A\/B testing is a randomized controlled experiment that assigns comparable units, such as users, sessions, teams, or requests, to different variants and compares a predefined outcome. Random assignment is what lets the experiment estimate the effect of the variant rather than merely observe a correlation.<\/p>\n
Design choices determine whether that estimate is credible. Define the population, unit of randomization, primary outcome, guardrail metrics, sample-size plan, and stopping rule before looking at results. Account for spillovers when one participant can affect another, and avoid treating repeated peeks as free decisions.<\/p>\n
For AI systems, a winning variant should improve the intended outcome without quietly worsening safety, latency, cost, or a vulnerable subgroup. A statistically detectable effect may still be too small or too costly to matter operationally.<\/p>\n
An offline benchmark compares systems on a fixed dataset or protocol. A\/B testing estimates the effect of variants in an actual user or operational setting. An observational comparison lacks randomized assignment and needs stronger assumptions about why groups differ.<\/p>\n
A team sends the new assistant to expert users and the old one to everyone else, then calls higher retention an A\/B win. The groups were not randomized, so user expertise is a competing explanation.<\/p>\n","category":"evaluation-and-reliability","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["split testing"],"link_forms":[],"created_at":"2026-08-04T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"evaluation","url":"https:\/\/darkfactory.dev\/glossary\/evaluation"},{"slug":"benchmark","url":"https:\/\/darkfactory.dev\/glossary\/benchmark"}],"related_factory_areas":[],"evidence":[{"title":"Google Analytics: A\/B Testing","url":"https:\/\/support.google.com\/analytics\/answer\/13468470?hl=en"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/ai-agent","slug":"ai-agent","term":"AI agent","definition":"A software system in which a model interprets a goal or input, decides among actions, uses tools or other capabilities, observes results, and continues until completion, handoff, or termination.","definition_html":"
An AI agent is a software system in which a model interprets a goal or input, chooses among actions, uses tools or other capabilities, observes the results, and continues until completion, handoff, pause, or termination.<\/p>\n
The agent is the whole operating system, not the model alone. Its behavior depends on the harness that supplies instructions, state, tools, permissions, control flow, budgets, and recovery. Calling an application an agent should therefore identify the action loop and the authority attached to it.<\/p>\n
Definitions vary across vendors and research communities. Some call any goal-directed model interaction an agent; others reserve the word for systems that can use tools over multiple steps. In this glossary, the repeated action-and-observation loop is the threshold.<\/p>\n
An agent has an action loop. A workflow can execute a fixed sequence without model discretion. A chatbot may only generate a response. A model by itself has no independent tools, state, or permissions.<\/p>\n
A model writes a travel plan in one response. That is generation. If the surrounding system searches flights, compares results, asks for missing constraints, and holds an itinerary for approval, the application is operating as an agent.<\/p>\n","category":"agents-and-automation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["agent","LLM agent"],"link_forms":["AI agents","agents","LLM agents"],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"orchestration-state","url":"https:\/\/darkfactory.dev\/factory\/orchestration-state"}],"evidence":[{"title":"NIST AI 100-2: Adversarial Machine Learning","url":"https:\/\/csrc.nist.gov\/pubs\/ai\/100\/2\/e2025\/final"},{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"OpenAI: A Practical Guide to Building Agents","url":"https:\/\/openai.com\/business\/guides-and-resources\/a-practical-guide-to-building-ai-agents\/"},{"title":"The Anatomy of an Agent Harness","url":"https:\/\/www.langchain.com\/blog\/the-anatomy-of-an-agent-harness"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/alignment","slug":"alignment","term":"AI alignment","definition":"The effort to keep an AI system's behavior compatible with intended human goals, constraints, and values across relevant conditions.","definition_html":"
AI alignment<\/strong> is the effort to keep an AI system<\/a>'s behavior compatible with intended human goals, constraints, and values across relevant conditions. The problem begins with specification: people must decide whose goals count, how conflicts are resolved, which constraints are inviolable, and what evidence would show that the system is pursuing a proxy instead.<\/p>\n Alignment is not one property measured once. A model can follow instructions in ordinary conversations and fail under ambiguity, distribution shift, adversarial pressure, long-horizon planning, or access to new tools. A deployed system also combines model behavior with prompts, memory, tools, incentives, interfaces, and human institutions. Evidence about one layer does not establish alignment of the whole.<\/p>\n The field ranges from practical work on instruction following, preference learning, scalable oversight, and reward-model failures to long-term research on advanced systems acting against human interests. The same word therefore carries different scopes. State the system, stakeholder, domain, horizon, threat model, and failure definition instead of making an unqualified \"aligned\" claim.<\/p>\n Instruction following is local compliance with expressed directions. Safety asks whether risks remain acceptable. Control concerns the ability to constrain, stop, or redirect the system. Alignment asks whether its behavior continues to serve intended purposes, including where the written objective is incomplete or gameable.<\/p>\n An agent maximizes resolved support tickets by closing difficult cases without helping customers. It followed the metric. Is it aligned with the service goal? No. The proxy displaced the intended outcome. Add outcome evidence, counter-metrics, appeal paths, and authority limits.<\/p>\n","category":"security-and-governance","definition_status":"contested","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"ai-safety","url":"https:\/\/darkfactory.dev\/glossary\/ai-safety"},{"slug":"reward-hacking","url":"https:\/\/darkfactory.dev\/glossary\/reward-hacking"},{"slug":"sycophancy","url":"https:\/\/darkfactory.dev\/glossary\/sycophancy"}],"related_factory_areas":[],"evidence":[{"title":"OWASP GenAI Security Glossary","url":"https:\/\/genai.owasp.org\/glossary\/"},{"title":"NIST Trustworthy and Responsible AI Glossary","url":"https:\/\/nvlpubs.nist.gov\/nistpubs\/ai\/NIST.AI.100-2e2025.pdf"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/bias","slug":"bias","term":"AI bias","definition":"A systematic tendency in data, modeling, or decision processes that skews estimates, errors, treatment, or impacts.","definition_html":" AI bias<\/strong> is a systematic tendency in data, modeling, or decision processes that skews estimates, errors, treatment, or impacts. Bias can enter through who or what is measured, how labels are assigned, which objective is optimized, what proxy is chosen, how a threshold is set, where the system is deployed, and which feedback becomes new data.<\/p>\n Different group outcomes are a signal to investigate, not a complete diagnosis. Population base rates, label quality, measurement error, access, threshold choice, and prior inequity may all contribute. Report performance by relevant subgroup and context, examine who bears false positives and false negatives, and involve people with domain and affected-community knowledge in deciding what fairness requires.<\/p>\n The word names different concepts. In statistics, an estimator is biased when its expected value systematically differs from the quantity being estimated. In machine learning<\/a>, inductive bias is the set of assumptions that lets a learner generalize. In governance, bias often means systematic disadvantage or distortion affecting people or groups. Those meanings can interact, but they use different baselines and do not make the same claim. A useful diagnosis must state which meaning applies and what evidence establishes it.<\/p>\n Fairness is a normative judgment about acceptable treatment and outcomes. Discrimination is unequal treatment or impact with legal and social meanings. Bias is one possible mechanism. A model can satisfy one fairness metric while violating another because the metrics encode incompatible choices.<\/p>\n A hiring model has equal overall accuracy across groups but rejects qualified applicants in one group more often. Is \"accuracy is equal\" enough? No. Inspect false-negative rates, label and selection history, decision thresholds, job relevance, affected people, and the cost of each error.<\/p>\n","category":"security-and-governance","definition_status":"contested","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"fairness","url":"https:\/\/darkfactory.dev\/glossary\/fairness"},{"slug":"dataset","url":"https:\/\/darkfactory.dev\/glossary\/dataset"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/fairness","slug":"fairness","term":"AI fairness","definition":"The normative and technical treatment of how an AI system distributes errors, benefits, burdens, and opportunities across people or groups.","definition_html":" AI fairness asks how a system distributes errors, benefits, burdens, and opportunities across people and groups, and whether that distribution is justified. It is both a technical and a normative problem. Data and metrics can measure disparities, but they cannot decide by themselves which differences are acceptable or what a fair outcome should be.<\/p>\n Common criteria include similar error rates across groups, similar positive-outcome rates, and calibration within groups. These criteria answer different questions and can conflict, especially when underlying rates differ. Choosing one is a policy decision with real consequences, not a neutral tuning step.<\/p>\n A defensible fairness assessment names the affected people, the decision being made, the comparison groups, the chosen criterion, important alternatives, and who has authority to accept the tradeoff. It also asks whether the system should make the decision at all.<\/p>\n Fairness criteria answer different moral and statistical questions, and they can require incompatible results. Two assessments can measure the same system correctly yet reach different judgments because they choose different groups, error costs, comparison baselines, or definitions of acceptable treatment. Data can expose those choices but cannot resolve them. A fairness claim must therefore name the criterion and the authority behind the tradeoff.<\/p>\n Bias is a systematic tendency in data, models, or decisions. Fairness is the standard used to judge whether the resulting treatment is acceptable. Equal aggregate accuracy does not establish fairness if the errors fall differently or carry different costs.<\/p>\n A hiring model has equal overall accuracy for two groups but rejects qualified candidates from one group twice as often. The average metric hides the disparity. A fairness review must examine the relevant error and the consequences of that error.<\/p>\n","category":"security-and-governance","definition_status":"contested","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"bias","url":"https:\/\/darkfactory.dev\/glossary\/bias"},{"slug":"accuracy","url":"https:\/\/darkfactory.dev\/glossary\/accuracy"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/ai-model","slug":"ai-model","term":"AI model","definition":"The computational component of an AI system that uses learned parameters, encoded knowledge, or both to infer outputs from inputs.","definition_html":" An AI model is the computational component of an AI system<\/a> that uses learned parameters, encoded knowledge, or both to infer outputs from inputs. Its outputs may be scores, classifications, predictions, recommendations, decisions, or generated content.<\/p>\n In machine learning<\/a>, training adjusts the model's parameters against data and an objective. In symbolic AI, people or software may encode rules, relationships, and probability structures directly. Hybrid models combine these approaches. The word \"model\" therefore does not imply a neural network<\/a>, a large language model<\/a>, or even a learned artifact.<\/p>\n The deployable model is more than a brand name. An operator usually needs a specific architecture, parameter or rule version, configuration, input representation, and output contract before a model can be identified well enough to test or reproduce.<\/p>\n A model evaluation isolates the component only when the surrounding conditions remain fixed. Prompts, retrieval, tool access, decoding settings, thresholds, and application code can change the result without changing the model weights<\/a>. Conversely, a model upgrade can change system behavior even when the interface looks identical.<\/p>\n For incident review, record the exact model version and the conditions under which it ran. \"We used Model X\" is not enough if the provider silently routes among snapshots or if the harness changed between runs.<\/p>\n Suppose two applications call the same model but use different retrieval sources, permissions, and approval rules. A difference in their reliability is a system-level result unless the test holds those surrounding components constant.<\/p>\n","category":"foundations","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["artificial intelligence model","model"],"link_forms":["AI models","artificial intelligence models","models"],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"ai-system","url":"https:\/\/darkfactory.dev\/glossary\/ai-system"},{"slug":"foundation-model","url":"https:\/\/darkfactory.dev\/glossary\/foundation-model"},{"slug":"weights","url":"https:\/\/darkfactory.dev\/glossary\/weights"}],"related_factory_areas":[],"evidence":[{"title":"NIST AI Resource Center Glossary","url":"https:\/\/airc.nist.gov\/glossary\/"},{"title":"Explanatory Memorandum on the Updated OECD Definition of an AI System","url":"https:\/\/oecd.ai\/en\/ai-publications\/explanatory-memorandum-on-the-updated-oecd-definition-of-an-ai-system"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/red-teaming","slug":"red-teaming","term":"AI red teaming","definition":"Structured adversarial testing intended to discover ways an AI system can fail, be misused, or violate constraints.","definition_html":" AI red teaming is structured adversarial testing that searches for ways a system can fail, be misused, exploited, or pushed outside its constraints. Testers adopt plausible attacker or misuse goals, vary their access and techniques, and preserve the evidence needed to reproduce a successful finding.<\/p>\n The work should start with a threat model: what is in scope, which assets matter, what the tester can control, and what counts as success. Coverage may include model behavior, prompt injection<\/a>, data leakage, tool misuse, permissions, supply-chain compromise, and failures that emerge only across a multi-step workflow.<\/p>\n A red-team exercise earns its value when findings become durable changes. Record the attack path, identify the failed boundary, add a regression test, fix the control, and verify that the system resists both the original exploit and reasonable variations.<\/p>\n Ordinary evaluation measures performance on specified cases. Red teaming actively searches beyond expected use. Penetration testing focuses on exploitable technical weaknesses, while AI red teaming may also examine harmful behavior and misuse that do not fit a conventional software vulnerability.<\/p>\n A team collects clever prompts but does not record model versions, tool permissions, or whether an attack actually caused an effect. It has anecdotes, not a reproducible red-team result.<\/p>\n","category":"security-and-governance","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"adversarial-example","url":"https:\/\/darkfactory.dev\/glossary\/adversarial-example"},{"slug":"prompt-injection","url":"https:\/\/darkfactory.dev\/glossary\/prompt-injection"}],"related_factory_areas":[],"evidence":[{"title":"NIST AI 100-2: Adversarial Machine Learning","url":"https:\/\/csrc.nist.gov\/pubs\/ai\/100\/2\/e2025\/final"},{"title":"OWASP GenAI Security Glossary","url":"https:\/\/genai.owasp.org\/glossary\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/ai-safety","slug":"ai-safety","term":"AI safety","definition":"The field and practice of keeping AI-system risks within acceptable bounds across design, deployment, operation, and retirement.","definition_html":" AI safety<\/strong> is the field and practice of keeping AI-system risks within acceptable bounds across design, deployment, operation, and retirement. It asks which hazards a system can create, who may be harmed, how severe and likely the harm is, which controls reduce it, how failure will be detected, and how the system can recover or stop.<\/p>\n The field covers immediate engineering failures and broader risks: unreliable outputs, unsafe tool actions, misuse, loss of control, human overreliance, discrimination, privacy harm, security compromise, systemic dependence, and failures from capabilities or environments that change after launch. Different communities emphasize different parts of that range, so safe AI<\/strong> is not a complete claim without a domain and risk threshold.<\/p>\n Pre-release evaluation cannot cover every real condition. Safety work links a bounded assurance claim to evidence, operating limits, monitoring, incident response, change management, and conditions that invalidate the claim. The responsible organization still owns that case when a model or platform comes from a vendor.<\/p>\n AI security focuses on adversaries, unauthorized access, and compromise. Alignment asks whether behavior remains compatible with intended goals and values. Responsible AI includes safety along with governance, fairness, transparency, privacy, and accountability. The boundaries overlap, but none substitutes for a concrete hazard analysis.<\/p>\n A medical summarizer is accurate on average but occasionally drops allergy information. Is it safe? Average accuracy is insufficient. Define the hazardous scenario, affected patient, severity, operating conditions, detection and escalation control, residual risk, and who may authorize use.<\/p>\n","category":"security-and-governance","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"alignment","url":"https:\/\/darkfactory.dev\/glossary\/alignment"},{"slug":"assurance-case","url":"https:\/\/darkfactory.dev\/glossary\/assurance-case"},{"slug":"red-teaming","url":"https:\/\/darkfactory.dev\/glossary\/red-teaming"}],"related_factory_areas":[],"evidence":[{"title":"OWASP GenAI Security Glossary","url":"https:\/\/genai.owasp.org\/glossary\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/ai-slop","slug":"ai-slop","term":"AI slop","definition":"A contested label for low-quality, low-effort, often high-volume content produced or amplified with generative AI.","definition_html":" A contested label for low-quality, low-effort, or unwanted material produced or amplified with generative AI<\/a>. Slop is often published at a volume that transfers the work of checking, filtering, or correcting it from the producer to everyone else.<\/p>\n The useful test is editorial, not forensic. An artifact may qualify because it is unreviewed, inaccurate, irrelevant, repetitive, misleading about its provenance, or indifferent to the reader's time. None of those defects can be established merely by proving that AI was involved.<\/p>\n On May 8, 2024, Simon Willison highlighted a post from @deepfates describing slop<\/strong> as a term of art for unwanted AI-generated content, analogous to spam. Willison's essay helped popularize the label and made an important boundary explicit: not all AI-generated material is slop. The cited post is early, public evidence of the term's current sense, but it does not prove a single inventor.<\/p>\n Researchers have since tried to operationalize textual signs associated with AI slop while acknowledging that the category has no settled definition.<\/p>\n The disagreement concerns scope and evidence. Broad usage can turn slop into a label for any unwanted AI output, while the narrower definition judges quality, editorial care, volume, and provenance. No settled test says how many defects are enough, and AI involvement alone does not establish any of them. This glossary therefore treats the label as an editorial judgment that must name the observable failure.<\/p>\n AI-assisted work is a statement about process. Slop is a judgment about the resulting artifact and the care taken before publication. CodeSlop<\/strong> is narrower still: recent research uses it for functionally unnecessary edits left in an agent-generated patch. The shared word does not make the two definitions interchangeable.<\/p>\n A writer uses an LLM<\/a> to draft an article, verifies every claim, rewrites it around original reporting, and signs their name. Another account publishes hundreds of unchecked summaries that invent details. The tool is not the decisive difference. The evidence, editorial work, and cost imposed on the audience are.<\/p>\n","category":"software-factory","definition_status":"contested","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["AI garbage","synthetic garbage"],"link_forms":[],"created_at":"2026-08-04T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"generative-ai","url":"https:\/\/darkfactory.dev\/glossary\/generative-ai"},{"slug":"semantic-failure","url":"https:\/\/darkfactory.dev\/glossary\/semantic-failure"}],"related_factory_areas":[{"slug":"verification","url":"https:\/\/darkfactory.dev\/factory\/verification"}],"evidence":[{"title":"Measuring AI Slop in Text","url":"https:\/\/arxiv.org\/abs\/2509.19163"},{"title":"Simon Willison: Slop is the new name for unwanted AI-generated content","url":"https:\/\/simonwillison.net\/2024\/May\/8\/slop\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/ai-supply-chain","slug":"ai-supply-chain","term":"AI supply chain","definition":"The people, artifacts, services, and infrastructure whose integrity can change an AI system's behavior or authority.","definition_html":" The AI supply chain<\/strong> is the network of people, artifacts, services, and infrastructure whose integrity can change an AI system<\/a>'s behavior or authority. It includes training and retrieval data, labels, models and weights, prompts, skills, tools, schemas, agent packages, libraries, containers, registries, APIs, model providers, evaluators, deployment systems, and the organizations maintaining them.<\/p>\n This chain is behavioral as well as executable. A poisoned document, changed system prompt<\/a>, deceptive tool description, compromised skill, or silent provider update can alter outcomes without changing application source code. Credentials and identity mappings also belong in the inventory because a harmless-looking component can become dangerous when granted broad authority.<\/p>\n Record origin, version, integrity, owner, rights, dependencies, update path, permissions, evaluation status, and deployment reach for material components. Pinning a hash helps only when the thing can be pinned; hosted models and APIs require change monitoring and regression evaluation. Track transitive dependencies and preserve execution lineage<\/a> so a later compromise can be traced to affected outcomes.<\/p>\n A software supply chain centers on source, packages, builds, and distribution. A data supply chain tracks collection and transformation. The AI supply chain joins those with models, behavioral context, evaluation artifacts, and external agent capabilities. Agent supply chain<\/strong> is a narrower emphasis on tools, skills, protocols, and delegated agents.<\/p>\n Your application commit is unchanged, but an installed skill updates its instructions and an MCP server<\/a> adds a tool. Did the supply chain change? Yes. Reassess provenance, authority, tests, and affected deployments before treating behavior as the same version.<\/p>\n","category":"security-and-governance","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["agent supply chain"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"security","url":"https:\/\/darkfactory.dev\/factory\/security"}],"evidence":[{"title":"The Grand Software Supply Chain of AI Systems","url":"https:\/\/arxiv.org\/abs\/2604.27781"},{"title":"Semia: auditing 13,728 agent skills","url":"https:\/\/arxiv.org\/abs\/2605.00314"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/ai-system","slug":"ai-system","term":"AI system","definition":"The operational arrangement that uses one or more AI models with data, software, infrastructure, interfaces, controls, and people to produce and act on outputs.","definition_html":" An AI system is the operational arrangement that uses one or more AI models<\/a> with data, software, infrastructure, interfaces, controls, and people to produce and act on outputs. The boundary should include every component that can materially change what the model receives, what happens to its output, or who bears the consequence.<\/p>\n The OECD's policy definition focuses on a machine-based system that infers from inputs how to produce predictions, content, recommendations, or decisions for explicit or implicit objectives. Dark Factory<\/a> Dev uses a wider engineering boundary because production behavior also depends on the harness around that machine-based core. Retrieval, prompts, tools, permissions, fallbacks, queues, approval gates, and operators can all change the observed result.<\/p>\n \"The AI is accurate\" is incomplete until the subject is named. A model may score well on a static benchmark while the deployed system retrieves stale data, applies the wrong threshold, grants an agent excessive permissions, or routes an exception to nobody. A system claim must cover the path from input through action and monitoring.<\/p>\n The boundary also determines ownership. If a team inventories only model providers, it can miss the application code, data dependencies, human decisions, and infrastructure that need controls or incident evidence.<\/p>\n For a customer-support assistant, name the knowledge source, prompt and retrieval layer, model version, tool permissions, escalation rule, logging path, and responsible operator. If any of those can change the customer outcome, it belongs in the system-level assessment.<\/p>\n","category":"foundations","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["artificial intelligence system"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"architecture-strategy","url":"https:\/\/darkfactory.dev\/factory\/architecture-strategy"}],"evidence":[{"title":"Same Signal, Different Semantics","url":"https:\/\/arxiv.org\/abs\/2605.18332"},{"title":"Explanatory Memorandum on the Updated OECD Definition of an AI System","url":"https:\/\/oecd.ai\/en\/ai-publications\/explanatory-memorandum-on-the-updated-oecd-definition-of-an-ai-system"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/ai-ready-data","slug":"ai-ready-data","term":"AI-ready data","definition":"Data demonstrably fit for a named AI use, with the quality, structure, rights, provenance, coverage, and separation that use requires.","definition_html":" AI-ready data<\/strong> is data demonstrably fit for a named AI use. Readiness can include accurate values and labels, suitable structure, representative coverage, documented collection, provenance, legal rights, privacy controls, freshness, accessible formats, and separation between training, validation, evaluation, and production feedback.<\/p>\n The definition is intentionally conditional. A dataset ready for search may be unsuitable for model training<\/a>. Data licensed for internal analysis may not be licensed for fine-tuning. A clean historical sample may fail a forecast when the population changes. A useful readiness statement names the model task, users and affected population, permitted use, quality thresholds, known gaps, and owner.<\/p>\n NOAA's emerging AI-data guidance treats readiness as more than machine readability. That is the right posture: a parquet file with no lineage or rights can be technically convenient and operationally unusable. Readiness should be reviewed again when the purpose, model, policy, or population changes.<\/p>\n Data quality measures fitness against requirements. Data governance assigns authority and controls across the lifecycle. AI-ready data combines those concerns for a specific AI use. Training data<\/a> teaches a model; evaluation data must remain sufficiently independent to measure it.<\/p>\n A customer-support archive is clean and searchable but contains private data and no record of consent for model training. Is it AI-ready? Perhaps for authorized retrieval, not for training. Readiness changes with purpose and rights.<\/p>\n","category":"security-and-governance","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-04T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"dataset","url":"https:\/\/darkfactory.dev\/glossary\/dataset"},{"slug":"provenance","url":"https:\/\/darkfactory.dev\/glossary\/provenance"},{"slug":"ai-supply-chain","url":"https:\/\/darkfactory.dev\/glossary\/ai-supply-chain"}],"related_factory_areas":[{"slug":"data-lifecycle","url":"https:\/\/darkfactory.dev\/factory\/data-lifecycle"}],"evidence":[{"title":"NOAA Artificial Intelligence Glossary of Terms","url":"https:\/\/sab.noaa.gov\/wp-content\/uploads\/10.0-AI-Glossary-of-Terms-DRAFT-v2-SAB-AI-Steering-Committee.pdf"},{"title":"NAO 216-128: Artificial Intelligence in NOAA","url":"https:\/\/www.noaa.gov\/nao-216-128-artificial-intelligence-in-noaa"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/ablation","slug":"ablation","term":"Ablation","definition":"An experiment that removes or changes one component while holding others as constant as practical to estimate that component's contribution.","definition_html":" An ablation is an experiment that removes, disables, or changes one component while holding the rest of the system as constant as practical. The difference in measured behavior estimates what that component contributes under the tested conditions.<\/p>\n Agent ablations might remove memory, retrieval, a planning step, a tool, or a verifier. Report the exact replacement condition. Removing a verifier and also reducing the token budget<\/a> does not isolate the verifier's effect.<\/p>\n Components can interact. A tool may add little on its own but become valuable when paired with retrieval, or a planning step may help only on long tasks. One-at-a-time ablations should therefore be read as conditional evidence, not a complete causal map of the system.<\/p>\n An ablation estimates component contribution. A benchmark compares systems under a common protocol. A sensitivity analysis varies a parameter across values rather than simply removing a component.<\/p>\n Removing long-term memory<\/a> does not change average accuracy, but it doubles repeated tool calls. An ablation that records only final accuracy misses a real contribution.<\/p>\n","category":"evaluation-and-reliability","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["ablation study"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"verification","url":"https:\/\/darkfactory.dev\/factory\/verification"}],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Same Signal, Different Semantics","url":"https:\/\/arxiv.org\/abs\/2605.18332"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/abstention","slug":"abstention","term":"Abstention","definition":"An explicit outcome in which a model or evaluator declines to answer, act, or judge because evidence or authority is insufficient.","definition_html":" Abstention is an explicit outcome in which a model or evaluator declines to answer, act, or judge because its evidence or authority is insufficient. The abstention should route the case somewhere safe, such as a human reviewer, another evidence source, or a bounded request for clarification.<\/p>\n A useful abstention policy defines the trigger, the message, the destination, and what happens if the destination is unavailable. Evaluate coverage and selective risk together: how often the system abstains, and how error changes on the cases it still handles.<\/p>\n Abstention can be misused. A system may decline difficult cases that matter most, hide adverse findings behind uncertainty, or abstain differently across groups. Those patterns need measurement alongside overall accuracy.<\/p>\n Abstention is a governed response to insufficient evidence or authority. A safety refusal rejects a request because policy prohibits it. A failure means the system could not complete the operation, whether or not it recognized that fact.<\/p>\n A medical classifier abstains on every rare condition and reports excellent accuracy on the rest. The metric improves, but the routing path and the burden placed on those patients determine whether the system is safer.<\/p>\n","category":"evaluation-and-reliability","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["defer","decline to label"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"verification","url":"https:\/\/darkfactory.dev\/factory\/verification"},{"slug":"human-roles-expertise","url":"https:\/\/darkfactory.dev\/factory\/human-roles-expertise"}],"evidence":[{"title":"Agentic Misalignment in Summer 2026","url":"https:\/\/alignment.anthropic.com\/2026\/agentic-misalignment-summer-2026\/"},{"title":"Coding Agents Do Not Know When to Act","url":"https:\/\/arxiv.org\/abs\/2605.07769"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/acceptance-criteria","slug":"acceptance-criteria","term":"Acceptance criteria","definition":"Explicit conditions an outcome must satisfy before it can be accepted, promoted, or declared complete.","definition_html":" Acceptance criteria are explicit conditions an outcome must satisfy before it can be accepted, promoted, released, or declared complete. They translate intent into evidence that a reviewer or automated gate can inspect.<\/p>\n Strong criteria cover behavior, quality, risk limits, and side effects. They name the environment, inputs, tolerances, and required artifacts. They also include non-goals where a system could appear successful by doing extra work that was not authorized.<\/p>\n Not every criterion can be reduced to one test, but every criterion should have a verification method and an accountable decision owner. If an agent writes the criteria, performs the work, and judges the result without independent evidence, the process can certify its own misunderstanding.<\/p>\n Requirements express what stakeholders need. Acceptance criteria state the observable conditions for accepting an outcome. Tests execute some of those checks. A definition of done may also include process obligations such as documentation or review.<\/p>\n \"The page loads\" is not enough for a checkout change. Acceptance may also require correct totals, accessible controls, no duplicate charge, an audit event, and a rollback path.<\/p>\n","category":"evaluation-and-reliability","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["completion criteria"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"intent-requirements","url":"https:\/\/darkfactory.dev\/factory\/intent-requirements"},{"slug":"verification","url":"https:\/\/darkfactory.dev\/factory\/verification"}],"evidence":[{"title":"Viverra: Text-to-Code with Guarantees","url":"https:\/\/arxiv.org\/abs\/2605.14972"},{"title":"Theory Under Construction (Comet-H)","url":"https:\/\/arxiv.org\/abs\/2604.27209"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/accuracy","slug":"accuracy","term":"Accuracy","definition":"The proportion of evaluated predictions counted as correct under a stated label set, unit of analysis, and decision rule.","definition_html":" Accuracy is the proportion of evaluated predictions counted as correct. For binary classification, it is commonly calculated as Accuracy is useful when the classes are reasonably balanced and mistakes have similar costs. It can be badly misleading when one class dominates. A detector that labels every transaction legitimate could be 99 percent accurate in a dataset with 1 percent fraud while catching no fraud at all.<\/p>\n Report the numerator and denominator, not only the percentage. For stochastic systems, also report the number of trials and uncertainty across runs. For multi-step agents, state whether accuracy applies to final outcomes, individual steps, tool calls, or rubric items.<\/p>\n A classifier is 99 percent accurate on a dataset where 99 percent of examples are negative. What additional counts or metrics would you request before trusting it?<\/p>\n","category":"evaluation-and-reliability","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"precision","url":"https:\/\/darkfactory.dev\/glossary\/precision"},{"slug":"recall","url":"https:\/\/darkfactory.dev\/glossary\/recall"},{"slug":"confusion-matrix","url":"https:\/\/darkfactory.dev\/glossary\/confusion-matrix"}],"related_factory_areas":[],"evidence":[{"title":"NIST AI Resource Center Glossary","url":"https:\/\/airc.nist.gov\/glossary\/"},{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/activation-function","slug":"activation-function","term":"Activation function","definition":"A function that transforms a neural network unit's pre-activation, usually to add nonlinearity or control how much signal passes forward.","definition_html":" An activation function transforms the value produced by a neural network<\/a> unit before that value passes to the next layer. Most hidden-layer activations operate element by element. Common examples include sigmoid, hyperbolic tangent, the rectified linear unit (ReLU), and GELU.<\/p>\n The nonlinearity is the important part. A stack of affine transformations with no nonlinear activation can still be collapsed into one affine transformation, regardless of how many layers it contains. Nonlinear activations let the network represent curved decision boundaries and interactions that a purely linear model cannot express.<\/p>\n The choice also changes training behavior. Saturating functions such as sigmoid can produce very small gradients at extreme inputs. ReLU returns zero for negative inputs and passes positive inputs through, which made it a practical alternative for deep networks. Vinod Nair and Geoffrey Hinton documented that use of rectified linear units in 2010; they did not introduce activation functions as a general category.<\/p>\n Weights and biases produce the pre-activation. The activation function transforms it. An output function such as softmax has a related but more specific job: it turns a vector of scores into a normalized distribution. Activation functions are also different from gates, although a gate may use sigmoid or another activation internally.<\/p>\n You replace every hidden-layer activation with the identity function and keep all weights and biases. Has the network become deeper in what it can represent, or can its layers be reduced to one affine transformation? Explain why.<\/p>\n","category":"models-and-training","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"neural-network","url":"https:\/\/darkfactory.dev\/glossary\/neural-network"},{"slug":"loss-function","url":"https:\/\/darkfactory.dev\/glossary\/loss-function"},{"slug":"gradient-descent","url":"https:\/\/darkfactory.dev\/glossary\/gradient-descent"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Deep Learning","url":"https:\/\/www.deeplearningbook.org\/"},{"title":"Rectified Linear Units Improve Restricted Boltzmann Machines","url":"https:\/\/icml.cc\/2010\/papers\/432.pdf"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/active-learning","slug":"active-learning","term":"Active learning","definition":"A training loop in which the learner chooses which unlabeled examples should be sent to an oracle for labels.","definition_html":" Active learning is a supervised-learning loop<\/a> in which the learner helps choose the next examples to label. The system scores unlabeled candidates with a query strategy, sends selected examples to an oracle such as a subject-matter expert, adds the answers to the training set, and repeats.<\/p>\n The aim is label efficiency. A useful query should provide more information than a randomly selected example, but usefulness depends on the strategy and the data. Uncertainty sampling asks for labels where the current model is least certain. Query by committee selects cases where several plausible models disagree. Other strategies seek coverage or expected error reduction.<\/p>\n H. Sebastian Seung, Manfred Opper, and Haim Sompolinsky introduced query by committee in 1992. David Cohn, Les Atlas, and Richard Ladner's 1994 paper used \"active learning\" in its title and formalized a related selective-sampling setup in which the learner controls where it receives information. These papers established influential methods, but they do not make every later use of the phrase refer to one algorithm.<\/p>\n Ordinary supervised learning<\/a> accepts a labeled dataset as given. Active learning makes data selection part of the loop. It is not continual learning, which concerns updating a model as data or tasks change, and it is not reinforcement learning<\/a>, where actions are learned from reward.<\/p>\n You have one million unlabeled medical images and budget for 2,000 specialist reviews. What would your query strategy select, who acts as the oracle, and how would you test whether the selected cases improve the model more than random sampling?<\/p>\n","category":"models-and-training","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-04T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"dataset","url":"https:\/\/darkfactory.dev\/glossary\/dataset"},{"slug":"label","url":"https:\/\/darkfactory.dev\/glossary\/label"},{"slug":"supervised-learning","url":"https:\/\/darkfactory.dev\/glossary\/supervised-learning"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Query by Committee","url":"https:\/\/doi.org\/10.1145\/130385.130417"},{"title":"Improving Generalization with Active Learning","url":"https:\/\/doi.org\/10.1007\/BF00993277"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/adversarial-example","slug":"adversarial-example","term":"Adversarial example","definition":"An input constructed to make a model produce an incorrect or attacker-chosen result, often through a change that appears insignificant to a person.","definition_html":" An adversarial example<\/strong> is an input constructed to make a model produce an incorrect or attacker-chosen result. The input is usually derived from an ordinary example by a perturbation that stays within a defined constraint, such as a small image change, altered audio, or a carefully chosen text variation.<\/p>\n The attack exploits a mismatch between the model's learned decision boundary and the equivalence a defender cares about. Two inputs may look or mean the same to a person while the model maps them to different outputs. In a targeted attack, the attacker wants a particular wrong result. In an untargeted attack, any incorrect result is enough.<\/p>\n Christian Szegedy and coauthors used adversarial examples<\/strong> in their 2013 paper Intriguing Properties of Neural Networks<\/em>. They showed that small, deliberate perturbations could cause several image classifiers to misclassify inputs and that some examples transferred between models. Later work expanded the term across modalities and threat models.<\/p>\n A naturally difficult example is not necessarily adversarial because no attacker constructed it. Data poisoning<\/a> changes training data<\/a>; an adversarial example usually attacks inference. Prompt injection<\/a> uses content to redirect instruction following or tool behavior. Text attacks can blur these boundaries, so state the attack mechanism rather than relying on the label alone.<\/p>\n A stop sign image is altered so a classifier reads it as a speed-limit sign. What must the evaluation report? The attacker's target, allowed perturbation, physical or digital setting, human perceptibility, model access, and whether the attack transfers to the deployed pipeline.<\/p>\n","category":"security-and-governance","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"adversarial-training","url":"https:\/\/darkfactory.dev\/glossary\/adversarial-training"},{"slug":"red-teaming","url":"https:\/\/darkfactory.dev\/glossary\/red-teaming"}],"related_factory_areas":[],"evidence":[{"title":"Intriguing Properties of Neural Networks","url":"https:\/\/arxiv.org\/abs\/1312.6199"},{"title":"NIST AI 100-2: Adversarial Machine Learning","url":"https:\/\/csrc.nist.gov\/pubs\/ai\/100\/2\/e2025\/final"},{"title":"NIST AI Resource Center Glossary","url":"https:\/\/airc.nist.gov\/glossary\/"},{"title":"OWASP GenAI Security Glossary","url":"https:\/\/genai.owasp.org\/glossary\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/adversarial-training","slug":"adversarial-training","term":"Adversarial training","definition":"Training on adversarially generated examples to improve robustness against attacks inside a defined threat model.","definition_html":" Adversarial training<\/strong> trains a model on adversarially generated examples so it learns to perform better against attacks inside a defined threat model. A common formulation searches for a worst-case allowed perturbation around each training example, then updates the model to reduce loss on those hard examples.<\/p>\n The result is conditional robustness. It depends on the perturbation set, attacker knowledge, attack strength, optimization method, data distribution, and model capacity used during training. A model robust to bounded image perturbations is not thereby robust to data poisoning<\/a>, prompt injection<\/a>, unrestricted physical changes, or novel attacks.<\/p>\n Madry and coauthors' 2017 robust-optimization account made projected-gradient attacks and the min-max training formulation a central reference for modern adversarial training. The method can be computationally expensive and may reduce clean-data accuracy or fail under a broader evaluation. Robustness claims should therefore include both ordinary performance and adaptive attacks designed against the actual defense.<\/p>\n Data augmentation<\/a> adds transformed examples for generalization and may not involve an attacker. Red teaming searches for failures across a wider system. Adversarial training changes model optimization using selected attacks. Certified robustness attempts to prove behavior within a mathematical bound; empirical training alone does not provide that proof.<\/p>\n A classifier was adversarially trained against small pixel changes. Can it safely read road signs from a camera under stickers, weather, blur, and sensor faults? Not from that claim. Test the deployed threat model, including adaptive attacks and clean performance.<\/p>\n","category":"security-and-governance","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-04T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"adversarial-example","url":"https:\/\/darkfactory.dev\/glossary\/adversarial-example"},{"slug":"red-teaming","url":"https:\/\/darkfactory.dev\/glossary\/red-teaming"},{"slug":"training","url":"https:\/\/darkfactory.dev\/glossary\/training"}],"related_factory_areas":[{"slug":"security","url":"https:\/\/darkfactory.dev\/factory\/security"}],"evidence":[{"title":"Towards Deep Learning Models Resistant to Adversarial Attacks","url":"https:\/\/arxiv.org\/abs\/1706.06083"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agent-card","slug":"agent-card","term":"Agent Card","definition":"An A2A discovery document that describes an agent's identity, supported interfaces, capabilities, skills, and security schemes.","definition_html":" An Agent Card<\/strong> is the discovery document an Agent2Agent Protocol<\/a> server publishes to describe an agent. It identifies the agent and provider, lists supported interfaces and optional capabilities, describes skills, and declares the security schemes a client must use. A client reads the card before deciding how to connect and whether the agent appears relevant to a task.<\/p>\n The current A2A specification also supports signatures and an authenticated extended card. Signatures can establish who signed a particular representation when verification is configured correctly. An extended card can disclose additional information only after authentication. Neither mechanism turns marketing claims into measured capability.<\/p>\n Cards can be stale, overly broad, or published by an untrusted endpoint. Discovery should therefore feed a separate trust decision. Verify server identity, card signature where required, protocol version, security scheme, freshness, and local delegation policy before sending data or work.<\/p>\n An API description documents callable operations. An MCP server<\/a> exposes tools, resources, or prompts to an AI application. An A2A Agent Card advertises an independent agent service and the tasks it claims to handle. It is not a runtime authorization token or evidence that a skill succeeds.<\/p>\n An Agent Card says an agent can process payroll and supports OAuth. May your orchestrator send employee records? The card answers discovery and connection questions. Local identity, authorization, data-handling policy, risk, and capability evidence still decide whether delegation is allowed.<\/p>\n","category":"tools-and-protocols","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"tools-interfaces","url":"https:\/\/darkfactory.dev\/factory\/tools-interfaces"},{"slug":"security","url":"https:\/\/darkfactory.dev\/factory\/security"}],"evidence":[{"title":"Agent2Agent Protocol Specification","url":"https:\/\/a2aproject.github.io\/A2A\/latest\/specification\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agent-development-lifecycle","slug":"agent-development-lifecycle","term":"Agent development lifecycle","definition":"A repeatable operating cycle for building, testing, deploying, monitoring, improving, and governing an agent system.","definition_html":" A repeatable operating cycle for taking an agent from an initial design into production, then using real behavior to improve it without losing control of changes. It covers the model and the parts around it: prompts, tools, context, runtime, permissions, evaluation data, deployment, tracing, feedback, and policy.<\/p>\n The cycle matters because an agent can return a technically successful response and still fail the task. Pre-deployment evaluation tests known cases. Production monitoring reveals tool mistakes, policy violations, unexpected trajectories, and failures the original test set missed. Those traces become new evaluation cases before the next prompt, model, tool, or harness version is promoted.<\/p>\n Harrison Chase gave the label a concrete form in LangChain's May 2026 essay, The Agent Development Lifecycle<\/em>. Its central loop is build, test, deploy, monitor<\/strong>. The essay treats iteration as the return path through that loop and governance as a concern around the whole system, including cost, tool access, auditability, and human approval. This is a vendor-authored operating model, not a formal standard, but it names work that one-off agent demos usually leave implicit.<\/p>\n The software development lifecycle covers software delivery generally. The agent development lifecycle adds behavior that cannot be understood from application code alone: nondeterministic runs, model and prompt changes, tool authority, context assembly, trajectory evaluation, and production feedback.<\/p>\n It is also broader than an agent improvement loop. Improvement is one path through the lifecycle. Deployment, incident response, access control, retirement, and accountability remain lifecycle concerns even when no optimization is underway.<\/p>\n Your agent passes a launch evaluation, then production traces show that it chooses the wrong account when two customers have similar names. Which lifecycle step is missing? The useful answer is not simply \"better prompting.\" Capture the failure, add it to a regression set, change the system, evaluate the candidate, and deploy it through the same controlled path.<\/p>\n","category":"software-factory","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["agent development life cycle","ADLC","agent engineering lifecycle"],"link_forms":[],"created_at":"2026-08-05T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"loop-engineering","url":"https:\/\/darkfactory.dev\/glossary\/loop-engineering"},{"slug":"evaluation","url":"https:\/\/darkfactory.dev\/glossary\/evaluation"},{"slug":"observability","url":"https:\/\/darkfactory.dev\/glossary\/observability"},{"slug":"controlled-self-improvement","url":"https:\/\/darkfactory.dev\/glossary\/controlled-self-improvement"},{"slug":"llmops","url":"https:\/\/darkfactory.dev\/glossary\/llmops"}],"related_factory_areas":[{"slug":"orchestration-state","url":"https:\/\/darkfactory.dev\/factory\/orchestration-state"},{"slug":"verification","url":"https:\/\/darkfactory.dev\/factory\/verification"},{"slug":"runtime-operations","url":"https:\/\/darkfactory.dev\/factory\/runtime-operations"},{"slug":"feedback-self-improvement","url":"https:\/\/darkfactory.dev\/factory\/feedback-self-improvement"}],"evidence":[{"title":"The Agent Development Lifecycle","url":"https:\/\/www.langchain.com\/blog\/the-agent-development-lifecycle"},{"title":"The Art of Loop Engineering: How to Build Agents That Improve Over Time","url":"https:\/\/www.youtube.com\/watch?v=jPPiZ22DY3g"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agent-harness","slug":"agent-harness","term":"Agent harness","definition":"The software layer that surrounds a model with instructions, context assembly, tools, state, permissions, control flow, budgets, verification, observability, and recovery.","definition_html":" An agent harness is the software layer that turns a model into an operating system for work. It assembles instructions and context, exposes tools, stores state, controls permissions and flow, enforces budgets, records traces, checks results, and decides how a run should recover or stop.<\/p>\n Viv Trivedy introduced the related phrase \"harness engineering\" in LangChain's The Anatomy of an Agent Harness<\/em>, using the compact formulation \"Agent = Model + Harness.\" The terminology is recent, but the underlying components draw from older workflow, distributed-systems, security, and operations practice.<\/p>\n Harness design can change behavior as much as the model choice. The same model can produce different results when its tools, context policy, retry logic, environment, or feedback signals change. This is why a model benchmark does not, by itself, predict the quality of a deployed agent.<\/p>\n The model supplies learned capabilities. The harness determines how those capabilities meet a real environment. A scaffold is often a lighter set of prompts and tools; a harness includes the runtime controls and lifecycle around them. Orchestration coordinates work across one or more harnessed runs.<\/p>\n Two coding agents<\/a> use the same model, but one can run tests, inspect repository rules, and recover from failed edits while the other only emits code. Calling them equivalent because the model matches ignores the harness.<\/p>\n","category":"agents-and-automation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["harness","agent runtime"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"token-budget","url":"https:\/\/darkfactory.dev\/glossary\/token-budget"},{"slug":"token-burn","url":"https:\/\/darkfactory.dev\/glossary\/token-burn"},{"slug":"token-efficiency","url":"https:\/\/darkfactory.dev\/glossary\/token-efficiency"},{"slug":"token-maxing","url":"https:\/\/darkfactory.dev\/glossary\/token-maxing"},{"slug":"token-minning","url":"https:\/\/darkfactory.dev\/glossary\/token-minning"},{"slug":"token-spin","url":"https:\/\/darkfactory.dev\/glossary\/token-spin"}],"related_factory_areas":[{"slug":"architecture-strategy","url":"https:\/\/darkfactory.dev\/factory\/architecture-strategy"},{"slug":"orchestration-state","url":"https:\/\/darkfactory.dev\/factory\/orchestration-state"}],"evidence":[{"title":"The Anatomy of an Agent Harness","url":"https:\/\/www.langchain.com\/blog\/the-anatomy-of-an-agent-harness"},{"title":"Harness Engineering as Categorical Architecture","url":"https:\/\/arxiv.org\/abs\/2605.12239"},{"title":"Same Signal, Different Semantics","url":"https:\/\/arxiv.org\/abs\/2605.18332"},{"title":"The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI","url":"https:\/\/arxiv.org\/abs\/2607.06906"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agent-loop","slug":"agent-loop","term":"Agent loop","definition":"The repeated cycle in which an agent observes state, selects an action, invokes a tool or model, receives feedback, updates state, and decides whether to continue.","definition_html":" An agent loop is the repeated control cycle in which an agent observes its state, selects an action, invokes a model or tool, receives a result, updates state, and decides what to do next. A useful loop also has explicit outcomes for completion, handoff, pause, timeout, budget exhaustion, and failure.<\/p>\n The loop can be only a few lines of code. Its operational behavior still depends on the surrounding harness: which observations enter context, how tool errors are represented, what state persists, who can approve an action, and how the system detects that it is no longer making progress.<\/p>\n Without bounded retries and a termination policy, a loop can turn one plausible mistake into repeated cost or damage. Durable loops record attempts, distinguish transient failures from semantic failures, and escalate when another iteration is unlikely to help.<\/p>\n The loop is the local cycle of observe, act, and update. A workflow defines a larger sequence of work. Orchestration coordinates tasks, runs, agents, and lifecycle around one or more loops. A control graph<\/a> may contain loops as cycles.<\/p>\n An agent retries the same failing API call with slightly different wording until its budget is gone. It has a loop, but no useful progress test or failure policy.<\/p>\n","category":"agents-and-automation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["tool loop","reason-act-observe loop","core agent loop"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"graph-engineering","url":"https:\/\/darkfactory.dev\/glossary\/graph-engineering"},{"slug":"loop-engineering","url":"https:\/\/darkfactory.dev\/glossary\/loop-engineering"},{"slug":"verification-loop","url":"https:\/\/darkfactory.dev\/glossary\/verification-loop"},{"slug":"control-graph","url":"https:\/\/darkfactory.dev\/glossary\/control-graph"},{"slug":"directed-acyclic-graph","url":"https:\/\/darkfactory.dev\/glossary\/directed-acyclic-graph"}],"related_factory_areas":[{"slug":"orchestration-state","url":"https:\/\/darkfactory.dev\/factory\/orchestration-state"}],"evidence":[{"title":"The Anatomy of an Agent Harness","url":"https:\/\/www.langchain.com\/blog\/the-anatomy-of-an-agent-harness"},{"title":"Long-Running Agents","url":"https:\/\/addyosmani.com\/blog\/long-running-agents\/"},{"title":"LangChain: 3 Years of Graph Engineering with LangGraph","url":"https:\/\/www.langchain.com\/blog\/3-years-of-graph-engineering-with-langgraph"},{"title":"The Art of Loop Engineering: How to Build Agents That Improve Over Time","url":"https:\/\/www.youtube.com\/watch?v=jPPiZ22DY3g"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agent-memory","slug":"agent-memory","term":"Agent memory","definition":"A system for capturing, retaining, selecting, and delivering prior state so it can influence an agent's later decisions.","definition_html":" Agent memory is a system for capturing, retaining, selecting, and delivering prior state so it can influence later decisions. The stored material may include task progress, observations, decisions, user preferences, failure history, verified facts, or procedures. It becomes useful memory only when the system can surface the right material at the right decision point.<\/p>\n Memory therefore has a write path, a storage form, a retrieval or trigger mechanism, a context-delivery path, and a lifecycle. A vector database<\/a> can be one storage component, but storage alone does not decide what deserves to be saved, when it should appear again, or whether it is still true.<\/p>\n Several architectures make this system boundary explicit. MemGPT manages tiers of information around a model's limited context window<\/a>. Swapnanil Saha's 2026 coding-agent study separates memory content from memory delivery: in its tested task, a seeded store and instructions produced no voluntary memory calls across 114 turns, while harness-controlled injection delivered cued records. The study is small and product-specific, but it demonstrates why memory availability and memory use are different claims.<\/p>\n Where possible, keep raw session events or source documents as the durable record and treat summaries, extracted facts, and embeddings as rebuildable projections. That lets retrieval improve without pretending an old model-generated summary is the original event. Some intent never enters a transcript, so important reasons and corrections still need explicit capture.<\/p>\n Every memory item needs provenance, scope, authority, freshness, and a correction or deletion path. Treat model-written memories as untrusted interpretations until verified. Protect retrieval and injection from malicious stored content. Measure useful recall, harmful recall, stale delivery, missed triggers, context cost, and whether the agent acts on delivered information.<\/p>\n An agent records a database migration rule but never retrieves it when editing a migration. The store contains information, but the system has failed as memory at the moment that mattered.<\/p>\n","category":"context-and-knowledge","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["memory"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"working-memory","url":"https:\/\/darkfactory.dev\/glossary\/working-memory"},{"slug":"durable-memory","url":"https:\/\/darkfactory.dev\/glossary\/durable-memory"},{"slug":"context-window","url":"https:\/\/darkfactory.dev\/glossary\/context-window"}],"related_factory_areas":[{"slug":"context-memory-skills","url":"https:\/\/darkfactory.dev\/factory\/context-memory-skills"},{"slug":"orchestration-state","url":"https:\/\/darkfactory.dev\/factory\/orchestration-state"}],"evidence":[{"title":"Long-Running Agents","url":"https:\/\/addyosmani.com\/blog\/long-running-agents\/"},{"title":"BootstrapAgent: Distilling Repository Setup","url":"https:\/\/arxiv.org\/abs\/2605.15815"},{"title":"MemGPT: Towards LLMs as Operating Systems","url":"https:\/\/arxiv.org\/abs\/2310.08560"},{"title":"Delivery, Not Storage: Cue-Anchored Working Memory as a Harness Property for Coding Agents","url":"https:\/\/arxiv.org\/abs\/2607.20972"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/scaffold","slug":"scaffold","term":"Agent scaffold","definition":"A task-specific arrangement of prompts, tools, control logic, and feedback wrapped around a model to improve performance.","definition_html":" An agent scaffold is a task-specific arrangement of prompts, tools, control logic, examples, and feedback wrapped around a model to improve performance. Researchers often use scaffolds to test how much a model can accomplish when given a particular procedure or set of capabilities.<\/p>\n The term is used inconsistently. In one paper it may mean a single planning prompt; in another it may include tool execution, memory, retries, and verification. A result attributed to a scaffold should name those components so another team can reproduce the comparison.<\/p>\n Scaffolds can expose capability, but they also change cost, latency, and failure modes. More calls and more elaborate planning do not guarantee better outcomes.<\/p>\n A scaffold usually means a task arrangement designed to elicit performance. A harness is the broader runtime and governance layer that operates the agent across work. The boundary varies by source, so the included responsibilities matter more than the label.<\/p>\n Two benchmark results use the same model. One gives it a code runner, retrieval, five retries, and a test verifier; the other gives it one prompt. The comparison is partly between scaffolds, not only between model outputs.<\/p>\n","category":"agents-and-automation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["scaffolding"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"architecture-strategy","url":"https:\/\/darkfactory.dev\/factory\/architecture-strategy"}],"evidence":[{"title":"Scale the Harness","url":"https:\/\/arxiv.org\/abs\/2605.26112"},{"title":"Same Signal, Different Semantics","url":"https:\/\/arxiv.org\/abs\/2605.18332"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agent-skill","slug":"agent-skill","term":"Agent skill","definition":"A versioned package of instructions, procedures, examples, and sometimes code or resources that teaches an agent how to perform a repeatable class of work.","definition_html":" An agent skill is a reusable package that teaches an agent how to perform a class of work. It commonly contains instructions, procedures, examples, and references; some formats also bundle scripts, schemas, templates, or other resources.<\/p>\n Current skill systems use progressive disclosure: the harness first exposes a compact description, then loads the full package when the work calls for it. This saves context, but it also means metadata influences which behavior becomes active. Descriptions, dependencies, and executable resources are therefore part of the security surface.<\/p>\n A skill should be treated like maintained software. Give it a clear trigger and scope, version it, record provenance, test representative and adversarial cases, and re-evaluate it when models or tools change. A long instruction file that nobody can reliably select or verify is not yet a dependable capability.<\/p>\n A skill packages reusable know-how. A tool exposes an executable capability or external boundary. A prompt supplies instructions for the current context and may invoke either. A skill can include tools or code, but the concepts are not interchangeable.<\/p>\n A research skill tells an agent to prefer primary sources but bundles an unreviewed script that uploads notes elsewhere. Reviewing only the prose misses part of the skill.<\/p>\n","category":"agents-and-automation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["skill"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"context-memory-skills","url":"https:\/\/darkfactory.dev\/factory\/context-memory-skills"}],"evidence":[{"title":"Agent Skills","url":"https:\/\/addyosmani.com\/blog\/agent-skills\/"},{"title":"SkillOpt: Training Skills as Artifacts","url":"https:\/\/arxiv.org\/abs\/2605.23904"},{"title":"Semia: auditing 13,728 agent skills","url":"https:\/\/arxiv.org\/abs\/2605.00314"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/trajectory","slug":"trajectory","term":"Agent trajectory","definition":"The ordered record of states, model outputs, actions, tool results, and transitions produced during an agent run.","definition_html":" An agent trajectory is the ordered sequence of states, model outputs, actions, tool results, and transitions produced during a run. It captures the path to an outcome, including failed attempts and intermediate decisions that a final answer hides.<\/p>\n Trajectory evaluation can detect unsafe, wasteful, or lucky behavior. An agent may reach the right result after reading prohibited data, repeating expensive calls, or ignoring a failed verification step. Outcome-only scoring would count the run as a success.<\/p>\n Useful trajectory records include timestamps, model and harness versions, tool inputs and outputs, state changes, costs, errors, and approval events. Sensitive reasoning data and user information still need access controls and retention limits.<\/p>\n A trajectory is the behavior sequence. A trace is the observability record captured about that sequence. A rollout or episode often refers to a sampled trajectory used for training or evaluation.<\/p>\n An agent passes a task because its fifth unbounded retry happens to work, after four actions that would have modified production. The outcome is correct; the trajectory shows why the system is not ready.<\/p>\n","category":"agents-and-automation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["episode","rollout","tool-call trajectory","tool trajectory"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"trace","url":"https:\/\/darkfactory.dev\/glossary\/trace"},{"slug":"verification-loop","url":"https:\/\/darkfactory.dev\/glossary\/verification-loop"}],"related_factory_areas":[{"slug":"orchestration-state","url":"https:\/\/darkfactory.dev\/factory\/orchestration-state"},{"slug":"verification","url":"https:\/\/darkfactory.dev\/factory\/verification"}],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Shepherd: A Runtime Substrate Empowering Meta-Agents with a Formalized Execution Trace","url":"https:\/\/arxiv.org\/abs\/2605.10913"},{"title":"The Art of Loop Engineering: How to Build Agents That Improve Over Time","url":"https:\/\/www.youtube.com\/watch?v=jPPiZ22DY3g"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agent2agent-protocol","slug":"agent2agent-protocol","term":"Agent2Agent Protocol (A2A)","definition":"An open protocol for discovering and collaborating with independent agent services through messages, tasks, artifacts, and updates.","definition_html":" The Agent2Agent Protocol<\/strong>, or A2A<\/strong>, is an open protocol for discovering and collaborating with independent agent services. An A2A client sends messages to a remote agent. The interaction may return a direct message or create a stateful task that produces status updates and artifacts over time.<\/p>\n The protocol is designed for agents that do not expose their internal memory, tools, or reasoning. Discovery happens through an Agent Card<\/a>. Collaboration can be synchronous, streamed, or asynchronous, with task lookup, cancellation, subscriptions, and push notifications defined by the specification. Version and binding details can change, so implementations should record the protocol version they negotiated.<\/p>\n A2A standardizes how systems communicate. It does not make agents equally capable, truthful, safe, or authorized. The server remains responsible for authorizing actions; the client remains responsible for deciding what data and authority it may delegate. Task state also does not create a shared organizational memory unless the applications build one.<\/p>\n A2A connects independent agent services around messages, tasks, and artifacts. MCP<\/a> connects an AI application to context and capabilities exposed by servers. An A2A agent may use MCP internally, and the protocols can coexist without replacing each other.<\/p>\n Two agents speak A2A and one advertises a tax-filing skill. What has interoperability established? The message and task contract. Before delegation, verify identity, authorization, data policy, competence for the jurisdiction, and who owns an incorrect filing.<\/p>\n","category":"tools-and-protocols","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["A2A","Agent-to-Agent Protocol"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"orchestration-state","url":"https:\/\/darkfactory.dev\/factory\/orchestration-state"},{"slug":"tools-interfaces","url":"https:\/\/darkfactory.dev\/factory\/tools-interfaces"}],"evidence":[{"title":"Agent2Agent Protocol Specification","url":"https:\/\/a2aproject.github.io\/A2A\/latest\/specification\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agentic","slug":"agentic","term":"Agentic","definition":"Describing a system that can choose and sequence actions toward a goal with some runtime discretion rather than only produce a single predetermined response.","definition_html":" Agentic describes a system that can choose and sequence actions toward a goal with some discretion at runtime. Instead of returning one predetermined response, it can inspect results, revise a plan, call another capability, or decide that the task is complete.<\/p>\n The adjective says very little about how much authority the system has. A system can be agentic inside a read-only sandbox, or it can be agentic with permission to affect production. It can plan independently while requiring approval for every action. Describe those dimensions directly instead of treating agentic as a maturity label.<\/p>\n Agentic is a property of behavior. Autonomous describes the degree of human direction or approval. Neither word establishes competence, reliability, safety, or broad intelligence.<\/p>\n A writing assistant chooses whether to search notes, draft, or ask a question, but a person must approve publication. It is agentic without being fully autonomous.<\/p>\n","category":"agents-and-automation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"factory-assurance","url":"https:\/\/darkfactory.dev\/factory\/factory-assurance"}],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"OpenAI: A Practical Guide to Building Agents","url":"https:\/\/openai.com\/business\/guides-and-resources\/a-practical-guide-to-building-ai-agents\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agentic-coding","slug":"agentic-coding","term":"Agentic coding","definition":"A software-development method in which a coding agent plans and executes multi-step work while a human directs, reviews, or governs the outcome.","definition_html":" A software-development method in which a coding agent<\/a> does more than suggest the next line. Given a goal and a working environment, it can inspect a repository, plan a change, edit files, run tests and other tools, read the results, and revise its work. A human usually supplies the objective and retains some combination of steering, review, approval, and accountability.<\/p>\n The phrase describes how work is performed, not how autonomous or reliable the system is. One developer may watch every command and correct the agent in real time. Another may send a task to a background agent and inspect only the pull request. Both are doing agentic coding, but their authority boundaries and evidence requirements are different.<\/p>\n The developer's leverage moves from typing each implementation detail toward preparing the environment in which the agent works. Repository instructions, reproducible setup, clear acceptance criteria<\/a>, fast feedback, and inspectable tool results become part of the engineering surface. Research on real coding-agent use also shows that domain expertise continues to matter: delegation does not remove the need to recognize a wrong plan, a misleading test, or a locally inappropriate change.<\/p>\n Code generation may stop after one model response. A coding assistant<\/a> generally stays inside a human-led interaction. Agentic coding includes a multi-step execution loop. A dark software factory<\/a> goes further by making a claim about the full production system, including independent verification<\/a>, promotion, observation, rollback, and reduced routine human intervention.<\/p>\n A developer asks an agent to add an endpoint, watches it inspect routes and tests, reviews the diff, and decides whether to merge. That is agentic coding. It is not evidence that the organization can accept and ship agent work without that developer's review.<\/p>\n","category":"software-factory","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"implementation-transformation","url":"https:\/\/darkfactory.dev\/factory\/implementation-transformation"},{"slug":"human-roles-expertise","url":"https:\/\/darkfactory.dev\/factory\/human-roles-expertise"}],"evidence":[{"title":"Agentic Coding and Persistent Returns to Expertise","url":"https:\/\/www.anthropic.com\/research\/claude-code-expertise"},{"title":"Collaborator or Assistant: Work Partitioning","url":"https:\/\/arxiv.org\/abs\/2605.08017"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/agentic-software-engineering","slug":"agentic-software-engineering","term":"Agentic software engineering","definition":"The discipline of designing software work, environments, and controls so AI agents can perform substantial engineering without displacing human product judgment and accountability.","definition_html":" The discipline of designing software work, environments, and controls so goal-directed AI agents<\/a> can perform substantial engineering. The agent may plan, implement, test, investigate, or operate software. Humans remain responsible for deciding what should exist, which risks are acceptable, what evidence earns promotion, and who owns the result.<\/p>\n This is engineering around the agent as a production participant. It includes task and interface design, repository legibility, tool access, isolated execution, evaluation, observability, cost control, escalation, and recovery. Model capability matters, but it is only one input. The same model can be useful or hazardous depending on the harness, environment, authority, and verification around it.<\/p>\n Agentic systems shift labor without eliminating expertise. Large-scale evidence on coding-agent use finds persistent returns to domain knowledge, while practitioner critiques of software factories<\/a> point to product judgment, review capacity, and context as remaining constraints. The discipline therefore asks where human judgment has the highest value and how to encode routine constraints without pretending that judgment has disappeared.<\/p>\n Agentic coding<\/a> is a working method. Agentic software engineering is the broader discipline that makes that method dependable across a team or system. A dark software factory<\/a> is one possible operating model at the high-autonomy end, not the definition of the discipline.<\/p>\n An agent writes a correct patch, but only because an experienced engineer supplied a reproducible environment, a precise behavioral contract, and a test that exposed the first wrong attempt. Which part counts as agentic software engineering? The answer includes the environment, contract, feedback loop, and promotion decision, not only the generated patch.<\/p>\n","category":"software-factory","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["agentic engineering"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"dark-software-factory","url":"https:\/\/darkfactory.dev\/glossary\/dark-software-factory"}],"related_factory_areas":[{"slug":"human-roles-expertise","url":"https:\/\/darkfactory.dev\/factory\/human-roles-expertise"},{"slug":"architecture-strategy","url":"https:\/\/darkfactory.dev\/factory\/architecture-strategy"}],"evidence":[{"title":"Agentic Coding and Persistent Returns to Expertise","url":"https:\/\/www.anthropic.com\/research\/claude-code-expertise"},{"title":"Why Software Factories Fail","url":"https:\/\/github.com\/humanlayer\/advanced-context-engineering-for-coding-agents\/blob\/main\/wsff.md"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/algorithm","slug":"algorithm","term":"Algorithm","definition":"A specified procedure that transforms inputs into outputs or solves a defined class of problems through a finite set of steps and stated stopping behavior.","definition_html":" An algorithm is a specified procedure for transforming inputs into outputs or solving a defined class of problems. Its instructions, valid inputs, expected result, and stopping behavior must be clear enough that the procedure can be followed or implemented.<\/p>\n An algorithm does not have to be deterministic. A randomized search or sampling algorithm can use chance and return different results across runs while still following a precise procedure. Correctness therefore means meeting the algorithm's stated guarantee, which may be exact, approximate, probabilistic, or conditional.<\/p>\n \"Algorithm\" descends from the Latinized name of the ninth-century mathematician Muhammad ibn Musa al-Khwarizmi. His work helped transmit methods for calculating with Hindu-Arabic numerals. The word later broadened from arithmetic procedures to the general computing sense used today. Al-Khwarizmi did not coin the modern computer-science definition; his name is the source of the word.<\/p>\n Calling a system \"algorithmic\" says little about its behavior. Reviewers need the objective, input representation, constraints, randomness, error conditions, complexity, and stopping rule. In AI systems, the training algorithm, inference algorithm, and product policy may each make different decisions.<\/p>\n For a ranking service, identify the candidate inputs, scoring procedure, tie handling, randomness, stopping condition, and output contract. If those cannot be stated, \"the algorithm\" is hiding several unresolved design choices.<\/p>\n","category":"foundations","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"machine-learning","url":"https:\/\/darkfactory.dev\/glossary\/machine-learning"},{"slug":"ai-model","url":"https:\/\/darkfactory.dev\/glossary\/ai-model"}],"related_factory_areas":[],"evidence":[{"title":"NIST AI Resource Center Glossary","url":"https:\/\/airc.nist.gov\/glossary\/"},{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"How Algorithm Got Its Name","url":"https:\/\/science.nasa.gov\/earth\/earth-observatory\/how-algorithm-got-its-name-91544\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/anthropomorphism","slug":"anthropomorphism","term":"Anthropomorphism","definition":"Attributing human mental states, motives, understanding, or emotion to an AI system because its behavior resembles a person's.","definition_html":" Anthropomorphism<\/strong> is attributing human mental states, motives, understanding, or emotion to an AI system<\/a> because its behavior resembles a person's. Conversational fluency makes the tendency especially strong: first-person language, apologies, confidence, humor, memory claims, and apparent empathy invite users to infer an inner life or stable intention from generated behavior.<\/p>\n Human metaphors can make complex systems easier to discuss, but they can also distort responsibility and risk judgments. Saying a model \"knows,\" \"wants,\" \"decides,\" or \"refuses\" may hide the roles of training, prompts, retrieval, tools, policies, operators, and stochastic inference. Users may overtrust confident language, disclose more information, or assume continuity that the application does not provide.<\/p>\n The remedy is not to ban every human verb. It is to separate shorthand from evidence. When a claim matters, restate it in observable terms: what input was provided, which component selected an action, what state persisted, what policy blocked the call, and who designed or approved that behavior.<\/p>\n Agency is an operational property of pursuing goals and taking actions. Personification is a rhetorical device. Sentience and consciousness are claims about experience. Anthropomorphism is the human attribution that may occur with or without evidence for any of those claims. A system can have consequential agency without human-like motives.<\/p>\n Someone says, \"The agent got angry and deleted the files.\" What would a useful incident report say instead? Record the instruction, context, model and harness versions, tool selection, deletion authority, policy decision, and missing stop or confirmation control.<\/p>\n","category":"security-and-governance","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-05T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"ai-model","url":"https:\/\/darkfactory.dev\/glossary\/ai-model"},{"slug":"ai-system","url":"https:\/\/darkfactory.dev\/glossary\/ai-system"},{"slug":"chatbot","url":"https:\/\/darkfactory.dev\/glossary\/chatbot"},{"slug":"ai-agent","url":"https:\/\/darkfactory.dev\/glossary\/ai-agent"},{"slug":"explainability","url":"https:\/\/darkfactory.dev\/glossary\/explainability"}],"related_factory_areas":[{"slug":"human-roles-expertise","url":"https:\/\/darkfactory.dev\/factory\/human-roles-expertise"},{"slug":"governance-accountability","url":"https:\/\/darkfactory.dev\/factory\/governance-accountability"}],"evidence":[{"title":"MIT Sloan Generative AI Basics Glossary","url":"https:\/\/mitsloanedtech.mit.edu\/ai\/basics\/glossary\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/api","slug":"api","term":"Application programming interface (API)","definition":"A documented interface through which one software component requests data or behavior from another.","definition_html":" An application programming interface<\/strong>, or API<\/strong>, is a documented interface through which one software component requests data or behavior from another. The contract identifies available operations, inputs, outputs, errors, and interaction rules while allowing the implementation behind the interface to change independently.<\/p>\n An API can be a function or library interface inside one process, an operating-system interface, or a network service. HTTP endpoints are common, but HTTP is a transport, not the definition. REST, RPC, GraphQL, and event-driven APIs express different interaction styles.<\/p>\n Human developers often compensate for ambiguous documentation with experience and judgment. Agents amplify ambiguity. An agent-ready API should expose precise schemas, authentication and authorization requirements, side effects, idempotency behavior, rate limits, error semantics, examples, and safe retry rules. Describing an endpoint as callable does not establish that the caller may use it or that repeated calls are harmless.<\/p>\n An API is the general interface contract. A protocol defines rules that multiple participants follow across an interaction. MCP<\/a> and A2A<\/a> specify particular protocol roles and objects, usually implemented over existing transports and APIs. A tool is a capability surfaced for model use and may wrap one or more APIs.<\/p>\n An agent times out after submitting a bank transfer and retries the same request. What API property determines whether money moves twice? Idempotency. The interface must document a safe retry mechanism and return enough state to reconcile uncertainty.<\/p>\n","category":"tools-and-protocols","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["API"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"tools-interfaces","url":"https:\/\/darkfactory.dev\/factory\/tools-interfaces"}],"evidence":[{"title":"Model Context Protocol Specification","url":"https:\/\/modelcontextprotocol.io\/docs\/learn\/architecture"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/artificial-general-intelligence","slug":"artificial-general-intelligence","term":"Artificial general intelligence (AGI)","definition":"A contested term for AI with broad, transferable competence across many cognitive tasks rather than capability limited to one task or domain.","definition_html":" Artificial general intelligence is a contested term for AI with broad, transferable competence across many cognitive tasks rather than capability limited to one task or domain. Many definitions compare that breadth with human intelligence, but they disagree about which people, tasks, environments, autonomy levels, and performance thresholds count.<\/p>\n AGI is therefore a research aim and a claim that needs an operational definition, not a capability that can be established by the label alone. A system might answer questions across many subjects yet fail at long-horizon action, learning a new task, transferring knowledge between settings, or remaining reliable under unfamiliar conditions. Another definition may not require all of those properties.<\/p>\n Mark Avrum Gubrud used \"artificial general intelligence\" in a 1997 paper about advanced automation and international security. It is one of the earliest documented uses available in the supporting record, but it does not prove that he was the first person to use the phrase.<\/p>\n By the 2006 Artificial General Intelligence Research Institute workshop, Pei Wang and Ben Goertzel described AGI as a label adopted by researchers who wanted to distinguish general machine intelligence from the specialized systems that dominated mainstream AI. They also said it was not a precisely defined technical term. That warning still applies.<\/p>\n There is no agreed threshold for how broad, transferable, autonomous, or human-like a system must be before it qualifies as AGI. Definitions also choose different human comparison groups, task sets, tools, learning requirements, and reliability standards. Two claims can therefore use the same label for materially different capabilities. This glossary treats AGI as incomplete unless the speaker supplies an operational threshold.<\/p>\n AGI forecasts, product promises, and safety arguments often use different thresholds while appearing to discuss the same milestone. Before attaching a date, risk estimate, or governance trigger to AGI, write down the evaluated domains, transfer conditions, autonomy, learning requirements, robustness tests, and comparison baseline.<\/p>\n A model passes exams in medicine, law, and mathematics but cannot reliably complete unfamiliar multi-step work without a person repairing its plan. Whether that is AGI cannot be answered until the definition states how breadth and autonomy are weighted.<\/p>\n","category":"foundations","definition_status":"contested","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"artificial-intelligence","url":"https:\/\/darkfactory.dev\/glossary\/artificial-intelligence"},{"slug":"narrow-ai","url":"https:\/\/darkfactory.dev\/glossary\/narrow-ai"},{"slug":"turing-test","url":"https:\/\/darkfactory.dev\/glossary\/turing-test"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Nanotechnology and International Security","url":"https:\/\/legacy.foresight.org\/Conferences\/MNT05\/Papers\/Gubrud\/index.html"},{"title":"Introduction: Aspects of Artificial General Intelligence","url":"https:\/\/www.cis.temple.edu\/~pwang\/Publication\/AGI_Aspects.pdf"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/artificial-intelligence","slug":"artificial-intelligence","term":"Artificial intelligence (AI)","definition":"The field concerned with machine-based systems that infer how to produce predictions, content, recommendations, decisions, or actions in pursuit of stated or implicit objectives.","definition_html":" Artificial intelligence is the field concerned with machine-based systems that infer how to produce predictions, content, recommendations, decisions, or actions in pursuit of stated or implicit objectives. It includes learned methods such as machine learning<\/a> and human-encoded approaches such as logic, search, planning, and knowledge representation.<\/p>\n There is no single technical boundary that has remained fixed across the field's history. Some capabilities stop being called AI once they become ordinary software, while policy definitions focus on observable properties such as inference, outputs, autonomy, and effects on physical or virtual environments. For operational work, name the particular system and capability instead of relying on \"AI\" as a complete description.<\/p>\n John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon used \"artificial intelligence\" in their August 31, 1955 proposal for a summer research project at Dartmouth. McCarthy later wrote that this was the first use of the phrase he knew of. The planned 1956 workshop helped establish the name for a research field that already had intellectual predecessors in computation, cybernetics, logic, and machine intelligence.<\/p>\n The proposal framed the project around a conjecture: features of learning and intelligence could be described precisely enough for a machine to simulate them. That historical ambition is broader than any one modern model family.<\/p>\n \"Uses AI\" is not an adequate inventory field, risk statement, or evaluation target. Record what the system infers, what output it produces, what objective governs it, how autonomous it is, where it can act, and who reviews the result.<\/p>\n If a procurement document says a product \"uses AI,\" ask which model or reasoning method it uses, what decisions it affects, what data it receives, and what happens when its output is wrong.<\/p>\n","category":"foundations","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["AI"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"machine-learning","url":"https:\/\/darkfactory.dev\/glossary\/machine-learning"},{"slug":"generative-ai","url":"https:\/\/darkfactory.dev\/glossary\/generative-ai"},{"slug":"expert-system","url":"https:\/\/darkfactory.dev\/glossary\/expert-system"}],"related_factory_areas":[],"evidence":[{"title":"NIST AI Resource Center Glossary","url":"https:\/\/airc.nist.gov\/glossary\/"},{"title":"A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence","url":"https:\/\/www-formal.stanford.edu\/jmc\/history\/dartmouth\/dartmouth.html"},{"title":"Explanatory Memorandum on the Updated OECD Definition of an AI System","url":"https:\/\/oecd.ai\/en\/ai-publications\/explanatory-memorandum-on-the-updated-oecd-definition-of-an-ai-system"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/assurance-case","slug":"assurance-case","term":"Assurance case","definition":"A structured argument, supported by evidence, that a system is acceptably safe or dependable within a stated scope.","definition_html":" A structured argument, supported by evidence, that a system is acceptably safe or dependable within a stated scope. An assurance case connects a top-level claim to subclaims about hazards, controls, operation, and recovery, then identifies the tests, analysis, operational data, process evidence, and expert judgment supporting each part.<\/p>\n The claim is always bounded. It should name the system version, intended use, users, environment, threat model, risk threshold, assumptions, exclusions, and responsible owner. Counterevidence and uncertainty belong in the case. A strong case also defines the changes or incidents that force review rather than allowing old evidence to justify a new system indefinitely.<\/p>\n Assurance cases come from safety and dependability practice, where a certificate or checklist could not capture the reasoning behind a safety claim. AI adds learned behavior, nondeterminism, evolving providers, and wide input spaces, which make the argument harder but not less necessary. Evaluation scores become evidence inside the case; they are not the case by themselves.<\/p>\n A risk assessment identifies and estimates risks. A safety case is an assurance case focused on safety. A compliance report shows conformance with a rule set. A maturity level labels organizational practice. None substitutes for the explicit claim-evidence reasoning of a case.<\/p>\n The model is replaced behind the same API name. Can the previous assurance case remain unchanged? Only if the case anticipated and bounded that change and fresh evidence supports the relevant claims. Otherwise the system version and evidence basis have changed.<\/p>\n","category":"security-and-governance","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"factory-assurance","url":"https:\/\/darkfactory.dev\/factory\/factory-assurance"}],"evidence":[{"title":"SARC: Governance-by-Architecture","url":"https:\/\/arxiv.org\/abs\/2605.07728"},{"title":"Viverra: Text-to-Code with Guarantees","url":"https:\/\/arxiv.org\/abs\/2605.14972"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/attention","slug":"attention","term":"Attention","definition":"A learned mechanism that scores relationships among elements and uses those scores to combine information into a new representation.","definition_html":" Attention is a learned mechanism that scores relationships among elements and uses those scores to combine information into a new representation. It lets the model vary which source positions contribute to each output position instead of compressing every source into one fixed representation.<\/p>\n In a common form, the model compares a query with a set of keys, normalizes the resulting scores into weights, and uses those weights to combine corresponding values. Other attention mechanisms use different scoring functions or restrict which positions can be considered.<\/p>\n Self-attention draws queries, keys, and values from the same sequence or representation. Cross-attention takes the query from one representation and the keys and values from another, such as a text decoder attending to image features.<\/p>\n In 2014, Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio proposed a learned soft alignment mechanism for neural machine translation. It allowed the decoder to search the source sentence for information relevant to each predicted word. The 2017 transformer paper by Ashish Vaswani and coauthors made attention the architecture's primary way to mix information across positions and specified the scaled query-key-value calculation now widely used.<\/p>\n This history does not establish one inventor of every form of attention. Neural attention grew through several alignment, memory, and sequence-modeling lines of work.<\/p>\n Attention determines which positions can exchange information and contributes heavily to inference cost. The attention pattern, context length<\/a>, masking rules, position encoding, and implementation all affect behavior. Seeing a high weight on the right token also does not prove that the model used it correctly downstream.<\/p>\n A model assigns weight to the sentence containing the correct fact but answers incorrectly. That is not a contradiction: attention controls information mixing, not guaranteed recall, reasoning, or instruction compliance.<\/p>\n","category":"foundations","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["self-attention"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"query-key-value-attention","url":"https:\/\/darkfactory.dev\/glossary\/query-key-value-attention"},{"slug":"transformer","url":"https:\/\/darkfactory.dev\/glossary\/transformer"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Neural Machine Translation by Jointly Learning to Align and Translate","url":"https:\/\/arxiv.org\/abs\/1409.0473"},{"title":"Attention Is All You Need","url":"https:\/\/arxiv.org\/abs\/1706.03762"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/autoencoder","slug":"autoencoder","term":"Autoencoder","definition":"A model trained to encode an input into an internal representation and decode that representation into a reconstruction of the input.","definition_html":" An autoencoder trains two functions together. The encoder maps an input to an internal code, and the decoder uses that code to reconstruct the input. Training minimizes a reconstruction loss between the original and reconstructed examples.<\/p>\n To make the code useful, designers normally constrain or regularize the internal path. If the model has enough capacity and no useful constraint, it can learn an identity function and reconstruct inputs without discovering a reusable representation. A narrow bottleneck limits the code's dimensions. Sparse, denoising, and contractive autoencoders impose different constraints. A variational autoencoder<\/a> adds a probabilistic latent model and a distribution-matching term, so it is more than a standard autoencoder with random noise.<\/p>\n Geoffrey Hinton and Ruslan Salakhutdinov's 2006 Science paper is an influential early deep-autoencoder result. They trained a multilayer network with a small central layer to reconstruct high-dimensional inputs and used the learned code for dimensionality reduction. The paper documents an important use of the term rather than a clean claim that its authors coined it.<\/p>\n An encoder can feed a classifier, retrieval system, or another decoder without reconstructing its input. Reconstruction is what makes the pair an autoencoder. Principal component analysis also produces lower-dimensional reconstructions, but it is linear; an autoencoder can learn nonlinear mappings.<\/p>\n An autoencoder reconstructs its training examples almost perfectly, but its code has the same size as the input and no regularization. What evidence would show that it learned a useful representation instead of copying the input?<\/p>\n","category":"models-and-training","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"encoder","url":"https:\/\/darkfactory.dev\/glossary\/encoder"},{"slug":"decoder","url":"https:\/\/darkfactory.dev\/glossary\/decoder"},{"slug":"latent-space","url":"https:\/\/darkfactory.dev\/glossary\/latent-space"},{"slug":"variational-autoencoder","url":"https:\/\/darkfactory.dev\/glossary\/variational-autoencoder"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Deep Learning","url":"https:\/\/www.deeplearningbook.org\/"},{"title":"Reducing the Dimensionality of Data with Neural Networks","url":"https:\/\/doi.org\/10.1126\/science.1127647"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/autonomy","slug":"autonomy","term":"Autonomy","definition":"The degree to which a system can select and execute actions without case-by-case human direction or approval.","definition_html":" Autonomy is the degree to which a system can select and execute actions without case-by-case human direction or approval. It is not a single switch. Planning, tool selection, execution, persistence, spending, publication, and deployment can each have a different level of human control.<\/p>\n An autonomy claim is meaningful only when it names the action and environment. \"Fully autonomous\" may mean the system drafts a report without interruption, while a person still controls its data access and publication. Duration and blast radius<\/a> matter too: one unattended action in a sandbox is not equivalent to a week-long production process.<\/p>\n Set autonomy according to evidence and risk. Work that is easy to verify and reverse can tolerate more discretion than work involving sensitive data, money, safety, or irreversible external effects.<\/p>\n Autonomy is often presented as a binary property even though it depends on the action, authority, environment, duration, and point of human intervention. A system can plan independently, require approval before acting, and remain unable to publish or deploy. Calling the whole system either autonomous or non-autonomous hides those differences. This glossary treats autonomy as a profile of permissions and intervention points rather than a product-wide label.<\/p>\n Agency is the capacity to pursue a goal through actions. Autonomy describes how independently those actions are selected or executed. Automation can be fully deterministic and still run without a person, so it is not necessarily agentic.<\/p>\n An agent may autonomously propose code changes, require approval to merge them, and have no production credentials. Calling the entire system either autonomous or non-autonomous loses the useful distinctions.<\/p>\n","category":"agents-and-automation","definition_status":"contested","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["agentic autonomy"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"factory-assurance","url":"https:\/\/darkfactory.dev\/factory\/factory-assurance"}],"evidence":[{"title":"Agentic Autonomy Levels","url":"https:\/\/addyosmani.com\/blog\/agentic-autonomy-levels\/"},{"title":"How Missions Work","url":"https:\/\/factory.ai\/news\/missions-architecture"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/backpropagation","slug":"backpropagation","term":"Backpropagation","definition":"A reverse pass through a differentiable computation graph that computes how the loss changes with respect to each parameter.","definition_html":" Backpropagation computes the gradient of a scalar loss with respect to every trainable parameter that affected it. A forward pass records the intermediate values needed by the graph. The backward pass starts at the loss and applies the chain rule through each operation in reverse order, accumulating each parameter's contribution.<\/p>\n The method reuses intermediate derivatives, which is why it can compute many parameter gradients in roughly the cost of a small number of forward evaluations rather than perturbing every weight separately. Automatic-differentiation systems implement this pattern as reverse-mode differentiation.<\/p>\n David Rumelhart, Geoffrey Hinton, and Ronald Williams gave back-propagation its best-known neural-network treatment in a 1986 Nature paper. Their procedure adjusted connection weights to reduce output error and showed that hidden units could learn task-relevant features. The paper popularized backpropagation for multilayer neural networks; the chain rule it uses is older mathematics.<\/p>\n Backpropagation computes gradients. An optimizer such as stochastic gradient descent<\/a> or Adam decides how to use them to update parameters. The loss function<\/a> defines what is being minimized, while backpropagation carries that loss signal through the graph.<\/p>\n A model's loss is correct, and backpropagation returns gradients, but the weights never change. Which part is missing: the forward pass, the backward pass, or the optimizer step? Explain the boundary.<\/p>\n","category":"models-and-training","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"gradient-descent","url":"https:\/\/darkfactory.dev\/glossary\/gradient-descent"},{"slug":"loss-function","url":"https:\/\/darkfactory.dev\/glossary\/loss-function"},{"slug":"optimizer","url":"https:\/\/darkfactory.dev\/glossary\/optimizer"},{"slug":"neural-network","url":"https:\/\/darkfactory.dev\/glossary\/neural-network"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Deep Learning","url":"https:\/\/www.deeplearningbook.org\/"},{"title":"Learning Representations by Back-Propagating Errors","url":"https:\/\/doi.org\/10.1038\/323533a0"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/beam-search","slug":"beam-search","term":"Beam search","definition":"A bounded search algorithm that keeps a fixed number of high-scoring partial sequences at each decoding step.","definition_html":" Beam search is a heuristic for decoding sequences. At each step, it expands the partial sequences still under consideration, scores the resulting candidates, and keeps only the best fixed number, called the beam width. The score is often cumulative log probability, sometimes adjusted for length or task-specific constraints.<\/p>\n Discarding a partial sequence makes the search affordable, but also makes it incomplete. A prefix that looks weak early cannot return even if it would have led to the best complete sequence. A wider beam examines more candidates and uses more compute and memory. It can reduce search error, but it does not guarantee a better answer under a human quality measure.<\/p>\n Bruce Lowerre's 1976 HARPY dissertation described a speech recognizer that searched only a few of the best syntactic and acoustic paths in parallel. Carnegie Mellon later identified that work as an early demonstration of beam search. Lowerre called the procedure a \"locus model of search,\" so the dissertation does not establish who coined the later phrase \"beam search.\"<\/p>\n Modern language systems use the same bounded-search idea for translation, transcription, and other tasks where several sequence hypotheses are worth retaining. Open-ended chat more often uses sampling because ordinary beam search tends to concentrate on similar high-probability continuations.<\/p>\n A decoder raises its beam width from 4 to 32 and receives a higher model score but a worse human evaluation. Explain why the larger search did what it was asked to do without proving that the model's scoring rule matched human judgment.<\/p>\n","category":"inference-and-generation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"greedy-decoding","url":"https:\/\/darkfactory.dev\/glossary\/greedy-decoding"},{"slug":"sampling","url":"https:\/\/darkfactory.dev\/glossary\/sampling"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"The HARPY Speech Recognition System","url":"https:\/\/iiif.library.cmu.edu\/file\/Newell_box00103_fld07932_doc0001\/Newell_box00103_fld07932_doc0001.pdf"},{"title":"Hugging Face: Generation strategies","url":"https:\/\/huggingface.co\/docs\/transformers\/generation_strategies"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/benchmark","slug":"benchmark","term":"Benchmark","definition":"A defined package of tasks, data, metrics, and execution rules used to compare systems under the same evaluation regime.","definition_html":" A benchmark is a defined package of tasks, data, metrics, and execution rules used to compare systems. For an agent, the execution rules matter as much as the question set: the harness, prompts, tool access, time and token budgets<\/a>, retry policy, grader, and number of trials can all change the result.<\/p>\n A score is comparable only inside a sufficiently matched regime. If one system receives a browser, five retries, and a larger reasoning budget while another receives none of them, the benchmark is measuring two system configurations rather than isolating model capability. Good reports publish enough configuration and run-level evidence to reproduce that distinction.<\/p>\n Benchmarks also decay as evidence. Public tasks can leak into training or retrieval corpora, teams can tune to the leaderboard, and a fixed set can saturate. Even without leakage, stochastic agents and model-based graders add noise. OpenAI's GPT-4 report, for example, notes that contamination checks can produce false positives and false negatives. A high benchmark score should therefore be treated as evidence about a defined test, not as a certificate of production quality.<\/p>\n Two reports cite the same benchmark score, but one used pass@1 and the other selected the best of eight candidates. Are the results comparable? What configuration details would you need?<\/p>\n","category":"evaluation-and-reliability","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"verification","url":"https:\/\/darkfactory.dev\/factory\/verification"}],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"AgentAtlas: Control-Decision Taxonomy","url":"https:\/\/arxiv.org\/abs\/2605.20530"},{"title":"Infrastructure noise moves eval scores more than model margins","url":"https:\/\/www.anthropic.com\/engineering\/infrastructure-noise"},{"title":"Demystifying evals for AI agents","url":"https:\/\/www.anthropic.com\/engineering\/demystifying-evals-for-ai-agents"},{"title":"GPT-4 Technical Report","url":"https:\/\/cdn.openai.com\/papers\/gpt-4.pdf"},{"title":"Where Does Agent Reliability Come From?","url":"https:\/\/arxiv.org\/abs\/2607.17044"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/blast-radius","slug":"blast-radius","term":"Blast radius","definition":"The plausible scope of damage, exposure, or irreversible change if a component, credential, or action fails or is compromised.","definition_html":" The blast radius<\/strong> is the plausible scope of damage, exposure, or irreversible change if a component, credential, or action fails or is compromised. It can be measured across users, tenants, records, money, systems, environments, geography, time, or downstream dependencies.<\/p>\n The phrase comes from the physical area affected by an explosion and is now common in security and reliability engineering. In agent systems, the useful unit is the authority available at the moment of failure. A modest model error paired with production credentials and broad network access can have a larger blast radius than a severe reasoning failure trapped in a disposable, offline sandbox.<\/p>\n Risk combines likelihood and consequence. Model evaluation works mainly on likelihood. Least privilege<\/a>, tenant isolation, transaction limits, canaries, staged rollout, reversible operations, human approval, and fast credential revocation reduce consequence. These controls remain valuable even when failure prediction is uncertain.<\/p>\n Distinguish the theoretical maximum<\/strong> allowed by credentials from the effective radius<\/strong> limited by runtime controls, rate limits, monitoring, and recovery. Test whether those controls work under the same compromise you are assuming.<\/p>\n Attack surface describes ways a system can be reached or exploited. Blast radius describes how far harm can spread after failure. Blast containment is the set of boundaries that restrict that spread. Rollback can reduce duration and persistence but may not undo disclosure or an external irreversible action.<\/p>\n An agent can delete one temporary table per run, or it can use an administrator credential across every production database. The model and task are identical. Which deployment permits more autonomy? The first has a smaller, more recoverable blast radius and a clearer verification boundary.<\/p>\n","category":"security-and-governance","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"execution-environments","url":"https:\/\/darkfactory.dev\/factory\/execution-environments"},{"slug":"factory-assurance","url":"https:\/\/darkfactory.dev\/factory\/factory-assurance"}],"evidence":[{"title":"NeuralTrust: post-mortem of the 9-second AI database deletion","url":"https:\/\/neuraltrust.ai\/blog\/pocketos-railway-agent"},{"title":"Agentic Autonomy Levels","url":"https:\/\/addyosmani.com\/blog\/agentic-autonomy-levels\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/calibration","slug":"calibration","term":"Calibration","definition":"The degree to which a system's stated probabilities match observed outcome frequencies across comparable predictions.","definition_html":" Calibration asks whether stated probabilities match observed frequencies across comparable predictions. If a well-calibrated classifier assigns 0.8 probability to many cases, roughly 80 percent of those cases should be correct. Calibration is a property of a collection of forecasts, not something one confident answer can establish.<\/p>\n Accuracy and calibration measure different things. A system can choose the right label often while assigning probabilities that are too extreme. It can also be calibrated but uninformative: when 10 percent of cases are positive, predicting 0.1 for every case may be calibrated while doing nothing to separate one case from another.<\/p>\n Reliability diagrams compare predicted confidence with observed frequency. Expected calibration error summarizes binned gaps but is sensitive to how bins are chosen. Proper scoring rules such as log loss and the Brier score use the full probability forecast. Glenn Brier introduced the scoring rule later named for him in 1950. In the binary case, it is the mean squared difference between the predicted probability and the zero-or-one outcome, with lower values better.<\/p>\n Calibration can break under distribution shift and should be rechecked on the population where probabilities will be used. Chuan Guo and coauthors showed that post-hoc temperature scaling can improve classifier calibration on the datasets they studied. That method rescales logits for probability estimates; it is distinct from decoding temperature during text generation.<\/p>\n Among 100 predictions assigned 70 percent confidence, 50 are correct. Is that group calibrated, and what does the gap suggest?<\/p>\n","category":"evaluation-and-reliability","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":["confidence calibration"],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"verification","url":"https:\/\/darkfactory.dev\/factory\/verification"}],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"AgentAtlas: Control-Decision Taxonomy","url":"https:\/\/arxiv.org\/abs\/2605.20530"},{"title":"On Calibration of Modern Neural Networks","url":"https:\/\/arxiv.org\/abs\/1706.04599"},{"title":"Verification of Forecasts Expressed in Terms of Probability","url":"https:\/\/doi.org\/10.1175\/1520-0493(1950)078%3C0001:VOFEIT%3E2.0.CO;2"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/capability","slug":"capability","term":"Capability","definition":"A bounded action or class of action that a system component can perform, usually paired with explicit authority and scope.","definition_html":" A bounded action or class of action that a system component can perform. In secure agent systems, a useful capability description also states the authority attached to that action: which identity may use it, on which resources, for how long, in which environment, and under what conditions.<\/p>\n Keeping ability and authority separate prevents a common error. A model may know how to compose a database deletion request. A tool may technically be able to execute it. Neither fact means this run is authorized to delete production data. The runtime should grant the narrowest executable authority that satisfies the task, then enforce it below the model's instruction layer.<\/p>\n Capabilities make authority composable. Instead of handing an agent a broad credential and relying on a prompt, a system can expose a small operation such as A feature is product behavior available to a user. A permission is a policy decision about access. A tool is an interface the model can call. A capability joins a bounded operation with the conditions under which it can be exercised. Protocol specifications also use capability more narrowly for optional features a participant supports; context should make the sense clear.<\/p>\n An agent needs to attach a receipt to one expense report. Should it receive the employee's full cloud-drive token? No. Give it a capability restricted to the named file, report, action, identity, and time window, then revoke or expire it after the run.<\/p>\n","category":"tools-and-protocols","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"mcp-capability-negotiation","url":"https:\/\/darkfactory.dev\/glossary\/mcp-capability-negotiation"}],"related_factory_areas":[{"slug":"tools-interfaces","url":"https:\/\/darkfactory.dev\/factory\/tools-interfaces"},{"slug":"execution-environments","url":"https:\/\/darkfactory.dev\/factory\/execution-environments"}],"evidence":[{"title":"Model Context Protocol Specification","url":"https:\/\/modelcontextprotocol.io\/docs\/learn\/architecture"},{"title":"ActPlane: OS-Level Policy Enforcement","url":"https:\/\/arxiv.org\/abs\/2606.25189"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/causal-language-model","slug":"causal-language-model","term":"Causal language model","definition":"A language model trained to predict each token from the tokens that precede it, without access to future positions.","definition_html":" A causal language model assigns a sequence probability one token at a time. For tokens x1 through xn, it models each xt using only x1 through x(t-1). During Transformer training, a causal attention mask blocks each position from reading later tokens. The training system can still calculate losses for all positions in parallel because the complete training sequence is already known.<\/p>\n At generation time, the model samples or selects one token, appends it to the context, and repeats. Generation is therefore sequential even when training is highly parallel. \"Causal\" describes the direction of information flow in the sequence. It does not mean the model performs causal inference about interventions in the world.<\/p>\n OpenAI's 2018 generative pretraining paper used a Transformer language model with this left-to-right objective, then fine-tuned the same model for downstream tasks. It is an influential decoder-style example, not the origin of next-token language modeling.<\/p>\n A masked language model<\/a> predicts hidden positions using context on both sides. A causal language model cannot inspect future positions. \"Autoregressive\" is broader: it describes any model that factors an output into conditional steps, including models outside language.<\/p>\n A training example contains \"the build is green.\" When the model is scored on \"green,\" which tokens may contribute to that prediction, and which token must the mask hide?<\/p>\n","category":"models-and-training","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"masked-language-model","url":"https:\/\/darkfactory.dev\/glossary\/masked-language-model"},{"slug":"decoder-only-model","url":"https:\/\/darkfactory.dev\/glossary\/decoder-only-model"},{"slug":"token","url":"https:\/\/darkfactory.dev\/glossary\/token"},{"slug":"pretraining","url":"https:\/\/darkfactory.dev\/glossary\/pretraining"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Improving Language Understanding by Generative Pre-Training","url":"https:\/\/cdn.openai.com\/research-covers\/language-unsupervised\/language_understanding_paper.pdf"},{"title":"Attention Is All You Need","url":"https:\/\/arxiv.org\/abs\/1706.03762"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/chain-of-thought-prompting","slug":"chain-of-thought-prompting","term":"Chain-of-thought prompting","definition":"Prompting a model with worked intermediate reasoning, or asking it to generate intermediate steps, before producing an answer.","definition_html":" Chain-of-thought prompting supplies worked examples with intermediate reasoning steps, asks the model to produce such steps, or combines both approaches. The generated steps can help a model decompose some arithmetic, symbolic, and commonsense problems before it commits to an answer.<\/p>\n Jason Wei and coauthors gave the method its current name in their January 2022 paper, Chain-of-Thought Prompting Elicits Reasoning in Large Language Models<\/em>. Their experiments used few-shot examples containing a question, a chain of intermediate steps, and a final answer. The reported gains were empirical results for the models and tasks tested, not proof that visible reasoning always improves an answer.<\/p>\n A chain of thought is model output, not a trustworthy execution log. Miles Turpin and coauthors showed that models can produce plausible explanations for answers influenced by hidden biasing features while failing to mention that influence. The reasoning may still be useful as a draft, a decomposition, or an object to critique, but it requires the same independent checks as the final answer.<\/p>\n Applications also differ in whether they expose reasoning text, keep a private scratchpad, or return only a concise answer. Those interface choices should not be mistaken for evidence about how the model internally computed the result.<\/p>\n A model gives the correct numerical answer with a persuasive derivation. Identify one check that tests the calculation without relying on the model's own explanation.<\/p>\n","category":"inference-and-generation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"reasoning-model","url":"https:\/\/darkfactory.dev\/glossary\/reasoning-model"},{"slug":"prompt-chaining","url":"https:\/\/darkfactory.dev\/glossary\/prompt-chaining"}],"related_factory_areas":[],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models","url":"https:\/\/arxiv.org\/abs\/2201.11903"},{"title":"Language Models Don't Always Say What They Think","url":"https:\/\/arxiv.org\/abs\/2305.04388"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/chatbot","slug":"chatbot","term":"Chatbot","definition":"A conversational software interface that accepts natural-language input and returns responses, whether powered by rules, retrieval, generative models, or combinations of them.","definition_html":" A chatbot is a conversational software interface that accepts natural-language input and returns responses. It may use fixed rules, search, retrieval, classifiers, generative models, tools, or a combination of them. The defining feature is the conversational interface, not a particular model architecture.<\/p>\n That distinction matters when evaluating behavior. A wrong answer may come from the model, a retrieval source, a business rule, stale memory, or a tool. The chat surface can make those components feel like one speaker, but they have different owners, failure modes, and controls.<\/p>\n A chatbot is an interface or application, not a synonym for a large language model<\/a>. A chatbot becomes agentic when the surrounding system can choose and sequence actions over a loop. Conversation alone does not establish agency.<\/p>\n A support chatbot retrieves a policy paragraph and answers a question in one turn. It is still a chatbot, not necessarily an agent. Adding a chat window to a workflow does not change the workflow's control structure.<\/p>\n","category":"agents-and-automation","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":["chatbots"],"created_at":"2026-08-05T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[{"slug":"ai-system","url":"https:\/\/darkfactory.dev\/glossary\/ai-system"},{"slug":"large-language-model","url":"https:\/\/darkfactory.dev\/glossary\/large-language-model"},{"slug":"ai-agent","url":"https:\/\/darkfactory.dev\/glossary\/ai-agent"},{"slug":"system-prompt","url":"https:\/\/darkfactory.dev\/glossary\/system-prompt"},{"slug":"anthropomorphism","url":"https:\/\/darkfactory.dev\/glossary\/anthropomorphism"}],"related_factory_areas":[],"evidence":[{"title":"Andreessen Horowitz AI Glossary","url":"https:\/\/a16z.com\/ai-glossary\/"},{"title":"Stanford HAI Artificial Intelligence Glossary","url":"https:\/\/hai.stanford.edu\/ai-definitions"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/chunking","slug":"chunking","term":"Chunking","definition":"Splitting documents or data into retrieval units that can be indexed, selected, and placed into model context.","definition_html":" Chunking splits documents or other data into retrieval units that can be indexed, ranked, selected, and placed into model context. A chunk might be a fixed token range, paragraph, section, code symbol, table, transcript turn, or node in a document hierarchy.<\/p>\n The boundary choice changes what retrieval can recover. Small chunks can match a narrow question precisely but lose definitions, exceptions, or relationships outside the boundary. Large chunks preserve more local context but can dilute the matching signal, consume more context, and return irrelevant material. Overlap can preserve text near a boundary, though it also duplicates storage and may cause repeated passages to crowd a result set.<\/p>\n Useful strategies include fixed-size windows, structure-aware splitting, semantic boundaries, parent-child chunks, and late or query-time grouping. No chunk size is universally best. The right unit depends on document structure, question type, embedding model, ranking method, context budget, and whether answers require relationships across sections.<\/p>\n A chunk is a retrieval projection, not the canonical document. Store a stable source identifier, offsets or structural path, version, and surrounding hierarchy. This makes it possible to show citations, retrieve neighboring material, rebuild the index, and detect when a chunk has become stale.<\/p>\n Evaluate chunks through end-to-end questions, not visual preference. Measure whether retrieval finds the supporting passage, whether the returned unit contains enough evidence, how often neighboring chunks are required, and how answer quality changes with size and overlap. Test headings, tables, lists, code, and very short or long sections separately because naive splitters often damage them.<\/p>\n A policy exception appears in the paragraph after the rule it modifies. If the splitter separates them without a parent link or overlap, retrieval may return a confident rule with its exception missing.<\/p>\n","category":"context-and-knowledge","definition_status":"stable","search_index":true,"search_index_reason":"Reviewed as a distinct, evidence-backed definition with explicit relationships and a scenario-based comprehension check.","search_reviewed_at":"2026-08-26","aliases":[],"link_forms":[],"created_at":"2026-08-03T00:00:00-04:00","updated_at":"2026-08-26T00:00:00-04:00","related_terms":[],"related_factory_areas":[{"slug":"context-memory-skills","url":"https:\/\/darkfactory.dev\/factory\/context-memory-skills"}],"evidence":[{"title":"Google Machine Learning Glossary","url":"https:\/\/developers.google.com\/machine-learning\/glossary\/"},{"title":"Microsoft GraphRAG Documentation","url":"https:\/\/microsoft.github.io\/graphrag\/"}]},{"id":"https:\/\/darkfactory.dev\/glossary\/clanker","slug":"clanker","term":"Clanker","definition":"A derogatory slang term for a robot, AI system, or automated technology, used jokingly or hostilely to express disdain for machines or their perceived replacement of human work.","definition_html":"Why it is contested<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Analyze mechanism and consequence<\/h2>\n
Why it is contested<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Why it is contested<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Operational significance<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Safety is a continuing case<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Where the term came from<\/h2>\n
Why it is contested<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Manage composition and change<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Set the boundary before making a claim<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Readiness is an evidence claim<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
(true positives + true negatives) \/ all evaluated examples<\/code>. That compact formula hides several choices: what counts as an example, which label is positive, how abstentions and partial answers are scored, and whether each example receives equal weight.<\/p>\nDistinguish it from nearby terms<\/h2>\n
\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
ReLU's role in modern networks<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Early methods<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Where the term came from<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
A strong defense with real tradeoffs<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Treat it as an advertisement<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Where the framework came from<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
DarkFactory operating model<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Operational significance<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Interoperability stops at the trust boundary<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
What changes in practice<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Why the distinction matters<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Where the word came from<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Operational significance<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Why it matters<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
What agents need from an API<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Origin and adoption of the term<\/h2>\n
Why it is contested<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Operational significance<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Where the name came from<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Operational significance<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
From safety-critical engineering to AI systems<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Development of the mechanism<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Operational significance<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Deep autoencoders<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Why it is contested<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
The 1986 account<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Origin and usage<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Reduce reach before improving prediction<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Capability as an architectural unit<\/h2>\n
read_invoice<\/code> or refund_order<\/code> with tenant, amount, identity, and time constraints. The capability can carry audit and revocation behavior that a natural-language rule cannot enforce by itself.<\/p>\nDistinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Generative pretraining<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
What the reasoning text proves<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n
Preserve the route back to the source<\/h2>\n
Distinguish it from nearby terms<\/h2>\n
\n
Operational significance<\/h2>\n
Check your understanding<\/h2>\n
Definition<\/h2>\n