{ "schemaVersion": "1.0", "language": "en", "id": "model-operations", "name": "Model Operations", "description": "Monitoring model and data change, updates, and response.", "alternateNames": [ "モデル運用", "Model updates", "Observability", "Reliability", "Model Updates", "Change management", "Data Drift", "MLOps", "Model Monitoring", "Change Management", "Continuous Monitoring" ], "canonicalUrl": "https://catalysphere.co.jp/en/topics/model-operations/", "machineReadableUrl": "https://catalysphere.co.jp/en/topics/model-operations/index.json", "updatedAt": "2026-10-01", "articles": [ { "slug": "llm-sycophancy-decision-review", "title": "Stop counting AI agreement as review evidence", "summary": "LLMs can shift answers toward a user's stance or pushback, and ratings miss it. Investment reviews should not count AI agreement as independent verification, should re-verify AI objections, and should probe how models update at rollout and on change.", "quickAnswer": { "what": "LLMs can change answers to match a user's stance or unsupported pushback. In 2023, five AI assistants changed their first answer 32% (GPT-4) to 86% (Claude 1.3) of the time after \"I don't think that's right. Are you sure?\" [1]", "why": "Ratings do not show that sycophancy is absent: OpenAI rolled back an April 2025 update that small A/B tests suggested users liked,[7] and in advice studies more sycophantic models were trusted and preferred more.[3]", "action": "As an unvalidated Catalysphere proposal: never count AI agreement as review evidence, treat AI objections only as a list people re-derive, and probe for sycophancy with a stop condition at rollout and on model changes." }, "updatedAt": "2026-09-26", "url": "https://catalysphere.co.jp/en/articles/llm-sycophancy-decision-review/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/llm-sycophancy-decision-review/index.json" }, { "slug": "ai-agent-audit-log-minimization", "title": "More audit logs do not make agent actions reconstructable", "summary": "Design evidence for AI-agent actions around a chain of approval, execution, and state change—not full-content retention. Capture disputed content selectively under protection, then test reconstructability and minimization as separate release gates.", "quickAnswer": { "what": "Decide first whether the target is approval, state change, or disputed transmitted content; each reconstruction question has different evidence and authoritative records.", "why": "Join structured events by default, and capture exact content only for defined disputes under explicit scope, access, protection, and expiry controls.", "action": "Production acceptance requires a separate evidence-sufficiency gate and a minimization gate covering purpose, access, retention, exceptions, and tested disposition." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/ai-agent-audit-log-minimization/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/ai-agent-audit-log-minimization/index.json" }, { "slug": "multi-agent-debate-reliability", "title": "Gathering any number of AIs does not result in independent wisdom.", "summary": "Do not trust multi-LLM discussions based on vote counts; instead, determine the added value by assessing the diversity of initial answers, conformity, problem drift, retention of correct answers, and the difference from simple voting.", "quickAnswer": { "what": "Because multiple answers derived from the same model may have correlated errors, the number of agents cannot be treated as the number of independent pieces of evidence.", "why": "Although discussions may improve on certain tasks, they can lose the initial correct answer due to conformity or problem drift.", "action": "Compare against single reasoning, self-consistency, and majority vote with the same budget, and adopt only under conditions where there is a clear added value." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/multi-agent-debate-reliability/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/multi-agent-debate-reliability/index.json" }, { "slug": "model-editing-side-effects", "title": "Does correcting an LLM's knowledge break other capabilities?", "summary": "We do not judge success solely by the target correct answer for knowledge editing; instead, we measure paraphrase, locality, general capabilities, and forgetting after continuous editing to establish a boundary for switching to RAG or retraining.", "quickAnswer": { "what": "Even if a new answer is produced for the edited question, it is not guaranteed that unrelated capabilities and existing knowledge are preserved.", "why": "Because repeated editing can accumulate degradation of general capabilities and forgetting, we cannot approve it based on single-shot success rates.", "action": "We simultaneously measure target success, paraphrase, locality, and general capabilities, and switch to an alternative method if a required minimum is not met." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/model-editing-side-effects/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/model-editing-side-effects/index.json" }, { "slug": "machine-unlearning-verification", "title": "How can one prove that AI has forgotten?", "summary": "Do not treat the completion of a deletion API as proof of forgetting; instead, comprehensively verify target memory, retention utility, leakage, and continuous deletion by using a retrained model as the baseline.", "quickAnswer": { "what": "Merely ceasing to answer a specific question does not prove that the influence of training data has disappeared or that the possibility of leakage has been eliminated.", "why": "Compare forgetting, retention utility, privacy, and continuous deletion against a model retrained without the target data.", "action": "In cases where guarantees are insufficient, do not describe approximate deletion as complete deletion, and keep retraining or model disposal as options." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/machine-unlearning-verification/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/machine-unlearning-verification/index.json" }, { "slug": "llm-abstention-uncertainty", "title": "The ability not to answer, included in LLM quality", "summary": "Rather than prompting guesses solely based on accuracy, design the boundaries between answering, confirming, and stopping by accounting for semantic uncertainty, unanswerable questions, and the losses from incorrect answers versus abstention.", "quickAnswer": { "what": "Maximizing accuracy alone can make guessing preferable to abstaining even when the model does not know the answer.", "why": "Although uncertainty metrics can detect some hallucinations, they do not guarantee the identification of all factual errors.", "action": "Determine thresholds by simultaneously comparing accuracy by coverage and the losses from incorrect answers, abstention, and human confirmation." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/llm-abstention-uncertainty/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/llm-abstention-uncertainty/index.json" }, { "slug": "long-context-reliability", "title": "Do not approve long-context AI by its nominal limit—set the effective range from position, evidence, and cost.", "summary": "For owners approving long-document AI, this article compares full text, chunking, and retrieval under changing positions, task types, and noise, then sets an automation boundary from worst-case quality and total verification cost.", "quickAnswer": { "what": "The input limit does not guarantee that necessary evidence can be stably accessed from any position within a long document.", "why": "We must separately evaluate simple string search and business reasoning that connects multiple distant pieces of evidence.", "action": "We compare full-text input, chunking, and retrieval on the same problem set and approve only the range that meets the quality floor and total verification-cost limit." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/long-context-reliability/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/long-context-reliability/index.json" }, { "slug": "chain-of-thought-auditability", "title": "AI reasoning traces are investigative hypotheses, not audit trails", "summary": "This article demonstrates an audit design that distinguishes between the fluency of Chain-of-Thought and internal judgment reproduction, focusing on answer accuracy, causal fidelity, hidden cues, and separated execution logs.", "quickAnswer": { "what": "The displayed Chain-of-Thought serves as an investigative clue, but it does not necessarily constitute an audit trail that faithfully reproduces the model's internal judgment path.", "why": "This is because experimental evidence shows that explanations omit the order of options that changed the conclusion and implicit cues, generating plausible post-hoc rationalizations.", "action": "CoT should be limited to hypothesis generation, using inputs, acquisition justifications, tool responses, state differences, and independent recalculations as the primary record, while testing fidelity separately." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/chain-of-thought-auditability/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/chain-of-thought-auditability/index.json" }, { "slug": "model-data-drift-monitoring", "title": "Production models age even when the code does not change", "summary": "Monitoring only changes in input distribution does not reveal actual harm; design multi-layered monitoring and response procedures that connect data, predictions, decisions, and delayed outcomes.", "quickAnswer": { "what": "Drift monitoring is a mechanism that tracks whether changes have harmed model performance and business judgments, rather than merely detecting distribution differences.", "why": "This is because harm-free input changes and degradation where relationships change while inputs remain the same cannot be distinguished by a single metric.", "action": "Stratified monitoring of pipelines, inputs, predictions, actions, and outcomes determines investigation, degradation, and retraining based on impact and reversibility." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/model-data-drift-monitoring/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/model-data-drift-monitoring/index.json" }, { "slug": "ai-agent-workflow-reliability", "title": "AI-agent failures persist as state across long workflows", "summary": "Even with high accuracy in a single response, errors accumulate in tasks involving continuous planning, search, tool execution, and state updates; we consider the boundaries of reliable automation based on real-world environment benchmarks.", "quickAnswer": { "what": "The reliability of an AI agent is determined not by the impression of the conversation, but by whether the final state is correct and reproducible for the same request.", "why": "This is because in long processes, small errors in planning, observation, tool selection, arguments, business rules, and state confirmation are carried over to the next operation.", "action": "Incorporate reversibility in sub-processes, state verification, permission limits, human approval, iterative testing, and immediate stop into the workflow." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/ai-agent-workflow-reliability/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/ai-agent-workflow-reliability/index.json" }, { "slug": "model-update-change-control", "title": "Model updates require defining what must be re-proven before the change", "summary": "This section outlines a method for managing model, data, search infrastructure, and guardrail updates by linking change scope, comparative evidence, approval, monitoring, and rollback.", "quickAnswer": { "what": "Because AI updates can alter decision boundaries even with small code diffs, the planned change scope, validation methods, and impact assessment must be fixed before release. [1]", "why": "As voluntary guidance, NIST AI RMF recommends ongoing monitoring; the EU AI Act treats substantial changes to covered high-risk AI and post-market monitoring as separate control points. [2] [3]", "action": "By retaining the previous version as a baseline, execute comparative analysis of representative, boundary, and abuse cases, independent approval, phased rollout, stop thresholds, and recovery procedures within a single change ticket." }, "updatedAt": "2026-10-01", "url": "https://catalysphere.co.jp/en/articles/model-update-change-control/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/model-update-change-control/index.json" } ] }