{"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 interruption developer flow inline suggestions study","programming autocomplete cognitive load interruptions automation bias study"],"source_ids":["S1","S2","S4","S7"],"no_result_note":null},"closest_prior_art":{"queries":["adaptive timing code completion suggestions interruptibility IDE","When to Show a Suggestion human feedback AI-assisted programming","inline code suggestions quiet interval rejection test failure focus state"],"source_ids":["S2","S3","S4","S5","S8"],"no_result_note":null},"historical_terminology":{"queries":["attentive user interface interruptibility defer notifications task boundaries suggestions historical","site:microsoft.com research inline suggestions \"pause\" \"period of time\" code"],"source_ids":["S4"],"no_result_note":null},"products_practices_standards":{"queries":["GitHub Copilot disable suggestions temporarily cooldown rejection focus mode","IDE code completion delay debounce suggestions after typing settings","Copilot suggestions suppress after rejection debounce rapid typing debug mode manual trigger"],"source_ids":["S5","S6","S8"],"no_result_note":null},"non_english_regional":{"queries":["asistente código IA sugerencias distraen desactivar temporalmente autocompletado español","KI Codevervollständigung Vorschläge ablenkend Pause Tippen deaktivieren","AI コード補完 提案 邪魔 一時停止 入力中 表示しない","代码补全 建议 干扰 暂停 输入时 不显示 AI"],"source_ids":["S6"],"no_result_note":null},"composition_subproblems":{"queries":["AI code completion cooldown after rejection manual invoke status indicator diagnostics remain visible","Copilot suggestions suppress after rejection debounce rapid typing debug mode manual trigger","inline code suggestions quiet interval rejection test failure focus state","code completion state machine pause recent rejection unresolved diagnostics"],"source_ids":["S3","S5","S6","S8"],"no_result_note":null}},"sources":[{"source_id":"S1","title":"An Eye for AI: Eye-Tracking the Micro-Interruptions of GenAI Code Suggestions","url":"https://hasel.dev/publication/an-eye-for-ai-eye-tracking-the-micro-interruptions-of-genai-code-suggestions/","publisher":"Human Aspects of Software Engineering Lab, University of Zurich","date_or_year":"2026","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["A study of 33 developers combined eye tracking with fine-grained suggestion logs.","Approximately half of generated suggestions were not viewed, and more than 75% of viewed suggestions were unused.","Viewed suggestions averaged about 0.9 seconds of inspection and were characterized as flow-disrupting micro-interruptions."]},{"source_id":"S2","title":"Developer Interaction Patterns with Proactive AI: A Five-Day Field Study","url":"https://arxiv.org/abs/2601.10253","publisher":"ACM IUI / JetBrains Research and Delft University of Technology","date_or_year":"2026-01-15","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["A five-day field study examined 229 interventions across 5,732 interaction points with professional developers.","Interventions after declined AI edits had 62% dismissal and 31% engagement, while post-commit interventions had 52% engagement.","The implemented assistant used workflow-grounded triggers, lightweight cues, optional expansion, and on-demand invocation to preserve control.","Timing-dependent receptivity supports distinguishing focused implementation from receptive workflow boundaries."]},{"source_id":"S3","title":"When to Show a Suggestion? Integrating Human Feedback in AI-Assisted Programming","url":"https://ojs.aaai.org/index.php/AAAI/article/view/28878","publisher":"AAAI Press","date_or_year":"2024-03-24","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Conditional Suggestion Display from Human Feedback decides whether to show or withhold code suggestions using acceptance and rejection data.","A retrospective evaluation used 168,000 suggestions from 535 programmers.","The reported method could hide 25% of displayed suggestions while estimating that 95% of those hidden suggestions would have been rejected.","The authors warn that optimizing display decisions from acceptance alone can reduce suggestion quality."]},{"source_id":"S4","title":"Towards More Effective AI-Assisted Programming: A Systematic Design Exploration to Improve Visual Studio IntelliCode’s User Experience","url":"https://www.microsoft.com/en-us/research/wp-content/uploads/2023/05/icse-seip-2023.pdf","publisher":"Microsoft Research / IEEE ICSE-SEIP","date_or_year":"2023","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["Seven laboratory-study participants requested a snooze feature to pause inline suggestions for a period or session; one requested a 30-minute snooze.","Participants also proposed suppressing or varying suggestion frequency using recent acceptance behavior.","The paper reports that inline suggestions can interrupt code-authoring activity and that visual clutter can make developers lose context.","The broader design program studied 19 interfaces with 61 programmers and deployed selected IntelliCode interfaces at scale."]},{"source_id":"S5","title":"Better Control over Your Copilot Code Suggestions","url":"https://devblogs.microsoft.com/visualstudio/better-control-over-your-copilot-code-suggestions/","publisher":"Microsoft Visual Studio Blog","date_or_year":"2025-08-21","source_type":"FIRST_PARTY_PRODUCT","language":"English","claims_supported":["Visual Studio 2022 17.14.13 introduced generally available controls intended to maintain attention and a quieter editor.","A debounce option pauses completions during rapid typing and resumes after a typing pause.","Users can disable automatic completions and invoke them manually with keyboard shortcuts.","Next-edit suggestions can remain collapsed behind a margin indicator until the developer requests review, providing a visible availability cue."]},{"source_id":"S6","title":"IntelliCode 整行自动完成 - Visual Studio (Windows)","url":"https://learn.microsoft.com/zh-cn/visualstudio/ide/visual-studio-whole-line-completions?view=visualstudio","publisher":"Microsoft Learn","date_or_year":"2025-12-12","source_type":"OFFICIAL_GUIDANCE","language":"Chinese (Simplified)","claims_supported":["Official Chinese documentation describes automatic-versus-manual completion invocation.","It documents showing completions only after typing pauses for users who find the default distracting.","It documents disabling suggestions during debugging and hiding suggestion hint bars to minimize distraction.","The documentation provides regional terminology corroborating that distraction-oriented completion controls are productized outside English documentation."]},{"source_id":"S7","title":"Multi-line AI-assisted Code Authoring","url":"https://arxiv.org/abs/2402.04141","publisher":"ACM FSE / Meta","date_or_year":"2024-02-06","source_type":"PRIMARY_RESEARCH","language":"English","claims_supported":["CodeCompose supplied inline suggestions to tens of thousands of Meta developers.","The authors describe constantly moving multi-line suggestions as jarring and cognitively costly because developers must review and rework mixed human and generated code.","The source also provides counterevidence: multi-line suggestions doubled reported keystrokes saved from 9% to 17%, and fewer than 1% of engineers opted out."]},{"source_id":"S8","title":"Get started with GitHub Copilot completions","url":"https://learn.microsoft.com/en-gb/visualstudio/ide/visual-studio-github-copilot-extension?view=visualstudio","publisher":"Microsoft Learn","date_or_year":"2026-05-06","source_type":"OFFICIAL_GUIDANCE","language":"English (United Kingdom)","claims_supported":["Current Visual Studio documentation supports manual-only inline-suggestion invocation and explicit keyboard recall.","It supports debounce-based suppression while typing quickly because fast completions can interrupt typing.","Copilot completions are automatically suppressed while IntelliSense is active and resume when the IntelliSense interaction ends.","The product exposes margin indicators, keyboard controls, and appearance settings that preserve awareness and recoverability."]}],"problem_evidence":{"status":"PARTLY_SUPPORTED","finding":"The interruption component is directly supported: eye tracking found brief suggestion-driven micro-interruptions, field evidence found strong timing-dependent dismissal during focused work, and product research reports requests for snoozing intrusive suggestions. However, the stronger claims that relevant suggestions prevent problem-state formation or induce premature acceptance are not directly established. Meta's deployment also supplies counterevidence that useful suggestions can save keystrokes and retain users.","source_ids":["S1","S2","S4","S7"],"uncertainty":"Studies differ in suggestion type, task, developer experience, and outcome measures. Evidence does not isolate cumulative interruption from relevance, latency, task difficulty, or suggestion length, and it does not directly measure formation of a mental problem representation."},"adopter_evidence":{"status":"SUPPORTED","finding":"Individual developers are identifiable adopters because current IDEs expose per-user timing, manual-invocation, suppression, and appearance controls. Visual Studio/Copilot product teams are identifiable implementers, while an engineering lead or repository owner can authorize a project-level pilot.","source_ids":["S5","S6","S8"],"uncertainty":"Public sources establish available controls, not which organizational role would authorize telemetry in every workplace or how frequently developers use the optional quiet settings."},"implementation_evidence":{"status":"PARTLY_SUPPORTED","finding":"Most operations are already implemented in production: automatic silence during rapid typing or active IntelliSense, automatic resumption, manual recall, hidden suggestions with a visible indicator, and debugging-related suppression. No retained source demonstrates the complete proposed state machine combining fixed cooldowns after rejection, unresolved test failure, and rapid edits while preserving all critical diagnostics.","source_ids":["S5","S6","S8"],"uncertainty":"Documentation confirms feature availability and default IntelliSense suppression but provides no usage-rate or outcome data for the exact configuration. The specific rejection- and test-failure-triggered cooldown composition remains unverified."},"prior_art":{"disposition":"ESTABLISHED_PRACTICE","closest_analogues":[{"name":"Visual Studio Copilot quiet-editor and on-demand completion controls","source_ids":["S5","S8"],"same_problem":true,"same_causal_lever":true,"overlap":"Production controls suppress completions during rapid typing, permit manual-only invocation, collapse unsolicited next-edit suggestions behind a visible indicator, and automatically suppress then restore Copilot around an active IntelliSense interaction.","remaining_difference":"The documented product does not establish a fixed, experimentally calibrated cooldown after rejection or unresolved test failure, nor comparative effects on defects, undo, task time, and perceived control."},{"name":"Conditional Suggestion Display from Human Feedback (CDHF)","source_ids":["S3"],"same_problem":true,"same_causal_lever":true,"overlap":"CDHF selectively withholds suggestions using programmer telemetry and latent-state estimates, including acceptance and rejection behavior, to reduce rejected displays and verification time.","remaining_difference":"It is a utility and acceptance-prediction filter rather than a transparent bounded silence interval triggered by rapid edits, rejection, or failed tests; immediate manual override and a quiet-state cue are not its central intervention."},{"name":"IntelliCode snoozability design finding","source_ids":["S4"],"same_problem":true,"same_causal_lever":true,"overlap":"Developers explicitly requested time-bounded or session-bounded pauses, suppression, and frequency adaptation based on recent acceptance behavior.","remaining_difference":"The paper reports requested design features rather than an evaluation of an automatic cooldown state machine with quality and control outcomes."},{"name":"ProAIDE workflow-grounded proactive timing","source_ids":["S2"],"same_problem":true,"same_causal_lever":false,"overlap":"The system detects workflow states, minimizes presentation with a lightweight cue, requires optional user engagement, and shows that receptivity is lower after declined edits than at post-commit boundaries.","remaining_difference":"It selects better moments for proactive intervention rather than creating a suggestion-free interval; its after-decline behavior offers more assistance, opposite to the proposed rejection cooldown."}],"contrastive_claim_remaining":"Relative to existing Visual Studio debounce, manual-only, collapsed-suggestion, and IntelliSense-suppression controls, an automatic short cooldown specifically after suggestion rejection or unresolved test failure—while retaining compiler/security diagnostics, a quiet-state cue, and instant manual recall—reduces within-person interruption and rework without worsening task success or completion time.","contrastive_claim_falsifier":"In an adequately powered within-person crossover among developers who exhibit timing-related interruption, the added rejection/test-failure cooldown improves none of the preregistered interruption, undo/rework, defect, or control outcomes, or any improvement is offset by worse completion time, task success, or missed critical cues.","confidence":"HIGH","search_limitations":"This bounded search covered six lanes and eight opened direct sources through 2026-08-04, but it was not an exhaustive scholarly, patent, source-code, procurement, or product-telemetry review. Product documentation can change, Microsoft sources are disproportionately informative because Visual Studio exposes the closest controls, and public evidence does not reveal adoption rates for individual settings."},"researchability_gates":{"externally_supported_problem":{"status":"PASS","rationale":"Multiple direct studies and product investigations establish that inline or proactive code suggestions can impose cognitive cost, visual disruption, or poorly timed interruption, even though premature acceptance remains unproven.","source_ids":["S1","S2","S4","S7"]},"identifiable_adopter_or_authorizer":{"status":"PASS","rationale":"Developers can adopt the local controls, IDE product teams can implement the state machine, and engineering or repository leadership can authorize a project-work pilot.","source_ids":["S5","S6","S8"]},"distinct_testable_incremental_claim":{"status":"PASS","rationale":"The residual claim is narrowly distinguishable from existing debounce/manual modes and acceptance filters: test automatic bounded cooldowns after rejection and unresolved test failure with preserved diagnostics and override.","source_ids":["S2","S3","S5","S8"]},"bounded_next_evidence_step":{"status":"PASS","rationale":"An opt-in local crossover can compare current automatic eligibility or existing debounce controls against preregistered cooldown durations after rejection and failed tests, using interaction logs, task time, undo/rework, defects, missed cues, and control ratings.","source_ids":["S1","S2","S3"]},"no_unresolved_safety_or_authority_stop":{"status":"PASS","rationale":"The next step can be local, reversible, non-production, and participant-controlled. Critical diagnostics must remain visible, telemetry must exclude code content and personnel evaluation, and instant manual restoration must be verified before enrollment.","source_ids":["S6","S8"]},"adequate_search_evidence":{"status":"PASS","rationale":"All six required lanes were searched adversarially, including historical snooze/interruptibility terminology, current products, Chinese/Japanese/German/Spanish terms, and combined trigger-cue-override-diagnostic subproblems. Exactly eight direct sources were retained and opened, spanning multiple independent publishers and several primary or official sources.","source_ids":["S1","S2","S3","S4","S5","S6","S7","S8"]}},"strict_success":false,"screen_survival":false,"remaining_research_value":"MODERATE","recommended_next_step":"Run a preregistered, opt-in within-person pilot on non-production tasks. Compare the editor's existing pause-after-typing/manual controls with 5-, 15-, and 30-second cooldowns after explicit rejection and unresolved test failure; preserve all compiler, test, security, permission, and data-loss cues; record only event timing and outcomes; require a visible quiet-state indicator and verified keyboard override; stop on excess defects, time, or missed cues.","world_novelty_boundary":"The search establishes neither world novelty nor patentability. It finds that the broad problem–intervention package—temporarily withholding unsolicited AI code suggestions during contextually unsuitable activity, visibly preserving availability, and allowing manual recall—is already established in research and production IDE practice. The only bounded research boundary left here is the comparative effect of combining rejection- and failed-test-triggered cooldowns with explicit safety-preserving cues and override against today's debounce, manual, collapse, and predictive-withholding baselines."}