{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp05_complete_proposal_portfolio20_20260803","cell_id":"negative_space_design__linguistics_semiotics","arm":"COMPLETE_PROPOSAL_PORTFOLIO","candidate_id":"nsd_ls_004_blank_first_annotation_review","proposal_index":4,"version":0,"title":"Blank-First Review for Morphosyntactic Annotation","problem":"When a corpus team audits a disputed dependency relation or morphosyntactic label, the review interface may display the existing annotation, its confidence value, an automated suggestion, and earlier reviewer notes before the new reviewer examines the linguistic evidence. Even if the reviewer agrees, the workflow cannot distinguish an independent analysis from acceptance of the visible answer. The candidate problem is not excessive information generally; it is premature occupation of the judgment space by prior analyses that remain necessary later for comparison and repair.","actors":["Corpus annotators conducting targeted reviews","Adjudicators resolving disagreements","Annotation-scheme maintainers","Corpus lead responsible for releases and provenance","Researchers who rely on the resulting morphosyntactic annotations"],"observable_state":"For targeted review items, the interface prepopulates the disputed head, relation, feature, confidence, model output, or prior rationale before the reviewer records a first-pass judgment. The audit log contains no independently committed analysis made before that exposure, so apparent agreement with the inherited label has ambiguous provenance.","consequence":"Inherited mistakes or framing choices may survive review, agreement statistics may mix independent convergence with answer exposure, and adjudicators may receive fewer visible disagreements from which to identify unclear guidelines or genuinely ambiguous constructions.","affected_objective":"Obtain auditable first-pass morphosyntactic judgments that are independent of inherited labels while preserving complete access to prior analyses for comparison and adjudication.","intervention":"Create a blank-first review stage for bounded, targeted annotation audits. The reviewer initially sees the focal sentence, the highlighted token or arc, the applicable guideline section, and only the linguistic context designated as load-bearing for that task. A clearly bounded analysis bay states that the prior annotation is intentionally withheld until the first-pass action. The reviewer may enter a provisional analysis, mark multiple plausible analyses, request additional context, declare the evidence insufficient, or abstain. Requests reveal the specified linguistic context but not inherited labels. After a first-pass action is committed, the bay reintroduces the existing annotation, model suggestion, confidence, provenance, and earlier rationale. Agreement can then be confirmed with an evidence note; disagreement or unresolved uncertainty enters adjudication. Source data and prior labels remain unchanged until separately authorized adjudication is complete.","structural_mapping":[{"archetype_element":"Attention Competition Map","domain_realization":"Identify inherited labels, confidence scores, model suggestions, reviewer identities, and rationales that can compete with direct examination of the sentence and applicable annotation rule."},{"archetype_element":"Omission Candidate","domain_realization":"Temporarily withhold prior analytical answers and authority cues during the first-pass judgment while retaining them intact for later comparison."},{"archetype_element":"Protected Empty Space","domain_realization":"Reserve an explicitly blank analysis bay that prior labels, defaults, and suggestions cannot occupy before a reviewer action."},{"archetype_element":"Positive Form Relationship","domain_realization":"Use the blank bay to foreground the focal linguistic evidence and the reviewer’s own analysis rather than producing an unexplained empty panel."},{"archetype_element":"Absence Boundary","domain_realization":"Define the blank stage as beginning when an item opens and ending upon a provisional judgment, context-insufficiency decision, or abstention."},{"archetype_element":"Meaning-of-Absence Check","domain_realization":"Label the empty bay as deliberate withholding for independent review so it is not mistaken for missing annotation data or a loading failure."},{"archetype_element":"Context Preservation Frame","domain_realization":"Keep the sentence, task scope, relevant guideline, uncertainty affordances, and requested linguistic context available while only prior answers are withheld."},{"archetype_element":"Reintroduction Trigger","domain_realization":"Reveal all inherited analyses and provenance immediately after the reviewer commits an allowed first-pass action."},{"archetype_element":"Accessibility and Recoverability Guardrail","domain_realization":"Provide keyboard and screen-reader access, permit context requests and abstention, and ensure every hidden item returns without loss after the boundary."},{"archetype_element":"Clarity or Effect Test","domain_realization":"Test whether reviewers understand what is withheld, can obtain necessary context, and produce independently attributable judgments without unacceptable accuracy or workload costs."}],"mechanism_mapping":[{"mechanism_slug":"focus_mode_or_control_hiding","role":"Temporarily removes prior-answer controls and suggestions that compete with examination of the focal sentence, while guaranteeing their return after a defined first-pass action.","counterfactual_removal":"If inherited labels remain visible, the workflow cannot create an answer-free judgment interval; if they never return, reviewers lose evidence needed for comparison and repair."},{"mechanism_slug":"editorial_cut","role":"Separates load-bearing linguistic context from premature analytical material and keeps every withheld label, note, and provenance record recoverable.","counterfactual_removal":"Without this selective cut, the first-pass surface remains framed by inherited answers; indiscriminate removal would instead deprive reviewers of necessary context."},{"mechanism_slug":"empty_state_design","role":"Explains why the analysis bay is empty, distinguishes intentional withholding from absent or failed data, and presents valid exits such as analyze, request context, mark uncertainty, or abstain.","counterfactual_removal":"Without a typed empty state, reviewers may interpret the blank bay as a software fault or assume that a prior annotation does not exist."},{"mechanism_slug":"sparse_layout","role":"Limits the first-pass surface to the focal evidence, task-specific guidance, and response actions so the reviewer’s attention is not divided among prior analyses and administrative metadata.","counterfactual_removal":"Without priority-based thinning, authority cues and secondary metadata may continue to frame the linguistic decision even if the main label field is hidden."}],"causal_chain":["Visible inherited labels and automated suggestions occupy the reviewer’s judgment space before the linguistic evidence has been independently evaluated.","The blank-first stage withholds those answer-like cues while preserving the focal sentence, governing rule, and uncertainty options.","The protected empty analysis bay makes the absence deliberate, bounded, and distinguishable from missing data.","The reviewer must produce a provisional analysis, request context, declare insufficiency, or abstain before exposure to prior answers.","The system then restores inherited analyses and provenance for explicit comparison.","Disagreements and uncertainties become visible inputs to adjudication instead of being silently absorbed into apparent agreement.","If the intervention works, the audit can distinguish independently generated judgments from post-reveal confirmations while retaining full recoverability and source integrity."],"baseline":"Use the same targeted items, sentences, guideline sections, annotators, and adjudication rules in the ordinary reveal-first interface. Existing labels, model suggestions, confidence, and available rationales are visible when each item opens, and the reviewer may affirm or revise them directly.","nearest_rivals":["A written instruction asking reviewers to ignore the visible inherited label before deciding","Independent reannotation of complete sentences from scratch, which removes prior answers but changes task scope and workload","Hiding only model confidence while leaving the inherited label visible","Seeding known errors into a reveal-first audit to measure reviewer vigilance without protecting an answer-free judgment stage"],"remaining_contrastive_claim":"The proposal’s distinguishing feature is a protected and explicitly meaningful absence of inherited answers before an auditable first-pass action, followed by guaranteed reintroduction for comparison. It is not merely a cleaner interface or independent double annotation. If a reveal-first interface plus an instruction to judge independently yields equally attributable and reliable decisions without the blank stage, the intervention’s contrastive rationale fails.","authority_safety":{"decision_authority":"The corpus lead may authorize a reversible pilot and the adjudication lead may approve label changes. Individual reviewers control whether to analyze, request context, express uncertainty, or abstain. No blank-first response alone has authority to modify the released corpus.","authorized_first_step":"Prototype the staged interface on a read-only set of previously adjudicated items whose source sentences, prior labels, provenance, and reference decisions remain unchanged.","excluded_actions":["Hiding source text, task scope, applicable guidelines, known data corruption, or necessary linguistic context","Withholding consent, access, confidentiality, or provenance constraints","Forcing a categorical label when evidence is insufficient or multiple analyses are plausible","Treating disagreement with a prior label as proof that either analysis is correct","Writing provisional judgments directly into the production corpus","Using reviewer identity or status cues during the blank stage to imply a preferred answer"],"halt_rollback":"Halt if reviewers cannot distinguish intentional withholding from missing data, cannot obtain load-bearing context, are pressured into unsupported labels, or the staged interface exposes restricted information. Roll back by disabling blank-first mode; retain all source data, inherited annotations, provisional responses, and audit logs without promoting pilot decisions into the corpus."},"negative_tests":{"strongest_counterevidence":"In a counterbalanced comparison, reveal-first reviewers match the independently adjudicated reference and identify disputed cases at least as reliably as blank-first reviewers, while blank-first reviewers require more context repair, make more unsupported analyses, or experience the empty bay as an interface defect.","problem_falsifier":"The problem is falsified for the target workflow if reviewers already record a time-stamped independent judgment before seeing inherited labels or suggestions, or if the task legitimately requires the prior analysis as part of the evidence being evaluated.","intervention_falsifier":"The intervention is falsified if it does not improve attribution of independent judgments, fails to surface meaningful disagreement, reduces agreement with independently adjudicated references, or is matched by a simple instruction in the reveal-first interface.","risks":["Removing a prior parse may increase avoidable cognitive work on complex sentences.","The selected context frame may omit evidence required for a valid analysis.","Reviewers may enter a superficial answer merely to reveal the inherited annotation.","Prior exposure to the corpus or repeated constructions may compromise practical blinding.","More visible disagreement may reflect underspecified guidelines rather than better judgment.","The staged interface may disadvantage reviewers using assistive technology if reveal order and focus are not implemented accessibly.","Previously adjudicated reference labels may themselves contain unresolved analytical choices."]},"next_evidence_step":"Select 48 read-only, previously adjudicated target decisions spanning clear, ambiguous, and context-dependent cases, without changing their labels. Divide them into matched blocks and counterbalance blank-first and reveal-first conditions across 4 eligible reviewers so no reviewer sees the same item twice. Record first-pass decisions, context requests, abstentions, time to action, post-reveal revisions, evidence notes, and interface misunderstandings. Compare each condition with the existing adjudicated decision while separately examining newly surfaced disagreements rather than assuming the reference is infallible. Ask reviewers whether the withholding boundary and recovery path were clear. Use the bounded result only to reject the intervention, revise its context frame, or justify prospective evaluation.","prior_art_status":"UNSEARCHED","diversity_from_prior_proposals":"Proposal 1 protects a visible temporal gap inside conversation transcripts so analysts can perceive response timing. Proposal 2 protects a live conversational interval so a language consultant can respond before a fieldworker adds linguistic material. Proposal 3 preserves typed unsigned locations in a site-slot atlas so researchers can distinguish sign absence from missing observation and perceive spatial sign allocation. This proposal instead protects an epistemic judgment space during morphosyntactic quality review by withholding existing analytical answers until a reviewer commits an independent first-pass action. Its causal path is prior-answer removal to independently attributable judgment to post-reveal adjudication, rather than temporal-gap perception, elicitation-turn management, or spatial absence recording. It is independently adoptable as an annotation-audit protocol and requires none of the transcript, fieldwork, or signscape interventions in proposals 1 through 3.","revision_record":{"parent_version":null,"progress_targets_addressed":[],"conceptual_changes":[],"operational_changes":[],"evidence_changes":[],"claim_changes":[]}}