v0.25.1 — Evaluation Harness, Prompt Optimization, Workflow Generation

What's New

Evaluation Harness (Phase C6)

• 20 benchmark tasks across 7 configurations (single, pipeline, fan-out, fan-in, diamond, full, stress)
• MockEvaluator using real structural analysis — compile→decompile round-trip through all 4 compilers
• Structural variation analysis: measures how topology metrics change across task configurations
• Credit assignment: attributes evaluation outcomes to individual stage contributions
• Guided configs (Operon adaptive, Scion + Operon) have lower mean risk than unguided

Prompt Optimization + Workflow Generation (Phase C7)

• PromptOptimizer protocol with NoOpOptimizer reference implementation; EvolutionaryOptimizer extended protocol
• attach_optimizer — attach optimizer to any SkillStage for prompt-level tuning
• WorkflowGenerator protocol with HeuristicGenerator reference implementation; ReasoningGenerator extended protocol
• generate_and_register — generate workflow topology and register it in PatternLibrary

Documentation

• Fixed example filenames (105/106 _interface suffix)
• Corrected Protocol vs implementation descriptions
• Updated all docs, papers, and website

Numbers

• 22 convergence modules, 1,510 tests, 106 examples
• 6 external frameworks (Swarms, DeerFlow, AnimaWorks, Ralph, A-Evolve, Scion)
• 4 compilers, 5 adapters, 4 TLA+ specifications

Examples

• 104: Evaluation harness with MockEvaluator
• 105: Prompt optimization protocols
• 106: Workflow generation and registration

Install

pip install operon-ai==0.25.1

v0.25.0 — Production Runtime (Phase C5)

What's New

4 Compilers

• organism_to_swarms() → Swarms SequentialWorkflow/GraphWorkflow config
• organism_to_deerflow() → DeerFlow LangGraph session config
• organism_to_ralph() → Ralph event-driven hat/event config
• organism_to_scion() → Scion containerized grove with git worktrees

All compilers output serializable dicts — zero external framework imports.

Distributed Watcher

• DistributedWatcher with InMemoryTransport (single-process) and HttpTransport (webhooks)
• Transport-agnostic signal publishing and consumption

LangGraph Integration

• operon_watcher_node() — native LangGraph node for convergence detection
• Cursor-based incremental processing, pre-intervention convergence check
• Lowercase action strings matching Operon's canonical InterventionKind

Stats

• 103 examples (was 98)
• 1474 tests (was 1404)
• 20 convergence modules (was 14)
• 4 compilation targets (Swarms, DeerFlow, Ralph, Scion)

🤖 Generated with Claude Code

v0.24.1 — Ralph, A-Evolve, Convergence Paper

What's New

Ralph + A-Evolve Adapters

• parse_ralph_config() — event-driven hat configs to ExternalTopology
• parse_aevolve_workspace() — workspace manifests to ExternalTopology
• seed_library_from_ralph/aevolve — catalog seeding
• EvolutionGating.tla — TLA+ spec for evolution loop

Convergence Paper

• Standalone paper: "Convergence by Composition" (article/paper2/main.tex)
• Whitepaper section 13: convergence summary with TikZ 5-layer diagram
• Both papers build cleanly with tectonic

5-Layer Architecture

A-Evolve (evolution) → AnimaWorks (cognitive) → AsyncThink (thinking)
→ Ralph/DeerFlow/Swarms (orchestration) → Operon (structural)

Stats

• 98 examples, 1404 tests, 4 TLA+ specs, 5 orchestration targets

🤖 Generated with Claude Code

v0.24.0 — Convergence Package (C1-C4)

What's New

Phase C1 — Foundation Adapters

• operon_ai.convergence package with type-level bridges to Swarms, DeerFlow, and AnimaWorks
• ExternalTopology abstraction: all adapters produce it, analyze_external_topology() consumes it
• Epistemic theorems applied as structural linter for external workflows
• Examples 86-88

Phase C2 — Template Exchange

• seed_library_from_swarms/deerflow/acg_survey — seed PatternLibrary from external catalogs
• Bidirectional DeerFlow Markdown skill ↔ PatternTemplate bridge
• hybrid_skill_organism() — library-first + LLM generator fallback
• Examples 89-91

Phase C3 — Memory + Thinking Convergence

• PrimingView — multi-channel SubstrateView subclass (truly immutable via MappingProxyType)
• Memory bridge: AnimaWorks/DeerFlow → BiTemporalMemory
• HeartbeatDaemon — idle-time consolidation via WatcherComponent extension
• AsyncOrganizer — Fork/Join execution within stages with concurrency metrics
• prompt_optimizer hook on SkillStage (interface for future DSPy integration)
• Examples 92-95

Phase C4 — Formal Verification + Co-Design

• 3 TLA+ specifications: TemplateExchangeProtocol, DevelopmentalGating, ConvergenceDetection
• Zardini co-design formalization: DesignProblem, compose_series/parallel, feedback_fixed_point
• Example 96

Also in this release

• Paper restructured per peer review (12 sections, proper theorem environments)
• 70+ roborev findings fixed across runtime, docs, and convergence code
• Watcher state leak fix, CLI handler improvements, UTC timestamp compatibility

Stats

• 96 examples (was 85)
• 1346 tests (was 1168)
• 12 convergence modules (new package)
• 3 TLA+ specs (new)
• Paper: 12 sections, all proofs verified, zero undefined references

🤖 Generated with Claude Code

v0.23.3

CLI Stage Handler — External Tool Integration

Shell out to any CLI tool as an organism stage. The full Operon stack works unchanged on CLI-backed agents.

cli_handler()

Factory wrapping CLI commands as SkillStage handlers:

• input_mode: "arg" (append to command), "stdin" (pipe), "none"
• parse_output: optional callable for structured output (JSON, lines, custom)
• success_codes: configurable (default: only 0)
• Timeout detection with timed_out flag
• sanitize_task strips shell metacharacters by default

cli_organism()

Convenience for building a ManagedOrganism from a dict of {stage_name: command}.

_action_type convention

Handlers can now signal FAILURE to the organism by returning {"_action_type": "FAILURE"}, triggering halt_on_block and watcher RETRY interventions.

Example

from operon_ai import cli_handler, managed_organism, SkillStage

m = managed_organism(stages=[
    SkillStage(name="generate", role="Coder", handler=cli_handler("claude --print")),
    SkillStage(name="lint", role="Linter", handler=cli_handler("ruff check --fix")),
])
result = m.run("Fix the auth bug")

10 new tests (1151 total). Example 83.

v0.23.2

Pattern-First Ergonomics Pass

Makes the v0.19-0.23 subsystems as accessible as skill_organism() made v0.18.

managed_organism()

One-call factory wiring the full stack: library + watcher + substrate + development + social learning. Returns ManagedOrganism with .run(), .consolidate(), .export_templates(), .import_from_peer(), .scaffold(), .status(). Everything opt-in — omit a parameter and that subsystem is simply not configured.

consolidate()

One-call sleep consolidation. Pass a PatternLibrary, get back a ConsolidationResult. Creates AutophagyDaemon + SleepConsolidation internally with sensible defaults.

advise_topology() v2

Now accepts optional library and fingerprint parameters. When provided, returns the best matching template from the library alongside structural advice.

Also included

• 11 new tests (1141 total)
• Example 82: full-stack managed_organism demo

v0.23.1

Release Integration + Publication Polish

Phase 8 — the capstone. Completes the 8-phase roadmap.

Memory Adapters

• histone_to_bitemporal() — bridges HistoneStore markers into bi-temporal facts
• episodic_to_bitemporal() — bridges EpisodicMemory entries into bi-temporal facts
• One-way, non-destructive: source memory systems unchanged

Cross-Subsystem Integration Tests

• Substrate + watcher: facts recorded during organism run, watcher observes
• Adaptive assembly + consolidation: run → consolidate → promote templates
• Social learning + development: scaffold with developmental gating and critical periods
• Full lifecycle: telomere → development → library → adaptive → consolidation
• Memory adapters: histone/episodic → bi-temporal bridge with consolidation

Article Polish

• Abstract rewritten for full v0.19–v0.23 six-layer scope
• Conclusion updated with roadmap arc and future work

The Complete Arc

structure (v0.17–0.18) → memory (v0.19–0.20) → adaptation (v0.21) → cognition (v0.22) → development (v0.23) → integration (v0.23.1)

1130 tests. 81 examples. 8 HuggingFace Spaces. Roadmap complete.

v0.23.0

Developmental Staging + Critical Periods

Phase 7: the developmental layer.

DevelopmentController

• DevelopmentalStage enum: EMBRYONIC → JUVENILE → ADOLESCENT → MATURE
• Threshold-based transitions from telomere consumed fraction (configurable)
• Learning plasticity: 1.0 (EMBRYONIC) → 0.25 (MATURE)
• CriticalPeriod — time-limited learning windows that close permanently
• stage_reached() — ordinal comparison helper

Capability Gating

• Plasmid.min_stage — tools require minimum developmental stage
• Mitochondria.acquire() checks developmental stage before granting

Teacher-Learner Scaffolding

• SocialLearning.scaffold_learner() — mature organisms guide younger ones
• ScaffoldingResult — adoption + plasticity bonus + filtered template IDs

Also included

• 18 new tests (1119 total)
• Examples 80 (developmental staging) and 81 (critical periods)
• Developmental Staging Space
• Article: critical periods (§6), developmental staging implementation (§8)

v0.22.1

Social Learning + Curiosity Signals

Phase 6: agents that learn from peers and seek informative inputs.

SocialLearning

• SocialLearning — cross-organism template exchange via PeerExchange
• TrustRegistry — EMA-based epistemic vigilance (trust increases on success, decreases on failure)
• export_templates() / import_from_peer() / record_adoption_outcome()
• Provenance tracking: get_provenance() traces which peer shared each template

Curiosity Signals

• Watcher emits EPISTEMIC/curiosity signals when EpiplexityMonitor detects EXPLORING status
• New step 4.5 in intervention chain: high novelty on fast model → ESCALATE
• curiosity_escalation_threshold in WatcherConfig (default 0.6)
• Refactored epiplexity measurement to avoid double-calling measure()

Also included

• 30 new tests (1101 total)
• Examples 78 (social learning) and 79 (curiosity)
• Social Learning Space
• Article: social learning + epistemic vigilance (§6), implementation (§8)

v0.22.0

Cognitive Modes + Sleep Consolidation

Phase 5: the cognitive architecture layer.

CognitiveMode

• OBSERVATIONAL (System A) / ACTION_ORIENTED (System B) enum on SkillStage
• resolve_cognitive_mode() infers from mode when not explicitly set
• Watcher detects mode mismatches, mode_balance() reports A/B distribution

SleepConsolidation

• 5-step post-batch cycle: prune → replay → compress → counterfactual → promote
• Replays successful PatternRunRecords into EpisodicMemory with tier promotion
• Compresses recurring high-success patterns into consolidated PatternTemplates
• counterfactual_replay() diffs bi-temporal corrections to detect outcome-changing facts
• Promotes frequently-accessed ACETYLATION histone marks to permanent METHYLATION

Also included

• 18 new tests (1071 total)
• Examples 76 (cognitive modes) and 77 (sleep consolidation)
• Consolidation Space
• Article: cognitive modes (§6), consolidation implementation (§8)