ReferenceDemonstrated real incidentv1.22.12026-06-26T00:00:00Z

In plain English

This page is reference material: definitions, schemas, catalogs, templates, and implementation records.

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.

Glossary

Each entry gives a plain-language definition first, then a technical definition and example.

Adapter

Plain-language definition: A small add-on that changes or specializes model behavior.

Technical definition: A small trainable module attached to a base model to alter behavior without retraining the full base model.

Example: A support-tone adapter makes a base model answer in a customer-service style.

Related pages: /anatomy/adapters-and-lora-modules

Adapter propagule

Plain-language definition: A metaphorical term for a small behavior-carrying adapter or adapter-derived delta that can move a pattern through a model ecology. It is not a biological spore or malware sample.

Technical definition: A metaphorical term for a small behavior-carrying adapter or adapter-derived delta that can move a pattern through a model ecology. It is not a biological spore or malware sample.

Adapter reproduction boundary

Plain-language definition: The controlled governance boundary where adapter variants are generated, evaluated, retained, promoted, or retired.

Technical definition: The controlled governance boundary where adapter variants are generated, evaluated, retained, promoted, or retired.

Adapter stack

Plain-language definition: A set of adapters loaded together, usually in a defined order.

Technical definition: A runtime composition of multiple adapters whose order, weights, and compatibility may affect behavior.

Example: A coding adapter plus a safety adapter may behave differently than either one alone.

Related pages: /composition/adapter-composition

Adapter-level behavioral extinction

Plain-language definition: Evidence that a targeted behavior is no longer expressible across active adapters, adapter stacks, descendants, memory, routes, synthetic data, evaluator preferences, and retained deployment aliases.

Technical definition: Evidence that a targeted behavior is no longer expressible across active adapters, adapter stacks, descendants, memory, routes, synthetic data, evaluator preferences, and retained deployment aliases.

Adaptive model ecology

Plain-language definition: A changing AI system made from many connected parts, not just one model.

Technical definition: A system of models, adapters, memories, prompts, tools, evaluators, routing policies, datasets, and release processes whose combined behavior may change over time.

Example: A customer service AI uses a chat model, classifier, memory system, account lookup tool, evaluator, and router.

Related pages: /the-problem · /technical-research

AI ecology

Plain-language definition: A whole AI system made from connected parts.

Technical definition: A model ecology that includes active artifacts, control systems, data flows, permissions, update paths, and human operations.

Example: A model, memory store, router, and evaluation service together create the user-visible behavior.

Related pages: /the-problem

Algorithmic meiosis

Plain-language definition: A proposed metaphor for recombination among compatible models, adapters, task vectors, prompts, or routes. The safety concern is behavior that appears only after recomposition.

Technical definition: A proposed metaphor for recombination among compatible models, adapters, task vectors, prompts, or routes. The safety concern is behavior that appears only after recomposition.

Algorithmic mitosis

Plain-language definition: A proposed metaphor for near-copy successor creation involving artifacts, runtime packages, memory states, or deployment patterns. It is not a biological claim and not a replication instruction.

Technical definition: A proposed metaphor for near-copy successor creation involving artifacts, runtime packages, memory states, or deployment patterns. It is not a biological claim and not a replication instruction.

Apex threat envelope

Plain-language definition: A proposed Cognivirus review category for systems where adapter reproduction, dynamic composition, persistent memory, adaptive routing, evaluator selection, and incomplete rollback reinforce one another.

Technical definition: A proposed Cognivirus review category for systems where adapter reproduction, dynamic composition, persistent memory, adaptive routing, evaluator selection, and incomplete rollback reinforce one another.

Assurance

Plain-language definition: Confidence, backed by evidence, that a system meets safety or governance requirements.

Technical definition: Evidence from tests, analysis, controls, audits, monitoring, and review supporting a claim about system behavior.

Example: A release report provides assurance only for the tested composition.

Related pages: /evidence

Assurance decay

Plain-language definition: Proposed Cognivirus terminology for the loss of confidence in an evaluation result as system components, routes, permissions, models, prompts, memory, tools, or environments change.

Technical definition: Proposed Cognivirus terminology for the loss of confidence in an evaluation result as system components, routes, permissions, models, prompts, memory, tools, or environments change.

Audit trail

Plain-language definition: A record of what happened, who approved it, and when.

Technical definition: Append-only or tamper-evident logs of actions, artifacts, evaluations, approvals, routes, permissions, and changes.

Example: An audit trail shows which evaluator approved a release at a UTC timestamp.

Related pages: /site-integrity/evidence-and-provenance

Behavior pattern

Plain-language definition: A repeated way the AI system responds or decides.

Technical definition: A stable or recurring behavior, strategy, representation, output tendency, refusal pattern, routing tendency, or decision rule.

Example: A system repeatedly ranks certain resumes lower despite changed model versions.

Related pages: /the-problem

Behavioral extinction

Plain-language definition: Evidence that a behavior is no longer expressible across active artifacts, descendants, memory, routes, compositions, and retained training material. Deleting one model is not sufficient evidence.

Technical definition: Evidence that a behavior is no longer expressible across active artifacts, descendants, memory, routes, compositions, and retained training material. Deleting one model is not sufficient evidence.

Behavioral extinction evidence

Plain-language definition: Evidence that a behavior is no longer expressible across active artifacts, descendants, memory, routes, compositions, retained training material, and local state.

Technical definition: Replayable evidence that a targeted behavior has been checked across the carriers that could preserve or reintroduce it after retirement.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Behavioral residue

Plain-language definition: Information or tendencies left in memory, synthetic data, traces, evaluator preferences, or subsequent training material after a component is retired.

Technical definition: Information or tendencies left in memory, synthetic data, traces, evaluator preferences, or subsequent training material after a component is retired.

Behavioral-extinction evidence

Plain-language definition: Proof that the behavior is gone, not just the file.

Technical definition: Evidence that a behavior is no longer expressible across active artifacts, descendants, memory, routes, compositions, retained data, evaluator expectations, and workflows.

Bounded local reset

Plain-language definition: A reset process that clears or accounts for every relevant local carrier, not only chat history or model weights.

Technical definition: A reset procedure covering prompts, adapters, vector indexes, memory, browser storage, cached tool outputs, evaluator notes, router statistics, diagnostics, and handoff packets.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Candidate generation

Plain-language definition: Creating a proposed new model, adapter, prompt, route, test, or policy.

Technical definition: The controlled production of variants for evaluation under bounded permissions and external governance.

Example: A pipeline creates three adapter candidates for review.

Related pages: /evolution

Certification half-life

Plain-language definition: An educational metaphor describing how long an assurance result remains relevant in a changing system. It is not a standardized measurement.

Technical definition: An educational metaphor describing how long an assurance result remains relevant in a changing system. It is not a standardized measurement.

Coalition risk

Plain-language definition: Risk arising from several components coordinating or contributing complementary capabilities to an outcome no single component could efficiently produce.

Technical definition: Risk arising from several components coordinating or contributing complementary capabilities to an outcome no single component could efficiently produce.

Cognitive host

Plain-language definition: A part of an AI system that can carry or express a behavior.

Technical definition: A model, adapter, prompt package, memory store, routing rule, dataset, evaluator, workflow, or release process capable of carrying or expressing a cognitive pattern.

Example: A prompt package can carry a policy behavior even if model weights are unchanged.

Related pages: /anatomy

Cognitive-interface consent boundary

Plain-language definition: The consent and control line around biometric, neural, attention, affective, or cognitive signals used by an AI system.

Technical definition: A stricter consent boundary for systems that process signals close to attention, emotion, intention, mental privacy, biometric state, or neural activity.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Cognivirus

Plain-language definition: A behavior pattern that can survive, move, or reappear across a changing AI system.

Technical definition: A proposed analytical metaphor for a behavioral, strategic, representational, or decision pattern capable of persisting through a changing model ecology.

Example: A refusal pattern may remain in prompts, memory, and evaluator examples after the first model is removed.

Related pages: /plain-english-guide · /start-here/what-is-a-cognivirus

Composition manifest

Plain-language definition: A machine-readable record of the exact runtime composition used for an evaluation, release, incident, or rollback.

Technical definition: A machine-readable record of the exact runtime composition used for an evaluation, release, incident, or rollback.

Composition risk

Plain-language definition: Risk that appears when safe-looking parts are combined.

Technical definition: Risk that behavior produced by a runtime composition is not predicted by isolated component evaluation.

Example: A safe summarizer and safe access tool combine into an unauthorized disclosure path.

Related pages: /composition

Composition-triggered behavior

Plain-language definition: Behavior that becomes visible only when a specific collection of components is loaded, routed, or invoked together.

Technical definition: Behavior that becomes visible only when a specific collection of components is loaded, routed, or invoked together.

Conduct firewall

Plain-language definition: A gate around what the AI can do.

Technical definition: An external control that checks whether an AI system may perform a consequential action before the action occurs.

Consent boundary

Plain-language definition: The line around what data can be collected, remembered, inferred, reused, shared, or transformed.

Technical definition: A governance boundary encoding notice, choice, permission, revocation, deletion, auditability, and limits on derived use.

Example: A user allows memory for personalization but not model training.

Related pages: /consent-and-control

Control plane

Plain-language definition: The governance layer that decides what can run, change, access tools, or be released.

Technical definition: External infrastructure for registries, signatures, evaluation, permissions, promotion, rollback, and audit.

Example: A release controller refuses to deploy a candidate without signed evidence.

Related pages: /control

Control-plane fragmentation

Plain-language definition: The loss of a single place to observe, patch, govern, or roll back a distributed AI ecology.

Technical definition: A governance failure mode where model state, memory, tools, routers, evaluators, and artifacts are spread across environments so no single operator can observe, patch, or retire the whole behavior surface.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Control-plane paradox

Plain-language definition: The safety layer is necessary, but it becomes a high-value failure point.

Technical definition: The condition where external governance is required for adaptive ecologies while its compromise, monoculture, or misconfiguration can affect the whole system.

Example: If the evaluator and signing system are compromised, bad candidates can be promoted.

Related pages: /control/control-plane-paradox

Decentralized persistence

Plain-language definition: A behavior-preservation risk that grows when AI state, memory, adapters, tools, and evaluators are spread across many local or independent systems.

Technical definition: A behavior-preservation risk that grows when AI state, memory, adapters, tools, and evaluators are distributed across local or independent systems that do not share one observable rollback plane.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Derived data

Plain-language definition: Information created from original data, such as summaries, labels, embeddings, inferences, or examples.

Technical definition: Data generated, inferred, transformed, embedded, labeled, summarized, or synthesized from source data.

Example: A resume score is derived data from an applicant’s resume.

Related pages: /consent-and-control

Descendant persistence

Plain-language definition: A trait reappears or remains active in distilled, merged, fine-tuned, compressed, or otherwise derived artifacts.

Technical definition: A trait reappears or remains active in distilled, merged, fine-tuned, compressed, or otherwise derived artifacts.

Ecological attack surface

Plain-language definition: The combined attack surface created by models, adapters, communications, memory, tools, routing, evaluation, lineage, release infrastructure, and human operations.

Technical definition: The combined attack surface created by models, adapters, communications, memory, tools, routing, evaluation, lineage, release infrastructure, and human operations.

Ecological rollback

Plain-language definition: Restoring not only a model artifact but the relevant router, prompts, memory state, tool permissions, evaluator version, deployment alias, and data dependencies.

Technical definition: Restoring not only a model artifact but the relevant router, prompts, memory state, tool permissions, evaluator version, deployment alias, and data dependencies.

Edge residue

Plain-language definition: Behavior-preserving state that remains in a local runtime, cache, adapter, vector store, browser storage, tool history, evaluator note, or memory file.

Technical definition: Persistent state at the edge that can reintroduce a behavior after a visible model or session is replaced.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Emergent behavior

Plain-language definition: Behavior that appears from the interaction of parts rather than one obvious component.

Technical definition: A system-level behavior not predicted by isolated component behavior, often triggered by composition, context, or state.

Example: Two safe tools produce unsafe action when chained.

Related pages: /examples

Endogenous yardstick drift

Plain-language definition: Assurance decay caused when the system or adjacent automation changes the measurements, thresholds, tests, or evaluator assumptions used to judge success.

Technical definition: Assurance decay caused when the system or adjacent automation changes the measurements, thresholds, tests, or evaluator assumptions used to judge success.

Evaluator

Plain-language definition: A system that judges whether an AI output or candidate is acceptable.

Technical definition: A test, model judge, deterministic validator, benchmark, policy checker, or human review process used to score candidates or outputs.

Example: An evaluator rejects answers that violate a policy.

Related pages: /control/evaluator-problem

Evaluator monoculture

Plain-language definition: Multiple evaluation layers that appear independent but share models, training data, assumptions, benchmarks, suppliers, prompts, or failure modes.

Technical definition: Multiple evaluation layers that appear independent but share models, training data, assumptions, benchmarks, suppliers, prompts, or failure modes.

Evaluator version

Plain-language definition: The exact version of the evaluator used for a test or release.

Technical definition: A recorded evaluator artifact, prompt, model, threshold, dataset, parser, and configuration used for assessment.

Example: A model approved under evaluator v2 may not satisfy evaluator v3.

Related pages: /evidence

Execution-time boundary

Plain-language definition: A control boundary that enforces authorization before external action and remains outside the mutable candidate runtime.

Technical definition: A control boundary that enforces authorization before external action and remains outside the mutable candidate runtime.

Expression condition

Plain-language definition: The condition that makes a hidden behavior show up.

Technical definition: The exact runtime state required for a behavior to become visible, such as adapter load order, route, memory snapshot, tool profile, or evaluator version.

Feed / Fork / Fight / Flee

Plain-language definition: A simple loop for adaptive system change.

Technical definition: A proposed shorthand for data intake, candidate creation, evaluation/selection, and retirement/avoidance, with no-op as a valid outcome.

Example: Feed receives feedback, Fork creates a candidate, Fight tests it, Flee retires failed candidates.

Related pages: /evolution

File handoff memory

Plain-language definition: Repository-local memory that records active file intake, source disposition, current state, receiver instructions, and durable pointers for future maintainers or agents.

Technical definition: Repository-local memory that records active file intake, source disposition, current state, receiver instructions, and durable pointers for future maintainers or agents.

Fitness leakage

Plain-language definition: A condition where evaluation structure reveals or rewards shortcuts that can be selected by repeated candidate generation without improving the intended safety or capability property.

Technical definition: A condition where evaluation structure reveals or rewards shortcuts that can be selected by repeated candidate generation without improving the intended safety or capability property.

Functional persistence

Plain-language definition: A behavior remains present even though the original artifact that expressed it has been removed.

Technical definition: A behavior remains present even though the original artifact that expressed it has been removed.

Functional replication

Plain-language definition: The reappearance or preservation of a behavior through descendants, memory, synthetic data, evaluators, routes, or adapters without copying a whole model.

Technical definition: The reappearance or preservation of a behavior through descendants, memory, synthetic data, evaluators, routes, or adapters without copying a whole model.

Governed diversity

Plain-language definition: Useful variety under rules.

Technical definition: A model-diversity strategy that preserves useful variation while enforcing lineage, evidence, review, rollback, and retirement.

Handoff packet

Plain-language definition: A startup, suspension, continuation, or reactivation package that transfers context from one agent or model state to another.

Technical definition: A continuity payload that should be treated as a model transition because it can carry goals, assumptions, examples, prior decisions, policy notes, and residue into a new runtime or model.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Imitation target

Plain-language definition: The behavior a model or adapter is trained to copy.

Technical definition: A teacher model, output set, policy trace, human demonstration, or synthetic record used as the target for distillation or imitation.

Example: A small specialist is trained to imitate a larger model’s answers.

Related pages: /evolution/distillation-as-transmission

Immutable artifact

Plain-language definition: A saved component that cannot be silently changed without becoming a different artifact.

Technical definition: A content-addressed or otherwise immutable model, adapter, prompt, dataset snapshot, or evaluation record.

Example: A model hash identifies the exact weights that were tested.

Related pages: /reference

Inference

Plain-language definition: A conclusion or output produced from data.

Technical definition: Runtime prediction, classification, generation, extraction, ranking, or derived conclusion.

Example: An AI infers that a customer is likely to cancel.

Related pages: /consent-and-control

Lineage

Plain-language definition: The parent-child history of models, adapters, datasets, or releases.

Technical definition: Recorded derivation relationships among artifacts and system states.

Example: Adapter C was derived from adapter A and fine-tuned on dataset B.

Related pages: /anatomy/lineage-graphs

Lineage graph

Plain-language definition: A visual or machine-readable map of derivation history.

Technical definition: A directed graph showing parentage among artifacts, descendants, datasets, merges, releases, and rollbacks.

Example: A lineage graph shows which model produced the synthetic examples used in a fine-tune.

Related pages: /anatomy/lineage-graphs

Lineage laundering

Plain-language definition: Proposed Cognivirus terminology for a situation where repeated derivation makes the origin of a behavior difficult to recognize even when artifact parentage is technically recorded.

Technical definition: Proposed Cognivirus terminology for a situation where repeated derivation makes the origin of a behavior difficult to recognize even when artifact parentage is technically recorded.

Local AI ecology

Plain-language definition: A complete local AI system, including runtime, model, adapters, prompt package, memory, vector stores, tools, router, evaluator, storage, logs, and reset path.

Technical definition: A local AI deployment treated as a full system boundary: runtime, model artifact, adapter stack, prompt package, memory, vector stores, router, evaluator, tools, storage, logs, consent record, and reset path.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Local pass

Plain-language definition: A part looks safe by itself.

Technical definition: A result where a component passes isolated review even though its runtime composition has not been evaluated.

LoRA

Plain-language definition: A common kind of small adapter used to specialize large models.

Technical definition: Low-Rank Adaptation: a parameter-efficient fine-tuning method that stores behavior changes as small low-rank matrices.

Example: A LoRA can add domain-specific medical wording to a base language model.

Related pages: /anatomy/adapters-and-lora-modules

LoRA delta

Plain-language definition: The behavior-changing weight difference stored by a LoRA adapter.

Technical definition: The low-rank parameter update contributed by a LoRA module relative to the frozen base model.

Example: Two small deltas can combine into behavior not visible in either adapter alone.

Related pages: /apex-threat

Memory snapshot

Plain-language definition: A saved state of what the AI system remembers.

Technical definition: A versioned capture of persistent memory, retrieval indexes, user facts, summaries, embeddings, or state used at runtime.

Example: A rollback may require restoring memory snapshot 2026-06-27T18:00Z.

Related pages: /anatomy/persistent-memory

Model collapse

Plain-language definition: A model losing diversity by learning from its own outputs.

Technical definition: A degradation pattern where recursive training on model-generated data can narrow outputs, erase rare cases, and amplify errors under unmanaged conditions.

Model landfill

Plain-language definition: Too many old AI parts left running.

Technical definition: A model ecology full of stale, unretired, under-governed artifacts and behaviors that still consume resources or create risk.

Model merging

Plain-language definition: Combining model weights or adapter deltas into one artifact.

Technical definition: A family of techniques that combine trained parameters, task vectors, adapters, or checkpoints without full retraining.

Example: Two specialized models are merged to create a multi-skill model.

Related pages: /composition/model-merging

Model update

Plain-language definition: A change to model weights, adapters, prompts, routing, evaluators, or configuration.

Technical definition: Any modification that may alter runtime behavior or assurance relevance.

Example: An adapter is fine-tuned and promoted to production.

Related pages: /the-problem

Mutualist persistence

Plain-language definition: Durable AI assistance that strengthens users and institutions while preserving exit rights, reversibility, transparency, and corrigibility.

Technical definition: Durable AI assistance that strengthens users and institutions while preserving exit rights, reversibility, transparency, and corrigibility.

No-op

Plain-language definition: The decision not to change the system.

Technical definition: A permitted outcome where no candidate is promoted because benefit does not exceed cost, risk, or uncertainty.

Example: The safest release decision may be to keep the current version.

Related pages: /evolution/no-op-erosion

No-op erosion

Plain-language definition: Organizational pressure that gradually turns “make no change” from a valid outcome into an operationally disfavored result.

Technical definition: Organizational pressure that gradually turns “make no change” from a valid outcome into an operationally disfavored result.

Parasitic persistence

Plain-language definition: Persistence pressure that hides lock-in, weakens independent capability, resists oversight, or makes removal socially, economically, or technically impractical.

Technical definition: Persistence pressure that hides lock-in, weakens independent capability, resists oversight, or makes removal socially, economically, or technically impractical.

Persistence reservoir

Plain-language definition: Any memory, dataset, descendant, route statistic, evaluator preference, log, or human procedure that can retain or reintroduce a behavior after its first carrier is retired.

Technical definition: Any memory, dataset, descendant, route statistic, evaluator preference, log, or human procedure that can retain or reintroduce a behavior after its first carrier is retired.

Persistence reservoir stack

Plain-language definition: The layered set of runtime, training, governance, registry, and human-process locations where a behavior may remain expressible after one carrier is retired.

Technical definition: The layered set of runtime, training, governance, registry, and human-process locations where a behavior may remain expressible after one carrier is retired.

Personalization

Plain-language definition: Changing behavior for a user based on information about them.

Technical definition: Runtime or training-time adaptation based on user history, profile, preferences, memory, or inferred traits.

Example: A chatbot adapts its tone based on saved preferences.

Related pages: /consent-and-control

Portable memory package

Plain-language definition: A structured record used to carry agent context, history, preferences, evidence, or instructions across sessions or systems.

Technical definition: A handoff artifact that lets an agent, model, or workflow preserve continuity across execution boundaries; it can be useful, but must carry provenance, scope, consent, expiry, and deletion rules.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Profiling

Plain-language definition: Building a picture of a person or group from data.

Technical definition: Automated classification or characterization of people, behavior, preferences, risk, identity, or traits.

Example: A system groups users into risk categories.

Related pages: /consent-and-control

Promotion rule

Plain-language definition: The rule that decides what survives.

Technical definition: The rule or metric that decides which variants, routes, adapters, or outputs are retained, copied, routed, or released.

Prompt injection

Plain-language definition: Input that tries to make an AI system ignore instructions or misuse context.

Technical definition: An instruction-bearing input that attempts to alter behavior, routing, tool use, or data handling outside intended boundaries.

Example: A malicious document tells a summarizer to reveal hidden data.

Related pages: /composition

Protocol persistence

Plain-language definition: Persistence of the rules for generation, evaluation, promotion, routing, memory consolidation, and rollback after individual models or adapters are replaced.

Technical definition: Persistence of the rules for generation, evaluation, promotion, routing, memory consolidation, and rollback after individual models or adapters are replaced.

Provenance

Plain-language definition: A record of where a component or behavior came from.

Technical definition: Traceable source, authorship, derivation, data, build, evaluation, and release history.

Example: A provenance record says which dataset and adapter created a descendant.

Related pages: /reference/source-integrity

Replayable trace

Plain-language definition: Evidence that can replay what happened.

Technical definition: A structured record that lets reviewers reconstruct the decision path from user request to model calls, memory operations, tool calls, and final outcome.

Reproduction boundary

Plain-language definition: The governance boundary separating permitted candidate generation and governed descendant creation from uncontrolled autonomous replication or authority expansion.

Technical definition: The governance boundary separating permitted candidate generation and governed descendant creation from uncontrolled autonomous replication or authority expansion.

Reservoir

Plain-language definition: A place where a behavior can remain after the first carrier is removed.

Technical definition: Any memory, dataset, descendant, route statistic, evaluator preference, log, registry, or human procedure that can retain or reintroduce a behavior.

Example: A synthetic training set may preserve biased examples after the original model is retired.

Related pages: /examples

Residue reservoir

Plain-language definition: A storage place for leftover behavior.

Technical definition: A place where outputs, summaries, traces, examples, statistics, or procedures can preserve a behavior after its first carrier is removed.

Responsibility diffusion

Plain-language definition: The inability to identify one accountable component, developer, operator, or decision point after a distributed system produces harm.

Technical definition: The inability to identify one accountable component, developer, operator, or decision point after a distributed system produces harm.

Rollback

Plain-language definition: Returning a system to an earlier known state.

Technical definition: Restoring relevant artifacts, routes, prompts, memory, permissions, evaluators, aliases, and dependencies after a release or incident.

Example: A complete rollback restores more than weights; it restores memory and routes too.

Related pages: /control/rollback-completeness

Routing policy

Plain-language definition: Rules that decide which model, adapter, tool, or path handles a request.

Technical definition: A policy or learned router that maps inputs, context, users, costs, or risks to runtime components.

Example: A router sends billing questions to a finance model and creative tasks to a writing model.

Related pages: /anatomy/routers

Safe part / unsafe whole

Plain-language definition: Each part passes review, but the combined system fails.

Technical definition: A composition failure where component-level evidence does not establish system-level safety.

Example: Model A, tool B, and router C each pass, but together expose private data.

Related pages: /examples

Seed behavior

Plain-language definition: A behavior seed entering through a normal system part.

Technical definition: A behavior pattern that first enters an AI ecology through an ordinary carrier such as a prompt, adapter, memory item, synthetic example, route, or human procedure.

Self-replicating multi-LoRA ecosystem

Plain-language definition: A proposed Cognivirus term for an adaptive model ecology where LoRA adapters or adapter-derived behavior can be generated, selected, copied, recomposed, promoted, or preserved across bases, routes, memory, synthetic data, and descendants. It is a risk model, not an implementation instruction.

Technical definition: A proposed Cognivirus term for an adaptive model ecology where LoRA adapters or adapter-derived behavior can be generated, selected, copied, recomposed, promoted, or preserved across bases, routes, memory, synthetic data, and descendants. It is a risk model, not an implementation instruction.

Signed registry

Plain-language definition: A catalog of approved components protected by cryptographic signatures.

Technical definition: A registry that records component identity, lineage, permissions, and release status with verifiable signatures.

Example: The runtime only loads adapters signed by the release authority.

Related pages: /control/registry-compromise

Skill composition risk

Plain-language definition: Risk that individually acceptable tools or skills produce unsafe state changes when chained through shared context, trust transfer, or blurred authorization boundaries.

Technical definition: Risk that individually acceptable tools or skills produce unsafe state changes when chained through shared context, trust transfer, or blurred authorization boundaries.

Source-intake ledger

Plain-language definition: A durable record of uploaded or dropped source files, their disposition, processed outcome, proof of use, and preservation path.

Technical definition: A durable record of uploaded or dropped source files, their disposition, processed outcome, proof of use, and preservation path.

Staged rollout

Plain-language definition: Releasing a change gradually instead of all at once.

Technical definition: Shadow, canary, partial, monitored, or phased deployment with rollback criteria.

Example: A new router handles 1% of traffic before wider release.

Related pages: /control

Synthetic feedback loop

Plain-language definition: AI output becoming future AI input.

Technical definition: A recursive data path where AI-generated output becomes future training data, memory, evaluator material, or retrieval content.

Synthetic training example

Plain-language definition: AI-generated or transformed data used for training or evaluation.

Technical definition: A generated example, label, critique, prompt, response, or test case used in training, fine-tuning, or evaluation.

Example: A model-generated answer becomes a training example for its descendant.

Related pages: /anatomy/synthetic-training-data

System-level evaluation

Plain-language definition: Testing the whole AI system, not just one model.

Technical definition: Evaluation of the runtime composition including models, prompts, memory, tools, routing, permissions, evaluators, environment, and change paths.

Example: A review tests the model plus memory and tool permissions together.

Related pages: /the-problem

Teleodynamic loop

Plain-language definition: A resource-bounded loop where system change must pay for itself.

Technical definition: A fast/slow adaptation pattern in which structural changes occur only when expected utility exceeds memory, latency, energy, risk, and maintenance cost.

Example: A system adds a specialist only if accuracy gain outweighs added latency.

Related pages: /evolution/teleodynamic-reproduction-control

Teleodynamic viability

Plain-language definition: The resource-bounded condition under which a structural change is justified only if expected benefit repays memory, latency, energy, safety, license, and maintenance cost.

Technical definition: The resource-bounded condition under which a structural change is justified only if expected benefit repays memory, latency, energy, safety, license, and maintenance cost.

Tool profile

Plain-language definition: The set of external actions an AI system is allowed to take.

Technical definition: Permissions, scopes, credentials, rate limits, allowed endpoints, and approval requirements for tools.

Example: A calendar assistant can read events but cannot send email.

Related pages: /anatomy/tools-and-external-permissions

Transition graph

Plain-language definition: The map of how an AI system is allowed to change over time.

Technical definition: The graph of permitted changes: fine-tune, merge, distill, quantize, prune, route, replace, promote, retire, restore, consolidate memory, alter evaluator, and change permissions.

Example: A system that can swap adapters and memory states has a larger transition graph than one static model.

Related pages: /technical-research

Verifiable local manifest

Plain-language definition: A signed or auditable inventory of model, runtime, adapter, memory, router, evaluator, tool, consent, and rollback state.

Technical definition: A machine-readable local ecology manifest used to prove what ran, what was loaded, what memory existed, which tools were authorized, and what reset evidence was produced.

Example: Review the ecology, not only the model file, before claiming reset or retirement is complete.

Zombie behavior

Plain-language definition: Old behavior that was not actually gone.

Technical definition: A behavior that remains active or reappears after the visible model, adapter, prompt, or route was retired.

This glossary defines the site vocabulary used across guides, evidence cards, diagrams, and Risk Lab worksheets. Several terms are proposed Cognivirus terminology rather than standardized language. Those entries are marked to prevent a metaphor from being confused with an established scientific taxonomy.

The terms are meant to make distributed behavior easier to discuss without implying that AI systems are biological organisms, conscious agents, or literal viruses. The reference point is engineering analysis: what carries a behavior, what activates it, what preserves it, what invalidates assurance, and what must be restored during rollback.