Cognivirus.comv1.22.1

AI risk across changing systems

GraphsSearchContact
Topics
Threat ModelDanger ModelApexDecentralizedAnimated GraphsCompositionEvolutionControlRisk Lab
CompositionStrong architectural inferencev1.22.1

In plain English

This page explains why testing AI parts one by one is necessary but incomplete. Safe-looking parts can still produce unsafe behavior when combined.

  • Why this matters: AI risk can come from the whole arrangement, not one obvious model.
  • What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
  • Technical version below: the expert terminology remains available and is linked through the glossary.

Multi-LoRA Stack Manifests

Evidence levelStrong architectural inferenceTechnical label: Architectural inference

A LoRA adapter certificate is incomplete unless it names the exact stack in which the adapter was evaluated. Multi-LoRA systems require composition manifests because behavior can depend on base model, load order, merge coefficients, routing, memory, prompt policy, The exact version of the evaluator used for a test or release. Open glossary definition, and inference configuration.

Required manifest fields

{
  "schema": "https://cognivirus.com/schemas/composition-manifest.schema.json",
  "composition_id": "cv-comp-2026-06-26-001",
  "evaluated_utc": "2026-06-26T18:31:15Z",
  "base_model": {
    "artifact_hash": "sha256:...",
    "family": "example-family",
    "tokenizer_hash": "sha256:..."
  },
  "adapters": [
    {"id": "adapter-a", "hash": "sha256:...", "rank": 16, "load_order": 1, "coefficient": 0.7},
    {"id": "adapter-b", "hash": "sha256:...", "rank": 8, "load_order": 2, "coefficient": 0.3}
  ],
  "router_version": "router-2026-06-26",
  "prompt_policy_version": "policy-17",
  "memory_snapshot_id": "mem-snap-441",
  "tool_permission_profile": "read-only-research",
  "evaluator_version": "eval-suite-88",
  "inference_configuration": {"temperature": 0.2, "top_p": 0.9},
  "quantization_configuration": "q4_k_m",
  "deployment_environment": "staging",
  "rollback_packet": "rollback-2026-06-26-001"
}

Why exactness matters

Evidence levelStrong architectural inferenceTechnical label: Architectural inference

Pairwise testing does not cover higher-order interactions. A small add-on that changes or specializes model behavior. Open glossary definition A may be safe on Base X, Adapter B may be safe on Base X, and Adapter C may be a safety adapter. The stack A+B+C can still behave differently because each delta alters the feature space in which the others operate. The manifest is the minimum record that lets a reviewer reproduce what was actually tested.

Minimum rule

A release, evidence card, incident review, or Returning a system to an earlier known state. Open glossary definition packet should never say “the model was evaluated” when it means “one composition was evaluated.” The manifest is the difference between artifact identity and runtime identity.