{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp03_full320_20260801","cell_id":"computability_boundary_mapping__tech_ethics_ai_governance","trajectory_id":"R","attempt_index":0,"archetype_slug":"computability_boundary_mapping","domain_slug":"tech_ethics_ai_governance","decision":"CANDIDATE","problem_id":"universal_autonomous_agent_policy_compliance_certification","causal_lever_id":"enforceable_verifier_scope_and_abstention_contract","proposal":{"problem":"An AI-governance program promises a terminating, exact pre-deployment verdict on whether any submitted autonomous-agent program will ever violate a machine-readable policy over unbounded interactions. Timeouts, unsupported inputs, and inconclusive searches are forced into compliant/noncompliant labels, creating false assurance or unjustified rejection. The problem concerns conformance to the declared formal policy, not universal moral acceptability.","actors_substrate":["AI governance board and assurance team","agent developers and deployers","formal-methods reviewers","operators consuming certification verdicts","people subject to agent decisions or actions","arbitrary agent programs, policy specifications, and interaction models"],"observable_state":"A universal Boolean certification claim lacks a fixed program language, environment model, interaction bound, and proof of totality; analyzer timeouts or search exhaustion are nevertheless emitted as substantive compliance verdicts.","consequence":"Deployments may receive authoritative-looking clearance without a justified guarantee, while repeated engineering investment pursues an unrestricted exact verifier that may be impossible.","affected_objective":"Truthful, auditable, and operationally safe governance of autonomous-agent policy compliance.","structural_mapping":[{"archetype_element":"Universal exact terminating solver claim","domain_realization":"The certification service claims to decide policy-violation reachability for arbitrary agents and unbounded environments.","claim_kind":"HYPOTHESIS"},{"archetype_element":"Explicit class, representation, model, and quantifiers","domain_realization":"The governed object becomes a tuple of agent language, policy semantics, environment model, interaction horizon, and universal or bounded guarantee.","claim_kind":"INFERENCE"},{"archetype_element":"Constructive and impossibility evidence in parallel","domain_realization":"Reviewers seek a total verified checker for restricted classes while testing whether unrestricted reachability admits a valid halting-problem reduction.","claim_kind":"INFERENCE"},{"archetype_element":"Decidable regions and weaker fallbacks","domain_realization":"Enforceable finite-state fragments receive exact checking; broader inputs receive sound abstraction, bounded analysis, or explicit UNKNOWN.","claim_kind":"INFERENCE"},{"archetype_element":"Status-preserving interface","domain_realization":"COMPLIANT-WITHIN-MODEL, VIOLATION-WITNESS, UNKNOWN, OUT-OF-SCOPE, and SYSTEM-FAILURE remain distinct.","claim_kind":"HYPOTHESIS"},{"archetype_element":"Versioned boundary and reclassification","domain_realization":"Every verdict links to the policy, model, scope, proof, and triggers that invalidate it.","claim_kind":"INFERENCE"}],"component_map":[{"component":"Problem-Class Specification","status":"adapted","domain_realization":"Define formal policy-violation reachability for declared agent and environment classes."},{"component":"Instance Representation Contract","status":"adapted","domain_realization":"Version the agent language, policy syntax, environment encoding, and horizon."},{"component":"Computation Model Contract","status":"direct","domain_realization":"State machine model, allowed oracle or human inputs, and resource semantics."},{"component":"Solvability Guarantee Profile","status":"direct","domain_realization":"Distinguish exact decision, sound one-sided analysis, bounded completeness, and approximation."},{"component":"Quantifier and Scope Map","status":"direct","domain_realization":"Separate arbitrary agents and unbounded traces from restricted or bounded claims."},{"component":"Computability Status Lattice","status":"direct","domain_realization":"Classify each scope as decidable, recognizable, partial, relative, undecidable, or unresolved."},{"component":"Constructive Procedure Witness","status":"adapted","domain_realization":"Provide a checker with correctness and termination proof for each exact fragment."},{"component":"Reduction Preservation Contract","status":"direct","domain_realization":"Document the computable source-to-certification translation and answer preservation."},{"component":"Computability Impossibility Certificate","status":"direct","domain_realization":"Retain a reviewed reduction if unrestricted certification is proved undecidable."},{"component":"Assumption Register","status":"direct","domain_realization":"Record formalization, policy semantics, environmental closure, and oracle assumptions."},{"component":"Decidable Subclass Map","status":"adapted","domain_realization":"Map enforceable finite-state, bounded-horizon, or restricted-policy fragments."},{"component":"One-Sided Recognition Contract","status":"adapted","domain_realization":"Permit confirmed violation witnesses without interpreting failure to find one as compliance."},{"component":"Unknown and Nontermination Policy","status":"direct","domain_realization":"Expose UNKNOWN separately from compliant, noncompliant, timeout, and failure."},{"component":"Fallback Solution Contract","status":"adapted","domain_realization":"Label exact, abstract, bounded, and escalated modes with their actual guarantees."},{"component":"Computability Guarantee Record","status":"adapted","domain_realization":"Attach the formal scope and evidence to each certification mode."},{"component":"Recheck Trigger","status":"adapted","domain_realization":"Reassess when language expressiveness, policy semantics, model, bound, or external capability changes."},{"component":"Termination Condition","status":"direct","domain_realization":"Specify proof-based termination or a resource bound ending in UNKNOWN."},{"component":"Scope Boundary","status":"adapted","domain_realization":"Mechanically reject or quarantine inputs outside an enforceable fragment."},{"component":"Decision Record","status":"direct","domain_realization":"Version the chosen boundary, shipped guarantee, rationale, and expiry conditions."},{"component":"Uncertainty Residue","status":"adapted","domain_realization":"List unproved obligations, abstraction gaps, and model-to-world mismatch."},{"component":"Independent Proof Review","status":"direct","domain_realization":"Have an independent reviewer check reductions and constructive proofs."},{"component":"Complexity Follow-On Gate","status":"direct","domain_realization":"Assess feasibility only after a class is placed on the decidable side."}],"mechanism_dispositions":[{"slug":"abstract_interpretation_or_model_checking","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Use a validated over-approximation or finite model as a sound fallback, with false alarms and model limits exposed.","counterfactual_removal":"Broader inputs lose a useful guarantee-preserving fallback and are reduced to rejection or unsupported Boolean claims."},{"slug":"bounded_domain_exhaustive_search","disposition":"selected_supporting","contribution_type":"TEST_DESIGN","adaptation_or_rejection":"Exhaust finite trace and state bounds and prohibit generalization beyond them.","counterfactual_removal":"The pilot loses a complete bounded benchmark but the boundary intervention remains viable."},{"slug":"computability_boundary_decision_record","disposition":"selected_load_bearing","contribution_type":"OPERATIONAL","adaptation_or_rejection":"Version the boundary, shipped labels, assumptions, evidence, and recheck triggers.","counterfactual_removal":"Guarantees can drift silently as policies and agent languages change."},{"slug":"computational_complexity_analysis","disposition":"selected_supporting","contribution_type":"OPERATIONAL","adaptation_or_rejection":"Gate decidable fragments through time and memory feasibility analysis.","counterfactual_removal":"Correct fragments may still be operationally unusable, though their computability classification survives."},{"slug":"constructive_algorithm_and_correctness_proof","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Require totality and correctness proofs before issuing exact verdicts for a fragment.","counterfactual_removal":"Exact certification would rest on tests rather than a class-wide witness."},{"slug":"diagonalization_impossibility_proof","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Reject as redundant unless a direct self-reference construction is clearer than the selected reduction.","counterfactual_removal":"No change; the halting reduction supplies the proposed impossibility test."},{"slug":"enumeration_and_dovetailing","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Reject for the first operational design because indefinite execution conflicts with the certification workflow.","counterfactual_removal":"No change; bounded semi-decision preserves one-sided honesty while terminating."},{"slug":"fallback_mode_router","disposition":"selected_load_bearing","contribution_type":"OPERATIONAL","adaptation_or_rejection":"Route by enforceable scope to exact, sound-abstract, bounded, or escalation modes and label each result.","counterfactual_removal":"Weaker guarantees are likely to be flattened into one misleading certification label."},{"slug":"halting_problem_reduction","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Attempt a computable, property-preserving reduction to unrestricted policy-violation reachability; classify as unresolved until reviewed.","counterfactual_removal":"There is no principled basis for stopping pursuit of the unrestricted total verifier."},{"slug":"language_fragment_restriction","disposition":"selected_load_bearing","contribution_type":"CORE_CAUSAL","adaptation_or_rejection":"Define a mechanically enforceable agent and policy fragment admitting a total checker.","counterfactual_removal":"The exact mode lacks a safely enforceable decidable region."},{"slug":"many_one_reduction_proof","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Its obligations are incorporated into the specific halting reduction rather than maintained as a separate mechanism.","counterfactual_removal":"No material change if the specific reduction still proves totality, computability, and answer preservation."},{"slug":"promise_problem_restriction","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Reject an unenforced don't-care region; governance outputs require explicit rejection or UNKNOWN outside scope.","counterfactual_removal":"Safety improves because off-promise inputs cannot receive arbitrary authoritative answers."},{"slug":"proof_by_counterexample","disposition":"selected_supporting","contribution_type":"TEST_DESIGN","adaptation_or_rejection":"Use well-formed adversarial agents to refute over-broad analyzer claims, without treating failed search as proof.","counterfactual_removal":"The pilot loses a cheap falsification route but retains formal boundary evidence."},{"slug":"proof_checking","disposition":"selected_load_bearing","contribution_type":"SAFETY_GUARDRAIL","adaptation_or_rejection":"Independently check the exact theorem, proof steps, and premises supporting each boundary.","counterfactual_removal":"A subtle proof or formalization error could acquire governance authority."},{"slug":"reduction_direction_checklist","disposition":"selected_supporting","contribution_type":"SAFETY_GUARDRAIL","adaptation_or_rejection":"Gate the impossibility claim on source-to-target direction and explicit assumptions.","counterfactual_removal":"Independent review remains possible but a common invalid transfer becomes easier to miss."},{"slug":"semi_decision_with_explicit_unknown","disposition":"selected_load_bearing","contribution_type":"SAFETY_GUARDRAIL","adaptation_or_rejection":"Return violation witnesses when found and UNKNOWN at the declared budget; never convert timeout to compliance.","counterfactual_removal":"One-sided searches can be laundered into false negative or false assurance verdicts."},{"slug":"theorem_prover_guided_search","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Reject as a required dependency for the first test; proof discovery timeout would add another ambiguous non-answer.","counterfactual_removal":"No change because supplied proofs can still undergo independent checking."},{"slug":"turing_reduction_analysis","disposition":"considered_rejected","contribution_type":"NONE","adaptation_or_rejection":"Reject because relative oracle degrees do not establish the concrete one-sided and operational contract needed here.","counterfactual_removal":"No material change to the proposed class boundary."}],"causal_chain":["Formalize the certification claim, input class, computation model, policy semantics, and quantifiers.","Seek a constructive total checker for enforceable fragments and a reviewed reduction against the unrestricted claim.","Classify each scope without collapsing undecidable, unresolved, recognizable, and costly.","Route queries only to modes whose preconditions hold.","Preserve exact verdict, witness, UNKNOWN, out-of-scope, and system-failure states at downstream interfaces.","Version evidence and assumptions so model or language changes trigger reclassification.","Reduce unsupported universal clearance and redirect effort toward bounded, sound, auditable guarantees."],"baseline":"Run a universal analyzer or test suite, impose a timeout, and emit a Boolean compliant/noncompliant certification without a model-relative proof or abstention state.","nearest_rival":"Red-teaming plus runtime monitoring, which can discover violations and reduce harm but cannot establish that an unrestricted exact terminating certification procedure exists or does not exist.","authority_safety":{"affected_parties":["people evaluated, assisted, or acted upon by certified agents","agent operators and developers","reviewers accountable for certification","organizations relying on verdicts"],"decision_authority":"The designated AI governance owner may authorize an offline pilot; changing production certification labels requires joint approval from governance, formal assurance, and the accountable deployment owner.","authorized_first_step":"Offline, formalize one existing policy and agent language; test 30 bounded or finite-state fixtures including seeded counterexamples, compare exact, abstract, and UNKNOWN outputs, and obtain independent review before any production use.","excluded_actions":["No production deployment approval or denial from pilot outputs","No claim of moral acceptability beyond formal policy conformance","No timeout-to-compliant or timeout-to-noncompliant conversion","No unrestricted undecidability claim before the reduction and formalization are independently checked","No silent expansion of the accepted language or environment model"],"halt_rollback":"Halt if fragment membership cannot be enforced, a supposedly sound mode misses a seeded violation, output labels are collapsed downstream, or independent review finds a proof or formalization gap; withdraw affected guarantees and revert to explicit UNKNOWN or manual review."}},"negative_tests":{"strongest_counterevidence":"The actual certification language and environment may already be finite, bounded, effectively enumerable, and covered by a verified total checker; then the issue is complexity or implementation quality, not a computability boundary.","analogy_break":"Formal reachability is not ethical legitimacy: contested values, incomplete policies, affected-party rights, and real-world model error cannot be resolved by a computability classification. A proved checker certifies only the encoded property under its model.","failure_condition":"The proposal fails if scope restrictions cannot be mechanically enforced or if downstream governance treats UNKNOWN, out-of-scope, or abstraction alarms as definitive compliance judgments.","problem_falsifier":"The problem is falsified if records show that only fixed bounded instances are claimed, every accepted input is within an enforced finite model, a reviewed total-correct checker exists, and no universal or unbounded guarantee is communicated.","intervention_falsifier":"The intervention is falsified if, on the bounded pilot, boundary records and routed labels do not prevent seeded timeout, out-of-scope, and abstraction cases from receiving unsupported Boolean verdicts, or reviewers cannot reproduce the claimed classification from recorded assumptions and evidence.","risks":["A correct proof may certify the wrong formalization of policy or environment.","A restrictive fragment may exclude the cases creating the greatest social risk.","False alarms from coarse abstraction may cause alert fatigue or unjustified deployment denial.","Official boundary records may institutionalize an erroneous result.","UNKNOWN may be strategically interpreted as safe unless downstream action rules are enforced.","Complexity may render a decidable exact mode unusable at realistic scale."]},"null_rationale":null,"classification":{"candidate_kind":"DOMAIN_TRANSFER","prior_art_status":"UNSEARCHED","evidence_maturity":"HYPOTHESIS"},"revision_change_log":{"revision_kind":"ORIGINAL","prior_problem_id":null,"prior_causal_lever_id":null,"problem_changed":false,"causal_lever_changed":false,"conceptual_changes":[],"operational_changes":[],"repairs_addressed":[]},"confidence":0.9,"generator_notes":"Candidate fit rests on a formal, universal policy-conformance claim for arbitrary agent programs, not on deciding ethics itself. Empirical prevalence and pilot effects remain untested."}