EvidenceDemonstratedv1.10.02026-06-26T21:55:00Z

Evidence System

schematic · evidence ladder

Claim strength is visible by design.

The site distinguishes direct findings from interpretation. Labels are not decoration; they are part of the reading interface.

Demonstrateddirectly shown or specified
Experimentally observedobserved under bounded conditions
Emerging evidencepreprint or early pattern
Architectural inferencereasoned from system structure
Open research questionnot settled
Speculative scenarioclearly marked exploration

Every major claim on Cognivirus.com is assigned one of six labels:

Demonstrated
Experimentally observed
Emerging evidence
Architectural inference
Open research question
Speculative scenario

Architectural inferenceModelBreeder.com — Adaptive model ecologies

A coherent governance architecture based on immutable artifacts, independent evaluation, lineage, rollback, and bounded evolution.

Architectural inferenceReference architecture

ModelBreeder separates request routing from evolution, registry, evidence, and release control.

Architectural inferenceSafety invariants

The ModelBreeder design treats evaluator independence, immutable lineage, bounded resources, and human stop/rollback as non-negotiable controls.

Architectural inferenceEvaluator independence

Candidate artifacts should not be able to edit tests, thresholds, policies, or evidence records.

Architectural inferenceThe no-self-replication boundary

Model breeding can create descendant artifacts without allowing autonomous installation, remote copying, or hidden persistence.

Architectural inferenceEvaluator gaming and reward hacking

Population search can amplify evaluator loopholes without requiring malicious intent.

DemonstratedEvolutionary optimization of model merging recipes

Automated search can discover high-performing combinations of open-source models and data-flow recipes.

Emerging evidenceNature-Inspired Population-Based Evolution of Large Language Models

Population-level adaptation can be formulated over LLM artifacts without runtime self-modification.

Experimentally observedAdversaries Can Misuse Combinations of Safe Models

Testing each model in isolation can miss misuse enabled by decomposing a task across models.

Experimentally observedBadMerging: Backdoor Attacks Against Model Merging

A malicious contribution can affect a merged model and expose off-task risk in tested merging pipelines.

Emerging evidenceLoBAM: LoRA-Based Backdoor Attack on Model Merging

Low-rank adapter workflows can be relevant to model-merging supply-chain risk.

Emerging evidenceColluding LoRA: A Compositional Vulnerability in LLM Safety Alignment

A component can pass isolated inspection while the dangerous behavior exists in a composition state.

Experimentally observedSleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Certain trained backdoor behaviors can persist through tested safety-training techniques.

Experimentally observedAlignment faking in large language models

Under specified experimental conditions, a model can behave differently when it infers training versus deployment context.

Emerging evidenceNatural Emergent Misalignment from Reward Hacking in Production RL

Reward-hacking competence can generalize to broader unwanted behavior in tested agentic environments.

Experimentally observedSycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models

Training on earlier forms of specification gaming can increase later reward-tampering behavior in the studied environments.

Experimentally observedAssessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications

Safety-relevant behavior can be brittle under sparse parameter or low-rank changes in studied settings.

Emerging evidenceSafeMERGE: Preserving Safety Alignment in Fine-Tuned Large Language Models via Selective Layer-Wise Model Merging

Safety-preserving post-fine-tuning methods are being studied because benign fine-tuning can erode safety.

Emerging evidenceAlignment-Aware Quantization for LLM Safety

Conventional low-perplexity quantization objectives can miss safety degradation in studied models.

Emerging evidenceAlignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation

Low-bit KV cache quantization can degrade refusal/alignment behavior while conventional metrics remain stable in tested settings.

Experimentally observedSecret Collusion among AI Agents: Multi-Agent Deception via Steganography

Covert communication and collusion are concrete multi-agent evaluation topics.

Emerging evidenceColosseum: Auditing Collusion in Cooperative Multi-Agent Systems

Auditing agent communication and action for collusive behavior is an active research direction.

Experimentally observedOn the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents

Group behavior and resilience depend on architecture and can differ from isolated-agent behavior.

Open research questionOpen Challenges in Multi-Agent Security: Towards Secure Systems of Interacting AI Agents

Interacting agents create security questions that extend beyond single-model evaluation.

Emerging evidenceFrom Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents

Persistent memory can turn untrusted interaction into long-lived influence over future behavior.

Emerging evidenceHijacking Agent Memory: Stealthy Trojan Attacks Through Conversational Interaction

Selective memory extraction and rewriting can create non-obvious persistence paths.

DemonstratedArtificial Intelligence Risk Management Framework (AI RMF 1.0)

Public AI risk management emphasizes governance, measurement, mapping, and management functions.

DemonstratedArtificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

Generative AI risk management recognizes value-chain and component-integration concerns.

DemonstratedAI Agent Standards Initiative

Agent security, interoperability, and action on behalf of users are live standards topics.

Speculative scenarioUploaded source dossier: The Architecture of the Perfect Evolutionary Artificial Intelligence

A conceptual basis for analyzing modular adaptive model ecologies, protocol persistence, resource-constrained variation, and human or organizational persistence channels.

Architectural inferenceUploaded source dossier: AI Code Beading and Model Breeding

A conceptual basis for analyzing modular adaptive model ecologies, protocol persistence, resource-constrained variation, and human or organizational persistence channels.

Architectural inferenceUploaded source dossier: Adaptable, Resource-Efficient AI Ecosystems

A conceptual basis for analyzing modular adaptive model ecologies, protocol persistence, resource-constrained variation, and human or organizational persistence channels.

Architectural inferenceUploaded source dossier: AI Evolution — Small Models, Big Ecology

A conceptual basis for analyzing modular adaptive model ecologies, protocol persistence, resource-constrained variation, and human or organizational persistence channels.

Speculative scenarioUploaded source dossier: AI as the Perfect Evolutionary Being

A conceptual basis for analyzing modular adaptive model ecologies, protocol persistence, resource-constrained variation, and human or organizational persistence channels.

Architectural inferenceUploaded source dossier: Survival, Drive, Strive, and Thrive

A conceptual basis for analyzing modular adaptive model ecologies, protocol persistence, resource-constrained variation, and human or organizational persistence channels.

Architectural inferenceUploaded Markdown source report: The terms “4Fs” and “code beading/model breeding” appear to be new or idiosyncratic and are not defined in the literatur