{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp09_archetype_breadth150_20260804","cell_id":"additive_measure_space_design__computer_science","arm":"BREADTH_PROBE_ONE_SHOT","candidate_id":"additive_measure_space_design__computer_science__P1","proposal_index":1,"version":0,"title":"Partition-Invariant Change Coverage for CI Test Selection","problem":"A continuous-integration test selector estimates how much change-relevant code a candidate suite covers by summing each test's reported covered-unit count. Because tests frequently cover the same executable units, the sum treats overlapping coverage as disjoint mass. The apparent coverage therefore depends on how tests are grouped or ordered and can reward redundant suites over suites covering more distinct change-relevant code.","actors":["Developers submitting code changes","CI and test-infrastructure engineers","The automated test-selection service","Release engineers reviewing build evidence"],"observable_state":"For a fixed build and test set, the reported aggregate change coverage can exceed the number of declared executable units, and splitting, merging, or reordering equivalent test groups changes attributed totals even though the union of executed units is unchanged. Units that instrumentation cannot identify are commonly absent from both numerator and denominator without an explicit status.","consequence":"Under a fixed CI-time budget, the selector can allocate capacity to tests with duplicative coverage while omitting tests that exercise distinct change-relevant units; release reviewers also receive a total that cannot be reconciled to the covered code universe.","affected_objective":"Select tests that exercise the largest coherent subset of declared change-impact executable units within the CI-time budget while keeping uninstrumented or unresolved units visible.","intervention":"For each build, establish a finite counting measure over a declared universe of uniquely identified executable units in the change-impact slice. Treat every exactly identifiable subset as measurable, assign each unit one nonnegative unit of mass, set the empty set to zero, and compute suite coverage as the measure of the union of its tests' coverage sets rather than the sum of their individual measures. Partition the universe into disjoint coverage-signature cells, verify that cell counts recompose to the universe count, and place generated, dynamically created, ambiguously mapped, or uninstrumented units in a boundary queue rather than silently assigning them zero mass. Bind the resulting measure contract to the downstream test-selection algorithm and report its scope with every selection.","structural_mapping":[{"archetype_element":"Measurable Universe Scope","domain_realization":"The uniquely keyed executable units in the dependency-based change-impact slice for one repository, commit, build configuration, and instrumentation version."},{"archetype_element":"Measurable Subset Family","domain_realization":"Subsets whose unit membership is exactly representable from the build manifest and coverage observations, including unions, intersections, and complements relative to the declared universe."},{"archetype_element":"Empty-Set Zero Rule","domain_realization":"A test or suite exercising no units in the declared universe receives coverage mass zero."},{"archetype_element":"Nonnegative Size Assignment Rule","domain_realization":"Each uniquely identified executable unit contributes one count; a subset's measure is its number of distinct members."},{"archetype_element":"Disjoint Additivity Rule","domain_realization":"Counts of disjoint coverage-signature cells must sum to the count of their union, while overlapping per-test sets must first be unioned or disjointized."},{"archetype_element":"Normalization or Scale Anchor","domain_realization":"The build's declared universe count is the scale anchor; normalized coverage, when used, is union count divided by that count."},{"archetype_element":"Null-Set and Negligibility Policy","domain_realization":"No covered executable unit is treated as negligible merely because it has low selection frequency; unresolved or uninstrumented units are reported separately rather than folded into the zero set."},{"archetype_element":"Partition Consistency Register","domain_realization":"A per-build table records disjoint cells defined by identical test-coverage signatures and checks that alternative regroupings yield the same union measure."},{"archetype_element":"Measurability Boundary Queue","domain_realization":"Units lacking stable identity or trustworthy instrumentation mappings are queued with reason codes for instrumentation review."},{"archetype_element":"Integration and Downstream Use Contract","domain_realization":"The selector may optimize measured union coverage only for the named build scope and must expose boundary-queue size and measure version with its recommendation."}],"mechanism_mapping":[{"mechanism_slug":"partition_sum_table","role":"Disjointizes overlapping test coverage by assigning each executable unit to exactly one coverage-signature cell, then verifies recomposition totals.","counterfactual_removal":"Without the table, overlap corrections and missing cells cannot be reconciled, so a seemingly corrected aggregate may remain partition-dependent."},{"mechanism_slug":"finite_or_countable_additivity_test","role":"Checks on the finite build universe that every tested disjoint partition has a union count equal to the sum of its cell counts.","counterfactual_removal":"Without the additivity test, identity duplication or leakage across coverage files can survive into the selector's objective."},{"mechanism_slug":"measurable_family_closure_check","role":"Verifies that union, intersection, and relative-complement operations used by selection and marginal-coverage calculations remain within exactly representable subsets.","counterfactual_removal":"Without closure checking, the selector may calculate marginal coverage for subsets whose membership cannot actually be resolved."},{"mechanism_slug":"null_set_policy_register","role":"Separates true zero coverage from unknown, generated, dynamically mapped, and uninstrumented cases.","counterfactual_removal":"Without the register, missing evidence is indistinguishable from measured absence and can disappear from release review."}],"causal_chain":["Per-test coverage sets overlap while the baseline sums their individual counts.","The same executable unit is consequently represented as multiple units of apparent suite coverage.","Candidate suites containing redundant tests can receive inflated aggregate values, and equivalent regroupings can change those values.","A declared finite universe and stable unit identity make membership and the scale anchor explicit.","Disjoint coverage-signature cells convert overlapping observations into a partition, and counting measure makes their recomposition additive.","Union measure supplies the selector with partition-invariant distinct coverage while the boundary queue preserves unresolved units.","The selector can then compare suites against the same coherent coverage universe within its existing time constraint."],"baseline":"The selector ranks or evaluates tests using per-test changed-unit counts, often divided by estimated runtime, and adds those counts across a proposed suite. Overlap handling, the exact executable-unit universe, and treatment of uninstrumented units are implicit or applied after scoring.","nearest_rivals":["Exact set-union coverage computed without a declared universe or boundary register: removes direct double-counting but leaves denominator drift and missing-unit semantics unresolved.","Greedy set-cover test selection: supplies an optimization procedure, but its decisions still depend on the validity of the underlying universe, membership sets, and size rule.","Mutation-based test prioritization: measures whether tests detect injected behavioral changes rather than establishing an additive size measure over covered executable subsets.","Coverage-instrumentation standardization: improves observation consistency but does not itself define the subset family, additive union rule, normalization anchor, or null policy."],"remaining_contrastive_claim":"The candidate's distinctive claim is limited to making change coverage a scoped, nonnegative, additive measure over executable-unit subsets before optimization. It is not a new test-ranking algorithm, a claim that line execution implies fault detection, or merely a change to coverage collection.","authority_safety":{"decision_authority":"The repository's CI or test-infrastructure owner may define and run the shadow measure; release-policy owners retain authority over merge or deployment gates.","authorized_first_step":"Run an offline shadow replay on historical build artifacts and compare measure consistency and resulting test rankings without changing live CI selection or release decisions.","excluded_actions":["Automatically blocking merges or deployments","Dropping tests solely because their measured marginal coverage is zero","Treating boundary-queue units as uncovered, covered, or irrelevant without review","Combining measures from different commits, build configurations, repositories, or instrumentation versions","Using the measure as evidence that executed code is behaviorally correct"],"halt_rollback":"Halt the pilot if unit identity is unstable, recomposition checks fail, the declared universe cannot be reproduced, or boundary cases are silently coerced to zero. Roll back by discarding shadow recommendations and retaining the existing selector and logs unchanged."},"negative_tests":{"strongest_counterevidence":"The current selector already uses a reproducible build-scoped universe, stable unique unit identities, exact union coverage, explicit unresolved-unit reporting, and partition-recomposition checks; if so, the posited structural defect is absent.","problem_falsifier":"Across equivalent test regroupings, observed suite totals equal the exact union count, never exceed the declared universe, and every omitted unit has an explicit measurable-status reason.","intervention_falsifier":"The proposed measure passes its internal consistency checks but does not alter any suite comparison affected by overlap, or discrepancies are instead explained by stale or incorrect instrumentation observations that measure-space rules cannot repair.","risks":["Stable unit identifiers may change across compilers or builds, creating false membership differences.","A count measure treats executable units as equal despite differing behavioral or safety significance.","Restricting the measurable family may exclude dynamic or generated behavior important to releases.","The boundary queue may grow without ownership and become a ceremonial warning.","Optimizing distinct execution coverage can still select tests that execute code without detecting faults.","Computing coverage-signature cells may add storage and processing cost for large test matrices."]},"next_evidence_step":"For one repository, replay 20 completed pull-request builds with frozen manifests, coverage artifacts, runtimes, and baseline selections. Construct the declared universe and boundary queue; test empty-set, nonnegativity, closure, and partition recomposition; then compare baseline and measure-based suite rankings under the recorded CI-time budget. Record only whether overlap-driven inconsistencies occur, whether the measure resolves them, which selections change, and which units remain nonmeasurable; do not alter production CI.","prior_art_status":"UNSEARCHED","diversity_from_prior_proposals":"No comparison with other proposals was performed under runtime isolation; this candidate is independently situated in build-scoped CI test selection and executable-unit coverage accounting.","revision_record":{"parent_version":null,"progress_targets_addressed":[],"conceptual_changes":[],"operational_changes":[],"evidence_changes":[],"claim_changes":[]}}