{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp04_retrieval_first_paired20_20260802","cell_id":"computability_boundary_mapping__public_administration_policy","arm":"RETRIEVAL_FIRST","candidate_id":"C-H5-V0","hypothesis_id":"H5","version":0,"title":"Computability gate for algorithm procurement","problem":"Public buyers may accept vendor claims of universal, exact, terminating detection across open-ended fraud or misconduct patterns when vendor assertions, benchmarks, or mismatched impossibility analogies substitute for a model-relative solvability argument. The contractual guarantee can then persist or drift after award without reclassification.","actors":["Procurement officers and contracting authorities","Program officials requesting the algorithmic capability","Vendor technical and contracting teams","Independent computability reviewer","Legal, audit, and oversight bodies","Case subjects affected by system outputs","Taxpayers funding procurement and remediation"],"observable_state":"For one procured algorithmic capability, the solicitation or proposed contract promises correct, exact, terminating detection for every member of a declared or implicitly open-ended input class, while the procurement file lacks a computability-specific record defining the encoding, computation model, quantifiers, termination requirement, supporting proof, fallback states, and recheck triggers.","consequence":"An unsupported universal guarantee may be awarded, weaker benchmark or bounded evidence may be mistaken for class-wide proof, and later scope or functionality changes may silently extend the guarantee, exposing case subjects and public funds to misleading automation claims and repeated spending on an unattainable specification.","affected_objective":"Award only algorithmic capabilities whose contractual guarantees are supported for their declared input class and computation model, while keeping procurement delay within 10% of the conventional cycle.","intervention":"Insert a mandatory, independently reviewed computability-specific subcase into the existing pre-award assurance record for any exact terminating algorithmic guarantee. The record must define the problem class, enforceable input encoding, computation model and external capabilities, universal quantifiers, output states, and required termination guarantee; classify the claim as decidable, recognizable, partial, relative, undecidable, or unresolved; and attach model-matched evidence. An exact terminating guarantee may proceed only with a constructive procedure plus correctness and totality justification for the full declared class. A counterexample may refute an overbroad vendor claim but cannot establish general undecidability; an impossibility conclusion requires a scope-matched proof or reduction. Otherwise the buyer must narrow the class, weaken the guarantee, expose UNKNOWN or OUT_OF_SCOPE, or decline the capability. The versioned record becomes part of technical evaluation and the contract, and specified material changes reopen review.","structural_mapping":[{"archetype_element":"Problem-Class Specification","domain_realization":"The solicitation defines one algorithmic capability as a class of encoded fraud or misconduct cases rather than relying on examples or an unrestricted natural-language promise."},{"archetype_element":"Computation Model Contract","domain_realization":"The procurement record names the permitted algorithm, data, expert assistance, external services, interaction, and resource assumptions behind the vendor guarantee."},{"archetype_element":"Quantifier and Scope Map","domain_realization":"Reviewers identify whether terms such as every, any, never, and always range over a finite enforceable class or an open-ended class."},{"archetype_element":"Solvability Guarantee Profile","domain_realization":"The proposed contractual output is classified without collapsing exact decision, one-sided recognition, bounded analysis, approximation, UNKNOWN, and system failure."},{"archetype_element":"Constructive or Impossibility Evidence","domain_realization":"A total-exact claim requires a class-wide algorithm, correctness argument, and termination argument; an impossibility claim requires assumptions and proof direction matched to the procured target."},{"archetype_element":"Fallback Solution Contract","domain_realization":"Where total exactness is unsupported, the solicitation specifies a narrower fragment, bounded result, explicit UNKNOWN or OUT_OF_SCOPE response, or accountable human escalation without inheriting the original guarantee."},{"archetype_element":"Decision Record and Recheck Trigger","domain_realization":"The reviewed boundary, evidence, shipped guarantee, fallback, and material-change triggers are versioned, incorporated into the contract, and reopened when scope, language, model, data access, or external capabilities materially change."},{"archetype_element":"Independent Proof Review","domain_realization":"A reviewer separate from the vendor checks that the formal statement matches the procurement claim, the evidence covers the asserted class, and any reduction runs in the justified direction."}],"mechanism_mapping":[{"mechanism_slug":"computability_boundary_decision_record","role":"Creates the versioned contractual ledger linking the declared class, classification, evidence, approved guarantee, fallback, assumptions, and recheck triggers.","counterfactual_removal":"Without the record, a sound review can remain informal, cannot reliably constrain award language, and can drift after material changes."},{"mechanism_slug":"reduction_direction_checklist","role":"Screens any hardness or impossibility argument for an established source, source-to-target direction, total computable mapping, answer preservation, matching encoding, and explicit assumptions.","counterfactual_removal":"Without the checklist, a reversed or assumption-mismatched reduction may be treated as authority for narrowing or rejecting a procurement."},{"mechanism_slug":"proof_by_counterexample","role":"Tests universal vendor claims by exhibiting a well-formed in-scope case on which the claimed method fails, forcing the claim to be withdrawn or narrowed.","counterfactual_removal":"Without counterexample testing, reviewers lose an economical way to refute overbroad universals, although absence of a counterexample still would not prove the claim."}],"causal_chain":["A solicitation proposes an exact, correct, terminating guarantee for a declared algorithmic capability.","The gate makes the input class, encoding, computation model, quantifiers, and output semantics explicit before award.","The vendor must connect the universal guarantee to class-wide totality and correctness evidence rather than demonstrations or benchmarks alone.","An independent reviewer checks formal-to-operational fidelity and screens counterexamples, assumptions, and reduction direction.","Unsupported claims are marked unresolved and therefore cannot be awarded as exact terminating guarantees; the buyer must narrow, weaken, add explicit fallback states, or decline them.","The approved guarantee and its boundary are incorporated into the contract with versioned material-change triggers.","Consequently, fewer unsupported universal guarantees reach award, while a standardized subcase is intended to limit added review time."],"baseline":"Conventional public software or AI procurement using general assurance cases, supplier claims, performance evidence, independent testing where selected, contractual acceptance conditions, and later monitoring, but without a mandatory computability-specific classification or proof obligation for universal exact terminating claims.","nearest_rivals":["Government RFP software-assurance-case submission described by the Carnegie Mellon University Software Engineering Institute","Object Management Group Structured Assurance Case Metamodel","UK National Cyber Security Centre supplier-claims procurement assurance","Government of Canada automated-decision governance and material-change trigger"],"remaining_contrastive_claim":"Within public algorithm procurement, the still-distinct element is a mandatory computability-specific subcase that formalizes the input class and universal quantifiers and requires model-matched evidence of total decidability—or an appropriately scoped impossibility result—before an exact terminating guarantee can be awarded or renewed.","authority_safety":{"decision_authority":"The contracting authority retains the award decision, with the responsible procurement officer accepting the record only after independent technical review and applicable legal or oversight review; the reviewer may classify evidence and recommend narrowing but may not award, cancel, or adjudicate cases.","authorized_first_step":"Conduct a document-only retrospective coding exercise and reviewer tabletop on a bounded sample of completed solicitations; record hypothetical gate outcomes without changing contracts, vendor status, benefits, enforcement, investigations, or individual cases.","excluded_actions":["No automated eligibility, enforcement, fraud, misconduct, disciplinary, or adverse-case decision","No finding that a vendor or system is deceptive merely because a proof is absent","No inference of false, impossible, or undecidable from timeout, failed search, missing implementation, or poor scaling","No use of a counterexample to claim general undecidability","No extension of a theorem or reduction beyond its stated encoding, model, promise, or quantifiers","No disclosure of protected procurement, vendor, employee, or case-subject information","No contract award, rejection, cancellation, renewal denial, or public vendor rating during the first evidence step"],"halt_rollback":"Stop the exercise if reviewers cannot define the claimed class without inventing requirements, protected information cannot be separated, reviewer agreement is too low for interpretable coding, or the exercise begins influencing live procurement or individual decisions. Quarantine the affected records, discard hypothetical determinations from operational files, and retain only de-identified methodological findings."},"negative_tests":{"strongest_counterevidence":"The independent search found substantial collision: assurance cases, structured claim-evidence-assumption records, qualified or independent review, contractual acceptance conditions, continuing verification, and material-change reassessment already exist. It found no direct implementation of the narrow computability-specific combination, but also no evidence that public buyers repeatedly accept universal, exact, terminating detection claims as a measured procurement failure.","problem_falsifier":"In a preregistered sample of relevant solicitations and award records, no in-scope universal exact terminating claims are found, or every such claim already has model-matched class-wide evidence and an enforceable change trigger under conventional review.","intervention_falsifier":"In a prospective matched evaluation, the gate does not reduce independently coded unsupported universal guarantees at award, reviewer classifications are not reproducible, or median procurement cycle time increases by more than 10% relative to matched conventional solicitations.","risks":["The gate may duplicate established assurance work and add delay without addressing a prevalent problem.","Procurement staff may mistake missing proof for proof of impossibility.","A formally valid record may describe the wrong operational task and confer false authority.","Vendors may narrow the declared class strategically while marketing language remains broad.","Independent review may be unavailable, inconsistent, or captured by vendor-provided framing.","Trade-secrecy or documentation limits may prevent meaningful scrutiny.","An official record may become stale and launder guarantee drift if recheck triggers are not enforced.","Overly strict proof demands may exclude useful bounded, approximate, or human-assisted systems that are honestly labeled."]},"next_evidence_step":"Preregister definitions and sample 20 completed public solicitations or award records involving fraud or misconduct detection, selected without regard to outcome. Two independent reviewers will code each record for universal quantifiers, open-ended versus enforceably bounded input class, exactness, termination, evidence type, fallback states, and change triggers; disagreements will be adjudicated without treating silence as failure. Then apply a standardized draft gate to the records containing candidate universal claims and measure reviewer agreement, the number of claims that would be upheld, narrowed, relabeled UNKNOWN or OUT_OF_SCOPE, or left unresolved, and estimated review hours. Proceed to a live matched pilot only if at least one unsupported in-scope claim is observed, reviewers can apply the classification reproducibly, and estimated added cycle time is no more than 10%.","prior_art_status":"SEARCHED_BOUNDED","revision_record":{"parent_version":null,"progress_targets_addressed":[],"conceptual_changes":[],"operational_changes":[],"evidence_changes":[],"claim_changes":[]}}