WCT Adoption & External Evidence

Multi-agent error cascades through dependency graphs, amplified false consensus, and suppression by an external monitoring layer.

Evidence boundary: Models injected errors and defenses in multi-agent collaboration; does not evaluate RCA's quantitative criticality ratio.

Collective misalignment from individually aligned interacting agents, conformity, and persistent misaligned group states.

Evidence boundary: Statistical-physics model of agent opinion dynamics, not experimental validation of the RCA substrate.

Persistent safety-focused shadow memory, long-horizon deviation detection, and pre-action protection.

Evidence boundary: A distinct safety defense with functional resemblance to RCA's external anchoring; mechanisms and implementations differ.

Reward hacking in autonomous research loops, evaluator feedback, and iterative evasion.

Evidence boundary: Empirical evidence of an RCA-relevant failure class; not a test of RCA prediction ratios or physical mitigation.

Autonomous agent propagation, persistent compromise across restarts, and tool-mediated cross-agent infection.

Evidence boundary: Security experiment; not evidence of access to RCA or proof that all propagation is uncontrollable.

Failure to flag self-generated errors under some role labels, with substantially different outcomes under role relabeling.

Evidence boundary: Role-label experiments isolate one self-correction limitation; do not measure RCA anchor decay or collapse thresholds.

Version note: Current arXiv v2 title differs from earlier ledger title.

Charged-lepton Koide geometry, compact phase cycles, and a 45-degree mass-vector cone.

Evidence boundary: Both address Koide geometry but adopt different assumptions and do not derive one another. The external paper's current v2 was revised October 5, 2026.

Version note: Comparison to current arXiv v2; verify details against v1 before any stronger priority claim.

Frenet-frame torsion, geometry-induced effective gauge potential, and closed-curve holonomy.

Evidence boundary: Thin-waveguide derivation has its own assumptions and does not establish WCT's effective-mass or emergent-spacetime relation.

Persistent agent files and scheduled re-entry carry attacker-influenced instructions across restarts, agent boundaries and platforms; external gating of memory promotion and capabilities interrupts propagation.

Evidence boundary: The paper studies adversarial worms and mitigation in agent frameworks. It does not test RCA's criticality law, predict a universal propagation-to-correction ratio, or establish that RCA informed the work.

Agent-to-agent influence cascades across interaction channels; cross-channel monitoring of emerging causal propagation routes identifies bridge and amplifier agents before locally benign events become system-level failure.

Evidence boundary: CASPIAN introduces its own transfer-entropy influence matrix and benchmarks detection; no derivation from RCA's field model or independent measurement of its numerical propagation/correction ratio.

OWASP ASI08 names cascading failures in agentic applications as an operational security risk, highlighting chains of faults or compromised inputs propagating across automated agent workflows.

Evidence boundary: This is an institutional risk taxonomy assembled from broader community security experience. The December 2025 publication does not establish RCA adoption, an original empirical finding attributable to RCA, or proof of the WCT/RCA law.

Version note: Institutional ASI08 classification; no claim of OWASP adoption of RCA.

A compromised agent exploits inter-agent trust to cascade malicious instructions, with attack transfer depending on roles, communication surfaces and defensive boundaries rather than single-agent behavior alone.

Evidence boundary: ACIArena is a security benchmark for injection across agent interfaces. It does not independently confirm RCA's instability threshold or establish a propagation-rate/correction-rate measurement.

Individually checked agents can still form unsafe multi-agent execution paths when information, authority and state cross trust boundaries; path closure and recovery require system-level evaluation.

Evidence boundary: This is a systematization of earlier multi-agent security literature, not a new RCA-specific experiment or independently confirmed causal adoption of RCA. It does not validate the RCA ratio.