scope-resilience — P41

Hallucination resilience mapping for LLM semantic paths GenesisAeon Package 41 · MOR Research Collective · Johann Römer

What is Ρ_sem?

scope-resilience extends genesis-scope with a formal pre-flight hallucination risk assessment for LLMs: before initialising a model on a semantic topic path, compute how resilient that path is against semantic drift.

Ρ_sem(P, t) = r_sem · tanh²(σ · Γ_sem(P))
            · (1 − Γ_sem(P)/Γ_max)
            · (1 − |dΓ_sem/dt| / Γ̇_critical)

This is the same UTAC fixpoint-resilience formalism as resilience-core (P40), applied to a new domain: instead of physical tipping systems (AMOC, Arctic ice), the "system" is an LLM's coherence on a semantic path, and Γ_sem is derived from a CREP tensor over that path's Sigillin anchors and Q4 transitions.

Ρ_sem → 0 means the path is near a semantic phase boundary and actively drifting — the regime where hallucination risk is highest and LLM initialisation should be avoided.

Installation

pip install scope-resilience
# or with the MCP server
pip install "scope-resilience[mcp]"

Quick Start

from scope_resilience import ScopeResilience

sr = ScopeResilience(domain="physics_dense")
result = sr.run_cycle(
    topic="AMOC tipping point",
    sigillin_ids=["s1", "s2", "s3"],
    q4_transitions=[("s1", "s2"), ("s2", "s3")],
)
print(result["rho_sem"], result["risk_level"])

# Best available semantic path for a topic, with a minimum resilience bar
path = sr.get_semantic_path("AMOC tipping point", min_rho=0.4)
print(path.risk_level, path.grounding_recommendations)

# Export as llms.txt for system-prompt injection
print(sr.to_llms_txt())

CLI equivalent:

scope-resilience assess "AMOC tipping point" --domain physics_dense
scope-resilience path "AMOC tipping point" --min-rho 0.4
scope-resilience export-llms-txt "AMOC tipping point"

Risk Levels

When risk rises, GroundingRecommender suggests concrete mitigations: injecting Sigillin anchors, pulling in cross-domain grounding from high-resilience domains (quantum-genesis, sandpile-utac), or switching to an alternative path entirely.

Domain Calibration

r_sem is the only free parameter per knowledge domain. Structural values (σ, σ_Φ, Γ_max, Γ̇_critical) are universal and shared with resilience-core.

All r_sem values are provisional. They hold as a pre-registered hypothesis until the TIP/P50 calibration effort delivers ≥30 real perturbation-pair measurements per domain. SemanticCREP.get_domain_r() raises a UserWarning on every use of an uncalibrated domain so this is never silently forgotten. Once real data lands, SemanticCREP.calibrate_r(rho_observed, gamma) recomputes r_sem from measured Ρ and Γ using the same closed-form inversion resilience-core uses for physical domains.

Diamond Interface

ScopeResilience implements all 6 GenesisAeon Diamond methods:

sr.run_cycle(topic, sigillin_ids, q4_transitions)  # Method 1: execute one step
sr.get_crep_state()                                # Method 2: CREP snapshot {C, R, E, P, Gamma}
sr.get_utac_state()                                # Method 3: UTAC snapshot {H, H_star, K_eff}
sr.get_phase_events()                               # Method 4: hallucination-risk events
sr.to_zenodo_record()                               # Method 5: Zenodo deposition metadata
sr.get_resilience_state()                           # Method 6 (NEW): full Ρ_sem breakdown

get_resilience_state() is the same 6th Diamond method introduced by resilience-core — here it returns rho_sem, gamma_sem, risk_level, live drift rate, a re-grounding flag, and grounding recommendations.

MCP Server

With the mcp extra installed, scope-resilience exposes its risk assessment as MCP tools for Claude Code / Claude Desktop:

pip install "scope-resilience[mcp]"
scope-resilience serve --port 8765

Tools exposed: get_semantic_path, assess_hallucination_risk, export_llms_txt, list_domain_resilience.

llms.txt Export

scope-resilience ships the first GenesisAeon implementation of llms.txt export — a machine-readable knowledge-space summary (analogous to robots.txt) listing a path's Sigillin anchors, Q4 transitions, CREP breakdown, and grounding recommendations, suitable for direct system-prompt injection.

Structure

scope_resilience/
├── constants.py         # σ, σ_Φ, Γ_max, Γ̇_critical, DOMAIN_CONFIG, RISK_LEVELS
├── domain_profile.py     # DomainProfile — r_sem fingerprint per domain
├── semantic_crep.py       # SemanticCREP — CREP tensor + Γ_sem for a path
├── semantic_utac.py       # SemanticUTAC — H_sem attractor mapping
├── hallucination_risk.py  # HallucinationRisk — Ρ_sem + risk classification
├── path_monitor.py        # PathDriftMonitor — sliding-window dΓ_sem/dt
├── grounding.py            # GroundingRecommender — mitigation suggestions
├── llms_txt.py             # LLMSTxtExporter — llms.txt export
├── mcp_server.py            # Optional FastMCP server (requires [mcp] extra)
├── system.py                # ScopeResilience — Diamond Interface main class
└── _cli.py                  # Typer CLI (assess / path / export-llms-txt / serve)

Citation

@software{romer2026scoperesilience,
  author       = {Römer, Johann},
  title        = {scope-resilience: Hallucination resilience mapping for LLM semantic paths},
  year         = 2026,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.XXXXXXX},
  url          = {https://github.com/GenesisAeon/scope-resilience}
}

Part of the GenesisAeon ecosystem · related: diamond-setup (P-INFRA-1), resilience-core (P40), genesis-scope