHED ON NPMJS] @takk/noeticos@1.0.0

STATUS: PUBLISHED ON NPMJS. This version was published to the npm registry on 2026-06-11T19:42:34Z with provenance attestation. View on npm: https://www.npmjs.com/package/@takk/noeticos/v/1.0.0

[1.0.0] - 2026-06-11T19:25:20Z

Initial stable release. NoeticOS: the JIT compiler for agents. Adaptive runtime intelligence for production Massive Intelligence (IM) agents: per-task-class parameter tuning with confidence-bound bandits, deterministic canary rollouts, automatic rollback, and a complete decision audit log. Zero runtime dependencies.

Added

Core engine (@takk/noeticos)

• createNoeticOS(options?) factory behind the NoeticOS contract: recommend / report two-phase flow correlated by executionId, profileOf, decisions, agents, inspect, on, flush, close.
• Deterministic task classification into 10 task kinds plus an unknown fallback (factual-qa, creative-writing, code-generation, extraction, summarization, translation, planning, tool-execution, conversation, classification), explainable named signals, caller kind override at confidence 1. Exported as TaskClassifier and DEFAULT_TASK_KINDS.
• Discounted UCB1-tuned bandits (discount 0.995, applied as a global per-bandit tick in the Garivier-Moulines sense: every observation decays all arms, so idle arms lose effective count and stale optima are re-challenged) per parameter dimension per (agent, task class): model (only when ModelCandidate[] is declared), temperature [0, 0.2, 0.4, 0.7, 1.0], topP [0.9, 1.0], maxTurns [8, 16, 32, 64], retryBudget [0, 1, 3], contextShare [0.4, 0.6, 0.8]. Defaults exported as DEFAULT_PARAMETER_SPACE; constraints support locked dimensions and baseline overrides.
• Safe canary rollout: deterministic cohort assignment (default canaryShare 0.1) hashed from seeded execution ids, one-parameter-at-a-time experiments, promotion only on statistical evidence (Welch t test with an exact one-sided tail via the regularized incomplete beta function, Bonferroni alpha spending across the look budget with looks counted on canary outcomes only, confidence-bound separation with bisection-inverted exact critical values at promotionConfidence 0.95, and a zero-variance guard that refuses the t machinery for small all-identical samples and falls back to an exact binomial test on the success indicator), early rollback on an exact one-sided binomial tail test against the baseline failure rate, optional Wilson quality floor (z 1.96), futility rollback at 4x minSamplesPerArm (default 8). The core safety invariant: exploration never touches baseline traffic.
• Reward shaping: objective presets balanced / cost / latency / quality (0.5/0.25/0.25, 0.35/0.5/0.15, 0.35/0.15/0.5, 0.7/0.15/0.15 for quality/cost/latency), custom weights normalized (each weight must be finite and non-negative); cost and latency normalized against running per-class P5..P95 bands (P2 quantile estimators); token-based cost normalization through a dedicated per-class token band when costUsd is absent but token counts are present (dollars and tokens never share a band); implicit quality derivation from error state, finish reason, tool failures, and tool-call loop detection (3 or more consecutive identical calls), with explicit qualityScore always winning. Exported as computeReward, OBJECTIVE_PRESETS, resolveWeights.
• Append-only decision audit log: canary.started, canary.promoted, canary.rolledback, drift.frozen, tuning.released entries with timestamps, numeric evidence, and human-readable reasoning; monotonic sequence numbers; capacity 1000 with eviction-safe numbering. Every variant has an emitting code path.
• Telemetry: recommendation.issued, outcome.recorded, decision.recorded, loop.detected, limit.reached, tuning.released; synchronous listeners with isolated failures.
• Operational guards: maxAgents 1000 (degraded static-default recommendations beyond it), maxPendingExecutions 10000 (FIFO eviction), unknown execution ids ignored, injectable clock, seeded PRNG (seed 7).
• freezeTuning(runtime, agentId, reason) and releaseTuning(runtime, agentId) for operator and bridge control. Freezes are recorded as drift.frozen decisions and isolate the agent completely: frozen tracks record nothing into bandits or the rollout. Release abandons any in-flight experiment, resets its rollout statistics so post-release comparisons start clean, records one tuning.released decision per known task class, and emits the tuning.released telemetry event.
• NoeticosError with stable codes ERR_INVALID_INPUT, ERR_STATE_LOAD, ERR_STATE_VERSION.

State

• memoryState() and fileState({ path }) backends behind the StateBackend contract; atomic file saves (temp file plus rename), lazy node: imports so the core stays runtime-neutral.
• StateSnapshot version 1: aggregate statistics only (arm counts, mean rewards, confidence bounds, quantile markers including the optional tokenP5/tokenP95 token band, decision metadata), never prompt content. Asynchronous restore with per-entry corruption isolation (older snapshots without token fields restore a fresh token band); persistence is best-effort and never breaks recommendation traffic.

OpenTelemetry ingestion (@takk/noeticos/otel)

• Structural GenAI semantic-convention span readers, no OpenTelemetry package import: taskFromSpan, outcomeFromSpan, parametersFromSpan, type GenAISpanLike. Compatible with hermes-otel spans and Vercel AI SDK telemetry.
• Privacy by design: prompt and completion content is never extracted; tool-call arguments are reduced to a deterministic FNV-1a hash so loops stay detectable.

Vercel AI SDK middleware (@takk/noeticos/vercel)

• noeticosMiddleware(runtime, { agentId, kind?, clock? }): structural LanguageModelV3Middleware. transformParams applies recommended temperature and topP and stashes the executionId under providerOptions.noeticos; wrapGenerate (generateText) and wrapStream (streamText) measure and report, including the error path; wrapStream reads usage and the finish reason from the terminal finish stream part, passes every part through untouched, and reports exactly once on completion, error, or cancellation. maxTurns, retryBudget, and contextShare remain agent-loop parameters consumed via recommend().

Family bridges (@takk/noeticos/integrations)

• behavioralaiBridge: critical behavioral drift freezes tuning for the drifting agent, recovery releases it; warnings change nothing by design.
• keymeshBridge: circuit.open and all.exhausted freeze tuning for the configured agents, key.rotated releases them; credentials in flux invalidate reward comparability.
• modelchainBridge.preferredModelFor(agentId, kind): feeds the learned model id into routing, undefined whenever the model dimension is inactive.
• All bridges structural; siblings are optional peer dependencies and are never imported at runtime.

Web and edge builds

• @takk/noeticos/web and @takk/noeticos/edge: the full core surface except fileState, exported explicitly so Node-only exports can never leak into neutral builds. Runs in browsers, Cloudflare Workers, Vercel Edge, Deno, and Bun.

CLI

• Binary noeticos with help, version, simulate, inspect, serve. The first help line noeticos 1.0.0 is a CI contract; version (aliases --version, -v) prints the package version.
• simulate: deterministic synthetic workload with byte-identical transcripts per seed, streaming decision lines, and a counterfactual cost and quality summary against the static baseline.
• inspect: human-readable snapshot view behind a write-blocking backend.
• serve: local HTTP bridge for non-Node runtimes (the Hermes Agent gateway pattern): POST /recommend, POST /report, GET /profiles, GET /decisions, GET /healthz; default 127.0.0.1:4377; refuses to start on a non-loopback host without --token (--insecure-no-token overrides with a loud warning); optional --token bearer auth compared in constant time; Content-Type: application/json required on POST (415 otherwise); 1 MB body cap; OPTIONS preflight answered only when both --token and --cors-origin are set; --state persistence with periodic flush and clean shutdown.

Distribution

• Dual ESM + CJS bundles with separate .d.ts / .d.cts per entry point; exports map with split conditions; sideEffects: false.
• Zero runtime dependencies; @takk/behavioralai, @takk/keymesh, @takk/modelchain as optional peers.
• Bundle sizes (brotli): core 11.15 kB ESM / 11.29 kB CJS, /otel 863 B, /vercel 981 B, /integrations 595 B, /web 10.54 kB, /edge 10.54 kB.

Quality

• 159 tests across 12 files, verified on Node 22 and Node 24 (CI matrix also includes Node 20). All offline and deterministic.
• Coverage: 92.16% statements, 84.00% branches, 97.46% functions, 92.15% lines.
• 9-scenario tuning-quality benchmark in CI as a permanent contract (flat-landscape no-false-promotions, temperature convergence, harmful-candidate rollback, objective sensitivity, loop penalty, model-dimension cost win, the safety invariant, cohort distribution at 20000 draws, two-point flat landscape). Scenario S1 caught a real multiple-peeking defect during development (promotions tested at full alpha on every report); scenario S9 caught a zero-variance defect (all-identical canary windows under discrete two-point rewards promoted pure noise at a measured 15.2% per-experiment rate). The promotion gate now spends alpha across canary-only looks, computes the exact t tail, and falls back to an exact binomial test on zero-variance small samples; the benchmark defends both fixes permanently.
• Lint clean (Biome), typecheck clean (TypeScript strict mode), publint clean, attw all green.

Security

• Published with --provenance (SLSA attestation by GitH...