Risk LabStrong architectural inferencev1.22.1

In plain English

This page provides local browser worksheets. They help plan reviews; they are not formal safety certifications.

  • 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.

Risk Lab

browser-only review console

Risk Lab

These worksheets help reviewers inspect the system between the models. They do not send entered data to a server, do not certify safety, and do not replace human review.

local browser-side onlybounded planning worksheetstraceable manifest-first reviewno-op valid outcome

Choose the review surface

Open a focused worksheet, or continue to the full console below.

1 · Select a surface2 · Mark what changed3 · Inspect gaps4 · Preserve evidence5 · Keep no-op available
schematic · assurance decay

Evidence decays when the evaluated state stops being the deployed state.

The worksheet logic is visualized as an evidence half-life: every material change pushes some previous claim from current, to review-needed, to stale, to unknown.

Full worksheet console

All tools on this page run in the browser. Entered data is not sent to a server by this implementation. The tools are planning worksheets, not certification methods.

Most Likely Threat Review

Planning worksheet only. It checks conditions for distributed behavior persistence; it does not calculate probability or certify safety.

Transition Graph Danger Review

Planning worksheet only. It checks whether the reviewed system includes the carriers, transitions, reservoirs, action boundaries, traces, and rollback dependencies that can preserve behavior.

Action Boundary and Conduct Firewall Checklist

Planning worksheet only. It checks whether consequential actions are gated outside the model's prompt and memory state.

Observability and Replay Coverage

Planning worksheet only. It checks whether traces are complete enough to reconstruct a decision path while preserving privacy boundaries.

Synthetic Feedback and Collapse Review

Planning worksheet only. It does not claim that all synthetic data is unsafe; it checks unmanaged recursion and diversity loss.

Promotion Rule Pressure Audit

Planning worksheet only. It asks what the system rewards, preserves, routes, and promotes.

Retirement Completeness and Zombie Behavior Review

Planning worksheet only. It checks whether a retired behavior still has active persistence paths.

Site Discovery Readiness Checklist

Planning worksheet only. It does not certify readiness, guarantee indexing or rankings, or validate production security.

Multi-LoRA Apex Threat Envelope

Qualitative worksheet only. It does not calculate real-world probability and does not provide replication instructions.

UAI File Handoff Memory Coverage

Planning worksheet only. It checks whether source intake has been copied, summarized, dispositioned, hashed, and reflected in .uai memory.

Composition Explosion Calculator

Formula: base models × (adapter versions + 1)optional adapters × router policies × prompt policies × memory states × tool profiles × evaluator versions × inference configurations. Not every mathematical combination is operationally valid.

Assurance Decay Worksheet

Planning worksheet only. It is not a formal certification method.

Behavioral Persistence Explorer

This simulates plausible persistence paths. It does not calculate real-world probability.

Ecological Rollback Planner

Evaluator Independence Checklist

Responsibility Diffusion Mapper

Enter an accountable owner for each role. Gaps indicate responsibility diffusion.

These tools are deliberately conservative. They expose where ordinary review language becomes imprecise: a "tested model" may actually mean one base model, one adapter stack, one router version, one A saved state of what the AI system remembers. Open glossary definition, one permission profile, and one evaluator release. Changing any of those pieces can change what the prior evidence means.

The outputs are not risk scores. They are prompts for engineering review, release gates, and incident-preparedness conversations. Use them to identify missing manifests, stale Confidence, backed by evidence, that a system meets safety or governance requirements. Open glossary definition, persistence reservoirs, rollback dependencies, evaluator coupling, and accountability gaps before a deployment decision is treated as routine.

Apex-threat worksheet

Evidence levelStrong architectural inferenceTechnical label: Architectural inference

The Multi-LoRA A proposed Cognivirus review category for systems where adapter reproduction, dynamic composition, persistent memory, adaptive routing, evaluator selection, and incomplete rollback reinforce one another. Open glossary definition worksheet is deliberately qualitative. It asks which reinforcing conditions are present: automated candidate generation, dynamic adapter composition, evaluator coupling, synthetic-data feedback, persistent memory, adaptive routing, tool permissions, incomplete rollback, and unknown lineage. Its result is not a certification score. It is a triage prompt for deciding whether ordinary component review is inadequate.

UAI File Handoff Memory Coverage

Evidence levelDemonstrated real incidentTechnical label: Demonstrated

This worksheet checks whether active source intake has been preserved, summarized, dispositioned, hashed, and reflected in .uai hot and long-term memory. It is a planning tool, not a UAIX certification result.

publication-integrity readiness

The Site Discovery Readiness Checklist reviews discovery files, metadata, manifests, source integrity, structured-data parity, and The decision not to change the system. Open glossary definition behavior.

Most likely threat review

Evidence levelStrong architectural inferenceTechnical label: Architectural inference

The Most Likely Threat Review Worksheet checks whether distributed behavior persistence conditions are present: composition-only expression, memory residue, A system that judges whether an AI output or candidate is acceptable. Open glossary definition pressure, router amplification, descendant inheritance, incomplete rollback, and human workflow capture.

v1.15 danger-model worksheets

v1.21.9 dashboard review worksheet

v1.22.1 decentralized persistence worksheet

v1.22.0 ModelBreeder risk-side worksheet

v1.22.0 ModelBreeder risk worksheet