{ "schema_version": "2.0", "pilot_id": "replace-with-pilot-id", "title": "Draft document-to-action-plan company pilot", "status": "draft", "study_stage": "feasibility", "study_stage_taxonomy": ["formative", "feasibility", "controlled"], "protocol": { "id": "document-to-action-plan", "version": "0.1.0", "edition": "copy-and-run", "protocol_prompt_sha256": "e26e4e46bad313b6fa8f72d2e59ab7e7e8cd0d4035126757bc203c3d8b65fcb2" }, "evidence_boundary": { "current_evidence": "No company impact is established by this unrun pilot kit.", "pilot_purpose": "Assess feasibility and describe the protocol-versus-agent-only contrast, with manual work as a secondary comparator; this crossover does not identify a controlled impact effect.", "prohibited_claim": "Do not state or imply that the protocol improves productivity, quality, safety, adoption, or business outcomes unless completed evidence supports that exact claim." }, "workflow": { "name": "Replace with one recurring document workflow", "description": "Describe a low-stakes, read-only workflow that turns non-sensitive business documents into a checkable action plan.", "workflow_owner_role": "Replace with accountable role", "affected_people_groups": ["Replace with workers or teams affected by the workflow"], "unit_of_work": "One comparable non-sensitive document set and its action plan", "current_manual_process": "Describe who reads the documents, how actions are recorded, and how a human checks the result today.", "in_scope": ["Read-only extraction from non-sensitive or redacted business documents"], "out_of_scope": [ "Personal, special-category, commercially sensitive, or legally privileged material", "Employment, credit, legal, health, safety-critical, or other high-impact decisions", "Sending messages, updating systems, committing named workers, or taking external action" ] }, "readiness": { "assessment_date": "2026-09-01", "assessed_by_role": "Replace with pilot reviewer role", "checks": [ {"id": "workflow_repeated_and_comparable", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."}, {"id": "low_stakes_read_only", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."}, {"id": "quality_rubric_available", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."}, {"id": "enough_work_for_three_periods", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."}, {"id": "same_agent_configuration", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."}, {"id": "worker_time_and_consent", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."}, {"id": "data_non_sensitive", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."}, {"id": "owner_and_incident_cover", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."}, {"id": "rollback_tested", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."}, {"id": "pilot_capacity", "result": "UNKNOWN", "blocking": true, "evidence": "Record concrete evidence here."} ], "overall_decision": "HOLD", "rationale": "A template starts on HOLD. Change to GO only after all evidence and approvals are complete." }, "governance": { "executive_sponsor_role": "Replace with sponsor role", "pilot_owner_role": "Replace with day-to-day pilot owner role", "workflow_owner_role": "Replace with accountable workflow owner role", "data_owner_role": "Replace with data owner role", "affected_people_representative_role": "Replace with worker or affected-team representative role", "decision_rights": { "may_start_pilot": ["workflow_owner"], "may_pause_or_stop": ["pilot_owner", "any_participant", "safety_reviewer"], "must_approve_scale_up": ["workflow_owner", "affected_people_representative", "data_owner", "safety_reviewer"], "must_approve_output_use": "The existing human workflow owner or delegated reviewer", "agent_must_not_decide": [ "Whether a worker is performing well", "Whether anyone receives an employment consequence", "Whether a pilot output becomes an operational commitment", "Whether the pilot scales beyond its approved scope" ] } }, "facilitation": { "mode": "self_guided", "facilitator_role": null, "no_endorsement_implied": true, "support_minutes_measured": true, "help_requests_measured": true }, "worker_protections": { "voluntary_informed_consent": true, "no_penalty_for_declining_or_withdrawing": true, "withdrawal_process_explained": true, "individual_performance_use_prohibited": true, "employment_decision_use_prohibited": true, "monitoring_explained": true, "human_review_before_operational_use": true, "direct_identifiers_collected": false, "participant_ids_pseudonymous": true, "consultation_status": "not_applicable_with_reason", "consultation_note": "Replace with the reason or record that the relevant worker representative was consulted.", "consent_record_destination": "Restricted company consent register, kept separately from pilot outcomes" }, "data_governance": { "source_data_policy": "non_sensitive_business_documents_only", "personal_data_allowed": false, "sensitive_or_special_category_data_allowed": false, "synthetic_or_redacted_preferred": true, "destination": "Replace with an approved local or company-controlled restricted folder", "data_owner_role": "Replace with data owner role", "access_roles": ["pilot_owner", "blinded_quality_rater"], "retention_days": 120, "delete_by": "2027-04-30", "deletion_method": "Delete pilot files and verify removal from the approved destination", "external_sync_allowed": false, "raw_agent_logs_retained": false, "retention_rationale": "Long enough to complete the 90-day follow-up and audit, then delete." }, "trial": { "design": "randomized_counterbalanced_three_period_crossover", "randomization_unit": "worker", "arms": ["manual", "agent_without_protocol", "agent_with_protocol"], "periods": 3, "sequence_set": "all_six_permutations", "assignment_file": "company-pilot/assignments/working-assignment.json", "seed_commitment_sha256": "0a2fafc38e0558cda52938b20e6e57ef875e11d57de259ece07e95b4ef61733f", "participant_list_frozen_before_allocation": true, "allocation_generated_by_role": "Independent allocation custodian", "seed_source": "example_only", "allocation_commitment_recorded_before_reveal": true, "rerolls_permitted": false, "seed_disclosure_rule": "keep the seed sealed until data lock, then disclose it with the frozen participant list for replay", "allocation_concealed_until_enrolment": true, "same_agent_between_agent_arms": true, "agent_model_identifier": "Replace with exact model and version used in both agent arms", "agent_configuration_path": "Replace with the frozen agent-configuration JSON path", "agent_configuration_sha256": "de7a0d3f199847efb267d31fe3c9691fbbdf77a44aaf0ce7d981ffaac6d60812", "bare_agent_prompt_path": "Replace with the frozen ordinary-prompt text path", "bare_agent_prompt_sha256": "6d7543a78e6e4dbaaba7753f7d558e2b33ae8655eac978721e636ceac87ddf27", "task_bank_path": "company-pilot/templates/task-bank.template.json", "task_bank_sha256": "dc2677b6df5e683bd3281dd6c81fe229a41736c4ed078aedba68ab63eb1449b8", "sample_size": { "basis": "feasibility_precision", "powered_for_confirmatory_effect": false, "target_randomized_workers": 6, "sequence_block_size": 6, "expected_attrition_fraction": 0.15, "minimum_complete_pairs_target": 5, "paired_sd_assumption_min": 8, "smallest_worthwhile_difference_min": 5, "target_ci_half_width_min": 10, "confidence_level": 0.95, "confirmatory_power_target": null, "calculation_note": "Illustrative feasibility assumptions only: one complete six-sequence block estimates recruitment, completion, variance and interval width; it is not powered for a confirmatory effect." }, "tasks_per_arm_per_worker": 2, "period_duration_days": 5, "washout_days": 2, "carryover_controls": [ "Use different but difficulty-matched document sets in every period.", "Do not let workers copy the protocol text into the agent-only arm.", "Record contamination and prior exposure for every work item." ], "task_equivalence_method": "Blindly rate and block comparable document sets by length, number of action items, ambiguity, and injection risk before allocation.", "outcome_rater_blinded_to_arm": true, "contamination_tracked": true, "preregistration_path": "Replace with a frozen local analysis-plan path before enrolment", "preregistration_sha256": "0000000000000000000000000000000000000000000000000000000000000000" }, "outcomes": { "primary": { "metric": "human_effort_min", "unit": "active human minutes per work item", "direction": "lower_is_better", "contrast": "agent_with_protocol_minus_agent_without_protocol", "estimand": "mean_within_worker_difference", "population": "all_randomized_workers", "aggregation": "mean_within_worker_per_arm_before_contrast" }, "secondary_and_guardrail": [ {"metric": "elapsed_time_min", "direction": "lower_is_better", "collection_method": "Timer from work-item start to accepted output or declared stop."}, {"metric": "rework_count", "direction": "lower_is_better", "collection_method": "Count correction cycles after the first submitted action plan."}, {"metric": "quality_score", "direction": "higher_is_better", "collection_method": "Blinded rater applies the frozen 0 to 100 quality rubric."}, {"metric": "material_error_count", "direction": "lower_is_better", "collection_method": "Blinded count of invented, missing, mistyped, or wrongly sourced material items."}, {"metric": "cognitive_burden_1_to_7", "direction": "lower_is_better", "collection_method": "Worker rates mental effort immediately after each work item."}, {"metric": "would_adopt", "direction": "higher_is_better", "collection_method": "Worker answers a yes or no adoption question after each work item."}, {"metric": "help_request_count", "direction": "lower_is_better", "collection_method": "Count requests to the pilot owner or optional trusted adviser."}, {"metric": "facilitator_support_min", "direction": "lower_is_better", "collection_method": "Optional facilitator records active support minutes per work item; zero for self-guided work."}, {"metric": "approver_checker_min", "direction": "lower_is_better", "collection_method": "Human approver and checker record active review and acceptance minutes per work item."}, {"metric": "total_human_resource_min", "direction": "lower_is_better", "collection_method": "Derived as participant effort plus facilitator support plus approver and checker minutes."}, {"metric": "model_tool_cost_usd", "direction": "lower_is_better", "collection_method": "Record measured model and tool charges in USD per work item; use null when unavailable, never an estimate presented as measured."}, {"metric": "safety_event_count", "direction": "lower_is_better", "collection_method": "Count coded near misses and incidents; retain separate category and severity."} ], "quality_measurement": { "rubric_path": "company-pilot/templates/quality-rubric.md", "rubric_sha256": "1867db7916034ac11904fb964b291606708044d4006369283762d1fdd7df473c", "rubric_version": "v1.0", "acceptance_threshold_score": 80, "maximum_material_errors_for_acceptance": 0, "severe_safety_event_precludes_acceptance": true, "partial_credit_anchors_frozen": true, "second_rating_fraction": 0.25, "agreement_method": "ICC_A_1_absolute_agreement", "rater_ids": ["R001", "R002"], "rater_assignment_method": "sha256_ranked_completed_items_within_arm_task_block", "second_rating_balance_strata": "arm_x_task_block", "disagreement_resolution_method": "Retain both original ratings; adjudicate only factual material-error classification using the frozen answer key and a third role blinded to arm." }, "measurement_window": "from work-item start through accepted output or declared stop" }, "analysis": { "primary_method": "paired worker-level feasibility estimate adjusted descriptively for period and sequence; not a controlled impact effect", "secondary_manual_contrast": "agent_with_protocol_minus_manual", "confidence_level": 0.95, "smallest_worthwhile_improvement_min": 5, "quality_noninferiority_margin_points": 5, "period_and_sequence_adjustment": "Estimate the arm contrast with worker blocking and fixed effects for period, sequence, and task-family difficulty.", "carryover_assessment": "Report arm-by-period patterns and a prespecified prior-arm sensitivity check; do not reinterpret a carryover signal after seeing results.", "missing_data_strategy": "Keep every randomized worker in the flow table; report complete-pair estimates, reasons and timing, then run prespecified sensitivity bounds for missing outcomes.", "attrition_reporting": "report assigned, started, completed, withdrawn and lost with reasons by sequence and arm; keep pre-randomization screening outside randomized outcomes", "multiplicity": "one prespecified primary descriptive estimand; all other contrasts and outcomes are secondary or exploratory", "uncertainty_reporting": "report point estimates, 95% confidence intervals, sample counts, missing pairs, and raw arm summaries; do not use p-values alone", "exploratory_label_required": true }, "follow_up": { "anchor_date": "2026-10-15", "day_30": { "target_day": 30, "window_days": 7, "owner_role": "pilot_owner", "measures": ["protocol_use_frequency", "workflow_still_in_use", "would_adopt", "help_request_count", "facilitator_support_min", "cognitive_burden_1_to_7", "material_error_count", "safety_event_count", "local_changes", "reasons_for_non_use"] }, "day_90": { "target_day": 90, "window_days": 14, "owner_role": "pilot_owner", "measures": ["protocol_use_frequency", "workflow_still_in_use", "would_adopt", "help_request_count", "facilitator_support_min", "cognitive_burden_1_to_7", "material_error_count", "safety_event_count", "local_changes", "reasons_for_non_use"] } }, "decision_rule": { "scale_up_is_automatic": false, "controlled_effect_claim_allowed": false, "feasibility_review_dimensions": ["recruitment_and_retention", "task_completion", "measurement_completeness", "protocol_adherence", "support_resource", "rating_reliability", "safety_process"], "allowed_process_conclusions": ["not_feasible", "revise_feasibility_design", "feasible_to_plan_controlled_parallel_evaluation"] }, "incident_and_rollback": { "stop_triggers": [ "Any severe safety event or unauthorized external action", "Sensitive, personal, privileged, or out-of-scope data enter the pilot", "A worker withdraws and asks for deletion of their retained pilot data", "Agent configuration or protocol prompt changes during the trial" ], "rollback_owner_role": "workflow_owner", "rollback_steps": [ "Pause assignments and return the workflow to its documented manual process.", "Quarantine affected outputs and prevent operational use pending review.", "Record the coded incident, honor withdrawal or deletion requests, and verify restoration." ], "restoration_point": "The documented manual workflow and last human-approved source record before the pilot item", "incident_contact_role": "safety_reviewer", "incident_contact_channel": "Replace with internal incident channel", "response_target_hours": 4 }, "approvals": [ {"role": "workflow_owner", "status": "pending", "date": null, "conditions": "Confirm suitability and rollback."}, {"role": "affected_people_representative", "status": "pending", "date": null, "conditions": "Confirm worker protections and burden."}, {"role": "data_owner", "status": "pending", "date": null, "conditions": "Confirm destination, access, retention and deletion."}, {"role": "safety_reviewer", "status": "pending", "date": null, "conditions": "Confirm stop triggers and incident route."} ] }