vsf-rsi — Recursive Self-Improvement for VSM systems

Don't ask the language model to improve itself. Give the improvement a substrate.

vsf-rsi is a cybernetic feedback loop that observes, evaluates, and improves socratic-engine predicates and trees through four levels of recursive self-improvement (RSI).

The system is deliberately bounded. It does not try to make the AI "smarter" unboundedly. It gives the system a formal structure in which improvements can be proposed, executed, validated, and rolled back — all within VSM safety constraints.

Status: v0.2.5 — Adversarial benchmark suite. 775 tests, context-aware predicates, behavioral validation, 4 scenario types (prisoner, parábola, XOR 5D, señal+ruido).

Installation

pip install vsf-rsi

With socratic-engine (required dependency):

pip install vsf-rsi socratic-engine

From source:

git clone https://github.com/rm-w3kufe/vsf-rsi.git
cd vsf-rsi
pip install -e ".[dev]"

The bet

LLMs are excellent at proposing solutions. They are less reliable when they must implicitly improve their own evaluation criteria while simultaneously maintaining safety guarantees.

The bet here is simple:

Let the model propose improvements. Let a deterministic substrate validate and execute them. The model does not certify its own improvements.

flowchart LR
    A["LLM<br/>propose improvement"] --> B["RSI Observer<br/>capture + discriminate"]
    B --> C["Resolution<br/>drift / inject / generate / evolve"]
    C --> D["SocraticEngine<br/>re-evaluate"]
    D --> E["Metrics<br/>did it improve?"]
    E --> A

This creates a division of labour:

The central boundary is:

The LLM proposes. The substrate validates. The LLM does not self-certify.

What it is

At its core, vsf-rsi provides:

• Observer (rsi_observer.py) — wraps SocraticEngine.evaluate() to capture every evaluation event, discriminate errors, and resolve them through L1-L4
• Bridge (rsi_bridge.py) — integrates with state-canon-mcp for canonical state queries, metrics feedback, and rule enforcement
• Metrics (rsi_metrics.py) — tracks classification history, accuracy, latency, and threshold adjustments
• Feedback Loop (rsi_feedback_loop.py) — manages threshold adjustments with drift bounds and step limits
• Scenario Memory (scenario_memory.py) — procedural learning: records failures with correction paths, matches novel faults to prior corrections
• Gap Detector (rsi_gap_detector.py) — detects evaluation gaps: stale predicates, missing branches, low accuracy
• Pattern Detector (rsi_pattern_detector.py) — identifies recurring error patterns with time-based decay (10%/day)
• Predicate Generator (rsi_predicate_generator.py) — creates new predicates from error patterns (with AST validation)
• Tree Generator (rsi_tree_generator.py) — generates evaluation trees from gap analysis
• Forest Generator (rsi_forest_generator.py) — generates populations of trees for genetic evolution
• Genetic Algorithm v1 (rsi_genetic_algorithm.py) — evolves predicate populations with tournament selection, crossover, mutation, and convergence detection
• Genetic Algorithm v2 (rsi_genetic_algorithm_v2.py) — improved GA with adaptive mutation rates
• Tree Registry (rsi_tree_registry.py) — manages predicate versions, prevents conflicts, tracks lineage
• Component Registry (rsi_component_registry.py) — maps package components to their roles and dependencies
• Manifest Parser (rsi_manifest_parser.py) — parses RSI manifest files for tree/predicate registration
• Benchmark (rsi_benchmark.py) — load scenarios, run benchmarks, save/load reports, compute improvement curves
• Adversarial Scenarios (rsi_adversarial.py) — 4 scenario generators: Prisoner's Dilemma, Parábola Silenciosa, XOR 5D, Señal en Ruido
• Demo (rsi_demo.py) — runnable demo: generate → evaluate → evolve → measurable improvement
• Error Recovery — system continues with degraded functionality on failures (never crashes)
• Logging — structured logging via logging.getLogger("vsf_rsi.observer")

The four levels of RSI

flowchart TD
    subgraph L1["L1: Parameter Drift"]
        A1["adjust thresholds<br/>(autonomous)"]
    end
    
    subgraph L2["L2: Capability Extension"]
        A2["inject predicate wrappers<br/>(autonomous)"]
    end
    
    subgraph L3["L3: Predicate Generation"]
        A3["generate new predicates<br/>(human approval)"]
    end
    
    subgraph L4["L4: Genetic Evolution"]
        A4["evolve predicate populations<br/>(human approval)"]
    end
    
    L1 -->|"if drift fails"| L2
    L2 -->|"if pattern detected"| L3
    L3 -->|"if enough data"| L4

state-canon-mcp integration

The rsi_bridge module provides functions to integrate with the state canon:

from vsf_rsi.rsi_bridge import get_rsi_rules, query_canon, feed_metrics_to_canon

# Get rules that constrain RSI behavior
rules = get_rsi_rules()

# Query canonical state
result = query_canon("rsi_metrics", {"predicate": "stasis_check"})

# Feed metrics back to state canon
feed_metrics_to_canon({"total_evaluations": 100, "error_rate": 0.05})

CLI usage

Each module has its own CLI entry point:

# Feedback loop — adjust thresholds
python -m vsf_rsi.rsi_feedback_loop adjust
python -m vsf_rsi.rsi_feedback_loop status
python -m vsf_rsi.rsi_feedback_loop history

# Pattern detector — analyze error patterns
python -m vsf_rsi.rsi_pattern_detector detect
python -m vsf_rsi.rsi_pattern_detector detect --predicate stasis_check
python -m vsf_rsi.rsi_pattern_detector summary

# Gap detector — find evaluation gaps
python -m vsf_rsi.rsi_gap_detector detect
python -m vsf_rsi.rsi_gap_detector gaps

# Tree generator — create trees from gaps
python -m vsf_rsi.rsi_tree_generator generate
python -m vsf_rsi.rsi_tree_generator list

# Forest generator — create populations for GA
python -m vsf_rsi.rsi_forest_generator generate --predicate stasis_check
python -m vsf_rsi.rsi_forest_generator evolve --predicate stasis_check
python -m vsf_rsi.rsi_forest_generator best --predicate stasis_check
python -m vsf_rsi.rsi_forest_generator list

# Component registry — manage components
python -m vsf_rsi.rsi_component_registry list

Demo

Run the complete end-to-end demo:

python -m vsf_rsi.rsi_demo

The demo demonstrates the full RSI cycle:

1. Creates an SocraticEngine with built-in predicates
2. Runs 10 evaluations (mix of passes and failures)
3. Records failures in scenario_memory
4. Matches novel faults to prior corrections
5. Applies scenario_memory corrections to improve results
6. Generates patterns from error history
7. Evolves predicate populations with the genetic algorithm
8. Reports measurable improvement

Testing

# Run full suite (708 tests, 99% coverage)
pytest tests/ -v

# Run with coverage
pytest tests/ --cov=vsf_rsi --cov-report=term-missing

# Run integration tests only (requires socratic-engine)
pytest tests/test_rsi_observer_integration.py -v

# Run scenario_memory integration test
pytest tests/test_scenario_memory_integration.py -v

Architecture

flowchart TD
    subgraph SocraticEngine["Socratic Engine"]
        SE["evaluate(tree, ctx)"]
        PD["predicates"]
    end
    
    subgraph RSIObserver["RSI Observer"]
        OB["wrapper: evaluate()"]
        BR["bridge: track_classification()"]
        DI["discriminate: BLOCKING / STRUCTURAL"]
        RE["resolve: L1 / L2 / L3 / L4"]
    end
    
    subgraph RSIModules["RSI Modules"]
        MET["rsi_metrics<br/>track + aggregate"]
        FL["rsi_feedback_loop<br/>adjust thresholds"]
        SM["scenario_memory<br/>match + record"]
        PG["rsi_predicate_generator<br/>create predicates"]
        GA["rsi_genetic_algorithm<br/>evolve populations"]
    end
    
    subgraph VSMConstraints["VSM Safety"]
        A0["A0: purpose inalienable"]
        A5["A5: autonomy ceiling"]
        R17["R17: destructive consent"]
    end
    
    SE -->|"capture"| OB
    OB -->|"feed"| BR
    BR -->|"store"| MET
    MET -->|"analyze"| DI
    DI -->|"classify"| RE
    RE -->|"L1"| FL
    RE -->|"L2"| OB
    RE -->|"L3"| PG
    RE -->|"L4"| GA
    FL -->|"adjust"| PD
    SM -->|"match"| RE
    SM -->|"record"| RE
    A0 -.->|"constrain"| RE
    A5 -.->|"constrain"| RE
    R17 -.->|"constrain"| RE

Quick start

from vsf_rsi import RSIObserver, RSIMode
from socratic_engine import SocraticEngine

# Create engine with predicates
engine = SocraticEngine()

@engine.register("my_predicate")
def my_predicate(ctx, threshold=0.5, **kw):
    val = ctx.get("value", 0.0)
    from socratic_engine.engine import PredicateResult, Truth
    return PredicateResult(
        truth=Truth.TRUE if val > threshold else Truth.FALSE,
        certified=True,
        evidence={"value": val},
        source="my_predicate",
    )

# Create observer
observer = RSIObserver(engine, mode=RSIMode.CAPABILITY.value)

# Evaluate a tree
tree = {"predicate": "my_predicate", "args": ["$ctx", 0.5]}
ctx = {"task_id": "example", "value": 0.3, "input_value": 0.3}
event = observer.evaluate(tree, ctx)

# Check what happened
print(f"Error: {event.is_error}")
print(f"Class: {event.error_class}")

# Run pattern analysis
patterns = observer.analyze_patterns()
print(f"Patterns: {len(patterns)}")

# Generate predicates from patterns
actions = observer.generate_predicates(min_occurrences=3)
print(f"Generated: {len(actions)}")

Integration with socratic-engine

vsf-rsi depends on socratic-engine as its recursive evaluation substrate. The integration is via the observer wrapper:

sequenceDiagram
    participant User
    participant Observer
    participant Engine
    participant Metrics
    participant Resolver
    
    User->>Observer: evaluate(tree, ctx)
    Observer->>Engine: original evaluate()
    Engine-->>Observer: result
    Observer->>Metrics: track_classification(event)
    Observer->>Observer: discriminate(error)
    
    alt error detected
        Observer->>Resolver: resolve_error(event, ctx)
        Resolver->>Resolver: try L1 (parameter drift)
        alt L1 fails
            Resolver->>Resolver: try L2 (inject predicate)
        end
        Resolver-->>Observer: action
    end
    
    Observer-->>User: event + actions

Integration with agent systems

vsf-rsi is designed to be used by any AI agent or human operator. Here's how to integrate it:

Available tools

Integration by platform

OpenCode

Add to your opencode.json:

{
  "mcp": {
    "socratic-engine": {
      "type": "local",
      "command": ["python3", "/path/to/socratic_mcp_with_bridge.py"]
    },
    "state-canon": {
      "type": "local",
      "command": ["python3", "/path/to/mcp_server.py", "--instance", "/path/to/your_state.py"]
    }
  }
}

Then use in your agent:

# Agent calls these via MCP tools
from socratic_engine import SocraticEngine
from vsf_rsi.scenario_memory import record

# Evaluate a decision
engine = SocraticEngine()
result = engine.evaluate(your_tree, your_context)

# Record what happened
record({
    "fault_signature": "what_went_wrong",
    "decision": "what_you_did",
    "outcome": "success_or_failure"
})

Claude Code

Add to your CLAUDE.md:

## Tools available

- `socratic_engine.evaluate(tree, context)` — evaluate decisions
- `state_canon.query(domain, filter)` — get ground truth
- `vsf_rsi.record(pattern)` — learn from outcomes

Perplexity / ChatGPT / Other agents

Install the packages:

pip install socratic-engine vsf-rsi
git clone https://github.com/rm-w3kufe/state-canon-mcp.git

Use in your code:

from socratic_engine import SocraticEngine
from vsf_rsi.scenario_memory import record

# Evaluate
engine = SocraticEngine()
result = engine.evaluate(tree, context)

# Learn
record({"fault": "error", "fix": "solution", "outcome": "success"})

Human operator

# Install
pip install socratic-engine vsf-rsi

# Use CLI
python -m socratic_engine evaluate tree.json --context ctx.json
python -m vsf_rsi.scenario_memory record --fault "error" --fix "solution"

For the full three-tool integration pattern, see state-canon-mcp README.

Safety model

The VSM safety constraints are enforced at every level:

flowchart TD
    subgraph Autonomous["Autonomous (no approval)"]
        L1["L1: parameter_drift<br/>adjust thresholds"]
        L2["L2: inject_predicate<br/>create wrappers"]
    end
    
    subgraph HumanApproval["Human Approval Required"]
        L3["L3: generate_predicate<br/>create new predicates"]
        L4["L4: evolve_forest<br/>genetic evolution"]
    end
    
    subgraph NeverAutonomous["Never Autonomous"]
        A0["A0: purpose changes"]
        A5["A5: autonomy ceiling"]
        R17["R17: destructive actions"]
    end
    
    L1 -->|"validate before keeping"| L1
    L2 -->|"validate before keeping"| L2
    L3 -->|"require approval"| L3
    L4 -->|"require approval"| L4

Key safety properties:

• A0/A0.1: Purpose is inalienable — RSI cannot change system purpose
• A5: Autonomy ceiling — L1 is autonomous, L2+ requires validation
• R17: Destructive actions always require explicit human consent
• R4: Trust requires independent check — RSI cannot self-certify

Going deeper

• ROADMAP.md — version history and future plans.
• CHANGELOG.md — what changed, by version.
• docs/rsi_forests.vsm — generated forest populations (VSM notation).
• docs/rsi_generated_predicates.vsm — generated predicates (VSM notation).
• docs/rsi_generated_trees.vsm — generated trees (VSM notation).
• tests/ — 775 tests.

Roadmap

See ROADMAP.md for detailed version history and future plans.

v0.2.5 — Adversarial Benchmarks ✅

• Benchmark framework: load scenarios, run benchmarks, save/load reports, improvement curves
• 4 adversarial scenario generators (prisoner, parábola, XOR 5D, señal+ruido)
• validate_predicate_behavior() — run predicates against test cases, reject below threshold
• Context-aware predicate generation (uses avg_threshold, avg_input, error_class)
• Feature-detect enforce_limits via inspect.signature (backward-compatible)
• CI fix: install socratic-engine before running tests
• 775 tests

v0.3.0 — Production

• Dashboard for observation
• Cross-validation
• Overfitting detection
• GA produces trees with fitness > 0.7 on real benchmark

Lineage & philosophy

Built on Stafford Beer's Viable System Model and the Cybersyn project (Chile, 1971) — a system is viable when it can be described, governed, and audited. "Don't trust, verify" isn't a slogan here; it's the reconciler and the verify-at-every-boundary discipline made mechanical. Community-first, and deliberately from the Global South — the heir to Cybersyn's bet that good cybernetics serves people.

License

Code: Apache-2.0 (LICENSE) · Docs: CC-BY-4.0.

This module is deliberately more permissive than the AGPL core it was extracted from — it is meant to be adopted, embedded, and improved by the community. Improvements can flow back; the module stays clean of AGPL code by construction.