{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp11_mechanism_context_external20_20260804","research_id":"eoa_inverse_innovation_exp11_external_scrutiny_20260804","cell_id":"negative_space_design__computer_science","opaque_id":"negative_space_design__computer_science__C","search_lanes":{"direct_problem":{"queries":["AI code completion interruptions developer focus inline suggestions study","AI coding assistant automation bias premature acceptance study","\"An Eye for AI\" \"Micro-Interruptions\" PDF"],"source_ids":["SRC1","SRC2","SRC3","SRC4","SRC8"],"no_result_note":null},"closest_prior_art":{"queries":["context aware suppress AI code completion after rejection cooldown","\"When to Show a Suggestion?\" Mozannar Bansal Fourney Horvitz 2024","\"conditional suggestion display from human feedback\"","\"Optimizing LLM Code Suggestions\" \"Feedback-Driven Timing\""],"source_ids":["SRC1","SRC2","SRC3"],"no_result_note":null},"historical_terminology":{"queries":["older terminology mixed-initiative code completion interruption management programmers","attention-aware notification deferral task boundaries interruption 2005 Iqbal Bailey","\"Reading Between the Lines\" AI-assisted programming verification cost 2022","\"Don't Complete It\" unhelpful code completion 2022"],"source_ids":["SRC1","SRC2","SRC7"],"no_result_note":"Older searches used the terms mixed-initiative interaction, attention-aware systems, interruption management, notification deferral, early rejection, and verification cost. The eight-source cap favored stronger code-specific and official evidence over additional generic interruption studies."},"products_practices_standards":{"queries":["GitHub Copilot pause suggestions snooze inline completions","site:code.visualstudio.com \"Snooze inline suggestions\"","site:learn.microsoft.com Visual Studio \"Show inline suggestions only after a pause in typing\"","ISO 9241 interruptions software notifications user control guidance"],"source_ids":["SRC5","SRC6","SRC7","SRC8"],"no_result_note":null},"non_english_regional":{"queries":["KI Codevervollständigung Vorschläge unterbrechen Entwickler Konzentration Vorschläge pausieren","AI コード補完 提案 一時停止 集中 割り込み IDE","AI 代码补全 建议 暂停 专注 打断 开发者","sugerencias código IA pausar interrupciones desarrolladores autocompletado"],"source_ids":["SRC5","SRC8"],"no_result_note":null},"composition_subproblems":{"queries":["code completion telemetry rapid typing rejection manual invocation filtering","adaptive code suggestion delay cognitive state debugging acceptance rejection","bounded pause inline suggestions visible status manual resume","AI coding assistant suppress suggestions preserve diagnostics accessibility"],"source_ids":["SRC1","SRC2","SRC3","SRC5","SRC6","SRC7","SRC8"],"no_result_note":null}},"sources":[{"source_id":"SRC1","title":"When to Show a Suggestion? Integrating Human Feedback in AI-Assisted Programming","url":"https://arxiv.org/abs/2306.04930","publisher":"AAAI / arXiv","date_or_year":"2023; revised and published 2024","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["CDHF uses Copilot interaction telemetry, including acceptance and rejection, to decide whether to display or withhold suggestions.","The retrospective evaluation used data from 535 programmers and found that a substantial fraction of suggestions predicted to be rejected could be hidden.","The work explicitly models programmer latent state and verification time, but warns that optimizing acceptance can favor lower-quality suggestions."]},{"source_id":"SRC2","title":"A Transformer-Based Approach for Smart Invocation of Automatic Code Completion","url":"https://pure.tudelft.nl/ws/portalfiles/portal/217678128/3664646.3664760.pdf","publisher":"Association for Computing Machinery / Delft University of Technology","date_or_year":"2024","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["The paper treats overly frequent or ill-timed code completions as intrusive and potentially disruptive to programming flow.","Its Code4Me implementation uses code context and IDE telemetry to predict whether to invoke completion, while treating manual invocation as evidence that assistance is wanted.","The approach was evaluated online with 34 developers and 74,000 invocations, demonstrating technical feasibility of telemetry-based suppression."]},{"source_id":"SRC3","title":"Optimizing LLM Code Suggestions: Feedback-Driven Timing with Lightweight State Bounds","url":"https://arxiv.org/abs/2511.18842","publisher":"IEEE/ACM ASE Workshops / arXiv","date_or_year":"2025","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["A deployed VS Code assistant varied suggestion delay within bounded ranges using recent acceptance and rejection plus a coarse implementing-versus-debugging state classifier.","The study reports higher acceptance and fewer near-instant blind rejections under static and adaptive delays than under no delay.","The evaluation demonstrates feasibility of bounded, state-sensitive suggestion timing but optimizes sub-second delay rather than event-triggered multi-second suggestion-free episodes."]},{"source_id":"SRC4","title":"An Eye for AI: Eye-Tracking the Micro-Interruptions of GenAI Code Suggestions","url":"https://conf.researchr.org/details/icse-2026/icse-2026-research-track/90/An-Eye-for-AI-Eye-Tracking-the-Micro-Interruptions-of-GenAI-Code-Suggestions","publisher":"ICSE 2026 Research Track","date_or_year":"2026","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Eye tracking and interaction logs from 35 professional and student developers found that roughly half of generated suggestions were not viewed.","More than 75% of viewed suggestions were not used.","The authors characterize viewed suggestions as brief micro-interruptions and call for more context-aware presentation that reduces cognitive overhead."]},{"source_id":"SRC5","title":"Autocompletado de Copilot - Visual Studio (Windows)","url":"https://learn.microsoft.com/es-es/visualstudio/ide/visual-studio-github-copilot-extension?view=visualstudio","publisher":"Microsoft Learn","date_or_year":"Updated 2026-05-08","source_type":"FIRST_PARTY_PRODUCT","language":"Spanish","claims_supported":["Visual Studio allows automatic inline completions to be replaced with manual invocation and offers a wait-for-pause debounce because fast completions can interrupt typing.","Copilot completions are automatically suppressed while IntelliSense is active and automatically resume after that state ends.","Manual keyboard invocation remains available, demonstrating recoverable user control over suggestion timing."]},{"source_id":"SRC6","title":"Inline suggestions from GitHub Copilot in VS Code","url":"https://code.visualstudio.com/docs/editing/ai-powered-suggestions","publisher":"Microsoft Visual Studio Code","date_or_year":"2026","source_type":"FIRST_PARTY_PRODUCT","language":"English","claims_supported":["VS Code provides a visible Snooze control that pauses inline suggestions in five-minute increments.","Users can cancel snooze to restore suggestions, and can enable or disable suggestions by language.","The product exposes a minimum-show-delay setting, showing that suggestion eligibility and timing are implementable editor controls."]},{"source_id":"SRC7","title":"Understanding Success Criterion 2.2.4: Interruptions","url":"https://www.w3.org/WAI/WCAG21/Understanding/interruptions","publisher":"World Wide Web Consortium Web Accessibility Initiative","date_or_year":"Updated 2025-09-16","source_type":"OFFICIAL_GUIDANCE","language":"English","claims_supported":["Users should be able to postpone or suppress interruptions except emergencies.","The guidance identifies distraction and discontinuity risks for users with cognitive, attention, visual, or assistive-technology needs.","It supports preserving urgent warnings and providing user-controlled postponement or manual updating."]},{"source_id":"SRC8","title":"AI Coding Assistants","url":"https://www.bsi.bund.de/SharedDocs/Downloads/EN/BSI/KI/ANSSI_BSI_AI_Coding_Assistants.pdf?__blob=publicationFile&v=7","publisher":"ANSSI and German Federal Office for Information Security (BSI)","date_or_year":"2025","source_type":"OFFICIAL_GUIDANCE","language":"English","claims_supported":["Official guidance identifies automation bias, uncritical acceptance, skill degradation, confidentiality, and security risks associated with AI coding assistants.","It recommends controlled enterprise access, clear organizational usage rules, and careful handling of sensitive code and credentials.","The guidance supports keeping human review and organizational authorization in scope rather than treating suggestion suppression as the only safety control."]}],"problem_evidence":{"status":"PARTLY_SUPPORTED","finding":"The timing-related interruption problem is externally supported: eye tracking identifies suggestion review as a micro-interruption, prior systems explicitly target intrusive or unwanted completions, and Visual Studio documentation acknowledges that suggestions appearing too quickly can interrupt typing. Evidence also supports verification cost and automation-bias risk. However, the stronger causal claims that cumulative relevant suggestions impair problem-state formation, cause premature acceptance, or increase defects and rework have not been directly established by these sources.","source_ids":["SRC1","SRC2","SRC3","SRC4","SRC5","SRC8"],"uncertainty":"Existing studies often use acceptance, rejection, gaze, or invocation metrics rather than task switching, mental-model quality, rework, defects, and experienced control. Effects may vary by expertise, task, suggestion quality, and whether the developer is exploring or implementing a known solution."},"adopter_evidence":{"status":"SUPPORTED","finding":"The identifiable immediate adopter is an individual developer using an AI-enabled editor, with repository owners or engineering leads able to authorize project use. First-party products already expose individual manual, delay, suppression, and snooze controls, while official security guidance calls for organizational usage rules and controlled access.","source_ids":["SRC5","SRC6","SRC8"],"uncertainty":"The exact allocation of authority varies by employer, repository policy, telemetry governance, accessibility accommodation, and whether code is processed locally or by a cloud service."},"implementation_evidence":{"status":"PARTLY_SUPPORTED","finding":"Major components are demonstrably implementable: telemetry-based conditional display, smart invocation, cognitive-state-dependent bounded delay, automatic suppression during an editor state, timed snooze, visible status, and manual restoration. No retained source demonstrates the complete package of automatically entering a clearly cued suggestion-free interval after rejection, rapid editing, or unresolved failure while preserving diagnostics and evaluating task-level control, rework, quality, and time.","source_ids":["SRC1","SRC2","SRC3","SRC5","SRC6","SRC7"],"uncertainty":"Rapid-edit and test-failure detectors may misclassify receptive moments; fixed intervals may delay useful completions; the cue may itself distract; and telemetry may expose sensitive behavioral or source-code information."},"prior_art":{"disposition":"ADJACENT_PRIOR_ART","closest_analogues":[{"name":"Conditional Suggestion Display from Human Feedback (CDHF)","source_ids":["SRC1"],"same_problem":true,"same_causal_lever":true,"overlap":"Uses acceptance/rejection telemetry and inferred programmer state to decide whether generated code suggestions should be displayed or withheld, targeting verification time and unwanted suggestions.","remaining_difference":"It selectively filters individual suggestions by predicted utility rather than creating an explicit, bounded quiet state after designated events; it does not supply the proposed visible quiet cue, immediate override protocol, diagnostic-preservation rule, or crossover test of control, rework, and quality."},{"name":"JonBERTa smart invocation for Code4Me","source_ids":["SRC2"],"same_problem":true,"same_causal_lever":true,"overlap":"Uses code context and IDE telemetry to suppress automatic invocation when assistance is unlikely to be wanted and preserves manual invocation as a strong user-intent signal.","remaining_difference":"It predicts invocation point by point and optimizes accepted or manually requested completions, rather than enforcing a recoverable cooldown after rejection, rapid editing, or unresolved diagnostics."},{"name":"Feedback-driven timing with lightweight state bounds","source_ids":["SRC3"],"same_problem":true,"same_causal_lever":true,"overlap":"Uses rejection history and an implementing-versus-debugging state estimate to impose bounded suggestion delays; a deployment reports fewer blind rejections and higher acceptance.","remaining_difference":"The quiet period is a sub-second presentation delay, not a visible event-triggered no-suggestion interval, and reported outcomes do not include task completion, defects, undo, experienced control, or missed critical cues."},{"name":"Visual Studio context suppression and VS Code timed snooze","source_ids":["SRC5","SRC6"],"same_problem":true,"same_causal_lever":true,"overlap":"Commercial editors already provide manual-only completion, wait-for-typing-pause behavior, automatic suppression during IntelliSense, bounded snooze, visible state, and manual restoration.","remaining_difference":"The controls are manual, debounce-based, or tied to IntelliSense rather than automatically entered after rejection, rapid-edit bursts, or unresolved test failure; they are not presented as a tested package with protected diagnostics and task-level outcome measures."}],"contrastive_claim_remaining":"For developers with verified timing-related disruption, automatically entering a clearly signaled, immediately reversible suggestion-free interval after recent rejection, rapid-edit bursts, or unresolved test failure will improve within-person interruption, experienced control, undo/rework, or quality relative to continuous eligibility and existing confidence/adaptive-delay controls, without exceeding preregistered completion-time or missed-cue costs.","contrastive_claim_falsifier":"In an adequately powered within-person comparison, the event-triggered quiet condition improves none of the preregistered interruption, control, rework, or quality outcomes, or any improvement is offset by slower completion, more defects, missed critical cues, or frequent manual overrides; equivalently, timing and density cease to predict disruption after task difficulty and suggestion quality are controlled.","confidence":"HIGH","search_limitations":"This was a bounded public-web search retaining exactly eight sources. It covered direct terms, prior-art terminology, older interruption and mixed-initiative terminology, products and standards, German/Japanese/Chinese/Spanish queries, and component combinations. It did not exhaust paywalled literature, private deployments, source-code history, patents, abandoned prototypes, or every regional market. The BSI PDF was indexed with substantive text but its direct open request returned HTTP 403."},"researchability_gates":{"externally_supported_problem":{"status":"PASS","rationale":"Multiple independent primary and first-party sources support unwanted, intrusive, or mistimed code suggestions and measurable attention or verification costs, although the proposed downstream defect and premature-acceptance pathway remains unproven.","source_ids":["SRC1","SRC2","SRC3","SRC4","SRC5"]},"identifiable_adopter_or_authorizer":{"status":"PASS","rationale":"Individual editor users can adopt and control the intervention; repository owners or engineering leads can authorize project use, and organizational security or privacy owners can authorize telemetry and cloud-processing conditions.","source_ids":["SRC5","SRC6","SRC8"]},"distinct_testable_incremental_claim":{"status":"PASS","rationale":"Close work filters individual suggestions or adjusts sub-second timing, and products provide manual snooze or suppression in other states. A distinct test remains for explicit, event-triggered quiet intervals after rejection, rapid editing, or unresolved failure, with visible status, manual recall, protected diagnostics, and task-level outcomes.","source_ids":["SRC1","SRC2","SRC3","SRC5","SRC6"]},"bounded_next_evidence_step":{"status":"PASS","rationale":"An opt-in, local crossover pilot on non-production tasks can compare continuous eligibility against one or more fixed cooldowns after rejection and rapid editing, preregistering interruption, control, undo, defects, completion time, manual overrides, and missed cues. Existing deployments show the necessary timing and telemetry mechanisms are feasible.","source_ids":["SRC2","SRC3","SRC5","SRC6"]},"no_unresolved_safety_or_authority_stop":{"status":"PASS","rationale":"No categorical stop is evident if participation and repository use are authorized, telemetry is minimized and kept within existing code-handling permissions, compiler/test/security/permission/data-loss warnings remain unsuppressed, and instant restoration plus rollback are verified. Accessibility and automation-bias concerns require monitoring but are manageable guardrails.","source_ids":["SRC7","SRC8"]},"adequate_search_evidence":{"status":"PASS","rationale":"All six adversarial lanes were searched. The eight retained sources span five independent publisher groups and consist entirely of primary research, official guidance, or first-party product documentation, including a Spanish-language first-party source and regional-language searches.","source_ids":["SRC1","SRC2","SRC3","SRC4","SRC5","SRC6","SRC7","SRC8"]}},"strict_success":true,"screen_survival":true,"remaining_research_value":"MODERATE","recommended_next_step":"Preregister and run a small opt-in within-person crossover pilot on non-production tasks. First screen participants for timing-related disruption; compare continuous eligibility, an existing adaptive-delay or pause-on-typing control, and fixed event-triggered quiet intervals after rejection and rapid-edit bursts. Preserve all non-AI diagnostics, show a persistent quiet-state cue, provide a keyboard override, keep telemetry local or appropriately authorized, and halt on preregistered bounds for defects, completion time, missed cues, or failed restoration.","world_novelty_boundary":"The search establishes adjacent prior art and a bounded incremental experimental claim only. It cannot establish world novelty, patentability, freedom to operate, market size, routine adoption of the exact package, or realized impact."}