{"schema_version":1,"experiment_id":"eoa_inverse_innovation_exp09_archetype_breadth150_20260804","cell_id":"agentic_control_loop_design__chemistry_materials","arm":"BREADTH_PROBE_ONE_SHOT","candidate_id":"agentic_control_loop_design__chemistry_materials__P1","proposal_index":1,"version":0,"title":"Bounded Operator Rescue Loop for Drifting Polymerization Batches","problem":"A reactor operator is held responsible for a polymer batch meeting its viscosity and conversion targets, yet the operator receives only fragmented process signals, cannot see the assumptions behind the recipe, and must seek approval for every corrective adjustment. By the time a process engineer interprets the deviation and authorizes a response, the batch may have passed the useful correction window. The operator therefore follows the nominal recipe or escalates ritualistically despite observing early signs of drift.","actors":["Reactor operator responsible for executing the batch","Process engineer responsible for the process model and validated operating envelope","Shift supervisor responsible for escalation and production coordination","Process-safety representative responsible for hard safety constraints","Quality laboratory staff providing in-process and final material measurements"],"observable_state":"At defined batch checkpoints, the operator can observe reactor temperature, pressure, agitator torque, actual versus planned feed delivery, elapsed stage time, and available in-process conversion or viscosity proxies. The relevant state is not any single reading but the signed deviation between these observations and stage-specific expected trajectories, together with sensor confidence and the remaining time in which an authorized correction could plausibly affect the batch.","consequence":"When emerging drift is observable but disconnected from an actionable model, bounded correction rights, and timely effect feedback, the batch can finish outside its specified viscosity, conversion, or molecular-weight proxy range, creating rework, disposal, schedule disruption, or escalation toward an operating limit.","affected_objective":"Complete each polymerization batch within its approved material-property specification while remaining inside validated thermal, pressure, feed, and equipment constraints.","intervention":"Create a checkpoint-based operator agency loop for deviations that remain inside a prevalidated correction envelope. At each checkpoint, a control card states the material-property goal, displays observed-versus-expected trajectories, exposes the current model assumptions, and offers a Safe Action Menu of already reviewed adjustments such as holding the next stage for a bounded interval, choosing among validated feed-rate bands, or requesting an additional approved sample. A Decision Rights Matrix specifies which choices the operator may execute, which require process-engineer confirmation, and which require immediate safe-state escalation. After each permitted action, the console presents the next relevant observations and records whether the predicted directional response occurred; the operator and process engineer then update the assumption register before the next batch. The intervention changes responsibility only where information, authority, execution capability, and feedback are actually present.","structural_mapping":[{"archetype_element":"Represented Goal","domain_realization":"A visible checkpoint goal combining the required final polymer property range with non-negotiable reactor safety limits."},{"archetype_element":"World Model","domain_realization":"A stage-specific account of how temperature, feed delivery, time, torque, and available analytical proxies are expected to relate to conversion and viscosity drift."},{"archetype_element":"Observation Channel","domain_realization":"Time-aligned reactor signals, sampling results, sensor-health indicators, and deviation from the expected batch trajectory."},{"archetype_element":"Action Repertoire","domain_realization":"A finite menu of prevalidated holds, feed-band selections, approved additional sampling, escalation, and safe-state responses."},{"archetype_element":"Selection Policy","domain_realization":"Choose an action only when the deviation pattern, sensor confidence, correction window, and predicted directional response meet the rule attached to that action."},{"archetype_element":"Legitimate Action Boundary","domain_realization":"Operator discretion exists only within the validated recipe envelope; changes to chemistry, total charge, hard limits, or unreviewed sequences remain outside it."},{"archetype_element":"Execution Capability","domain_realization":"The batch-control interface permits authorized adjustments without waiting for a separate routine approval while preserving interlocks and audit logging."},{"archetype_element":"Effect Feedback Loop","domain_realization":"The next checkpoint compares the action's predicted directional effect with temperature, torque, feed, and analytical responses."},{"archetype_element":"Model Update Rule","domain_realization":"A prediction mismatch creates a flagged assumption for process-engineer review and can narrow or suspend the corresponding menu option for later batches."},{"archetype_element":"Proportional Accountability Frame","domain_realization":"The operator is accountable for eligible choices and required escalation, not for raw-material variation, unavailable measurements, blocked controls, or outcomes outside the mapped correction window."},{"archetype_element":"Override and Escalation Path","domain_realization":"Low-confidence sensing, conflicting indicators, approach to a hard limit, or an unavailable safe action transfers the decision to the supervisor and process-safety pathway."},{"archetype_element":"Agency Health Signal","domain_realization":"Each deviation records whether a goal, trusted observation, feasible action, authority, execution path, and timely feedback were simultaneously available."}],"mechanism_mapping":[{"mechanism_slug":"agency_loop_map","role":"Map one polymer batch from property target through observations, inferred state, available corrections, decision authority, execution, and measured response to expose missing couplings.","counterfactual_removal":"Without the map, the organization may add alarms or assign ownership while leaving the operator unable to connect a deviation to a legitimate action and its effect."},{"mechanism_slug":"safe_action_menu","role":"Convert a validated operating envelope into a small set of condition-linked operator actions with explicit prerequisites and exclusions.","counterfactual_removal":"Without the menu, discretion remains symbolic or expands into unsafe improvisation."},{"mechanism_slug":"decision_rights_matrix","role":"Assign each deviation class to operator action, engineer confirmation, or immediate escalation and tie accountability to that allocation.","counterfactual_removal":"Without explicit rights, approval-seeking latency and retrospective blame can persist even when useful observations are available."},{"mechanism_slug":"model_assumption_register","role":"Expose the assumptions connecting process signals, batch state, selected action, and predicted directional response.","counterfactual_removal":"Without registered assumptions, unexpected responses cannot reliably update the model or future action selection."},{"mechanism_slug":"action_effect_feedback_review","role":"Compare the predicted and observed response at the next checkpoint and decide whether the action rule remains usable.","counterfactual_removal":"Without action-linked review, final quality results arrive as judgment rather than feedback capable of changing the next decision."}],"causal_chain":["The batch begins with a visible material-property target and hard operating constraints.","Stage-specific observations are compared with an explicit expected trajectory rather than displayed as disconnected readings.","The operator classifies the deviation only if sensor confidence and the remaining correction window satisfy documented conditions.","The selection rule identifies a prevalidated action whose predicted directional effect addresses the inferred drift.","The Decision Rights Matrix allows the operator to execute that action locally or routes the case to escalation.","The control system executes the bounded action while retaining existing interlocks and logs the decision context.","The next relevant observations are compared with the predicted response.","Agreement supports continued use of the model-action link; disagreement flags the assumption, narrows or suspends the option, and informs the next batch review.","Accountability is assigned according to the observations, options, authority, and resources that were actually available."],"baseline":"Operators run a fixed master recipe, watch alarms and trend screens, document deviations, and contact a supervisor or process engineer for adjustments. Final laboratory results determine whether the batch passed, but they are weakly linked to the specific local observations, delayed approvals, and actions that produced the outcome.","nearest_rivals":["A tighter fixed recipe with narrower alarms and mandatory escalation","Statistical process-control charts that detect drift but do not grant action rights","A permanently staffed centralized process-engineering approval desk","A fully automated model-based controller that removes routine operator selection","Additional final-product testing without changing in-batch decisions"],"remaining_contrastive_claim":"The testable contrastive claim is that a deviation response becomes materially different from alarms, recipe tightening, or decision-rights clarification alone when the same bounded operator loop couples a represented property goal, an inspectable process model, feasible actions, legitimate execution authority, action-specific feedback, and an explicit model-update rule. This claim concerns the presence and use of the coupling, not a claimed effect size.","authority_safety":{"decision_authority":"The reactor operator may choose only actions already validated for the current chemistry, equipment, batch stage, and state envelope. The process engineer owns model and menu revisions; process safety owns hard constraints and interlocks; the shift supervisor owns operational escalation. No intervention component overrides existing emergency authority.","authorized_first_step":"Conduct an offline replay of one completed deviation using recorded data and a draft control card; participants may classify the state and select hypothetical menu actions but may not alter a live recipe or control system.","excluded_actions":["Changing monomer, initiator, catalyst, solvent, or additive identity or total charge","Disabling alarms, interlocks, trips, containment controls, or required sampling","Exceeding validated temperature, pressure, agitation, feed, or hold-time bounds","Using an action when sensor confidence or batch-stage identity is unresolved","Applying the menu to a chemistry, scale, vessel, or recipe version for which it was not reviewed","Assigning the operator responsibility for outcomes outside the mapped controllability boundary"],"halt_rollback":"Suspend the loop and revert authority to the existing approved procedure and escalation chain if indicators conflict, sensor confidence falls, an action produces the opposite directional response, or the trajectory approaches a hard limit. In any later live pilot, the established safe-state or emergency procedure takes precedence; the menu itself cannot prescribe improvisational recovery."},"negative_tests":{"strongest_counterevidence":"Recorded deviations show that operators already receive an interpretable state estimate, possess timely authority over effective bounded actions, observe their effects, and update subsequent choices; remaining failures instead arise from factors that neither the operator nor the proposed action set can observe or influence.","problem_falsifier":"For sampled deviation episodes, no consequential delay or broken link can be found between goal, observation, model, feasible action, authority, execution, and feedback, or the hypothesized correction window consistently closes before any reliable deviation is observable.","intervention_falsifier":"In offline replay, qualified operators using the control card cannot select an eligible action before the recorded correction window, cannot state the action's predicted directional effect, or repeatedly require information and authority outside the proposed boundary; a later authorized pilot would also falsify the intervention if action-linked feedback fails to change subsequent selection or causes more boundary violations.","risks":["The simplified world model could encourage confident but incorrect intervention in a nonlinear reaction.","Operators could infer permission beyond the validated menu or apply a rule to the wrong chemistry or scale.","More sampling or holding could itself perturb the process or consume the correction window.","Management could use the agency-health record to intensify blame instead of correcting missing controllability.","A menu designed from successful cases could omit rare precursors to unsafe excursions.","Added checkpoint work could overload the operator during abnormal conditions.","Process engineers could fail to retire an action after contradictory feedback, allowing model error to persist."]},"next_evidence_step":"Using only one completed off-spec or near-limit batch and its existing time-stamped process record, run a 90-minute facilitated replay with one operator, one process engineer, and one process-safety reviewer. Freeze the replay immediately before the historical escalation, populate an Agency Loop Map, and require the operator to identify the goal, inferred state, trusted observations, eligible Safe Action Menu item, predicted directional response, authority class, and halt condition. Record whether every loop link can be specified from information that was actually available at that time. Do not operate equipment or change a recipe; the result determines only whether a narrowly bounded shadow-mode design is coherent enough for further review.","prior_art_status":"UNSEARCHED","diversity_from_prior_proposals":"No comparison with other proposals was performed under runtime isolation; this is a single domain-specific candidate centered on bounded human control of drifting polymerization batches.","revision_record":{"parent_version":null,"progress_targets_addressed":[],"conceptual_changes":[],"operational_changes":[],"evidence_changes":[],"claim_changes":[]}}