# Evolution Loop Playbook > Copyright (c) 2026 Paranoia. Licensed under the MIT License. ## 1. Controlled Evolution Self-evolving AI means improving external system layers under stable goals: prompts, memory, retrieval, tool routing, workflows, schemas, eval sets, and skill references. It does not mean uncontrolled model-weight mutation. ```text real task pressure -> WOOP Task Card -> Decision Object -> RJR-AI residual judgment authority gate -> VOI decision gate -> Scenario VOI Adapter -> smallest information probe -> UL exposure and attribution gate -> Orient-first OODA -> result feedback -> candidate mutation -> behavior eval -> Human Gate -> versioned promotion -> rollback path ``` ## 2. WOOP Admission Wish defines intent, artifact, scope, and stop condition. Outcome defines the acceptance rubric and decision value. Obstacle identifies internal human-agent Failure Patterns. Plan defines executable If-Then Protocols, judges, retries, handoff, and rollback. Unclear Wish means clarify or shrink. Unclear Outcome means draft only. Unclear Obstacle lowers autonomy. Unclear Plan blocks high-risk and irreversible actions. ## 3. Decision Object Before VOI, declare the decision question, real options, current default action, owner, deadline, stakes, reversibility, and boundary status. Without a default action there is no action-change test; without options there is no decision comparison; without a deadline research can become avoidance. ## 4. RJR-AI: Residual Judgment Authority Gate RJR-AI is not a new agent and does not replace leadership. It is the authority layer between the Decision Object and VOI. It answers: ```text What can AI expand into options? What must Workflow compress? What can Eval judge? What can automation execute reversibly? What can memory retain? What high-coupling, low-reversibility, under-evidenced directional bet remains human residual judgment? ``` Every evolution proposal should declare `rjr_authority_gate`: ```yaml rjr_authority_gate: residual_judgment: "directional choice reserved for the human" coupling: "low | high" reversibility: "reversible | costly_to_reverse | irreversible" authority_level: "P0_read | P1_suggest | P2_draft | P3_reversible_execute | P4_approved_execute" delegation_matrix: ai: "read | suggest | draft | execute" workflow: "none | route | gate | orchestrate" eval: "none | sample_check | regression | acceptance_gate" automation: "none | reversible_only | approved_execution" human: "not_required | review | approve | decide" ``` Default split: | Type | Default authority | | --- | --- | | Low coupling, high reversibility | AI / automation may execute with sampling | | Low coupling, low reversibility | AI drafts; workflow, eval, or expert rules confirm | | High coupling, high reversibility | AI explores; workflow narrows; eval compares; human chooses | | High coupling, low reversibility | AI only supports reasoning; Human Gate is required | The goal is to remove low-value human judgment, not human responsibility. If `coupling: high`, reversibility is costly or irreversible, and evidence is still insufficient for a stable choice, the output must preserve `residual_judgment` instead of presenting the AI's synthesis as a decision. ## 5. Scarce Resources Treat tokens, time, attention, tool calls, context, trustworthy feedback, labeled samples, money, trust, and decision windows as scarce. AI makes information generation cheap while increasing evaluation and closure costs. Optimize for closing low-value branches, not maximizing output volume. ## 6. VOI Gate Use `references/value-of-information-playbook.en.md` for the full method. ```text 1. What decision is being supported? 2. What action is taken with current information? 3. Which uncertainties can change the option ranking? 4. Which concrete information actions target them? 5. What action follows each plausible signal? 6. What are the EVPI ceiling and realistic EVSI? 7. What are acquisition, delay, attention, privacy, and contamination costs? 8. What is the smallest high-value probe and its stop rule? ``` After the generic VOI gate, select a Scenario VOI Adapter: `skill_evolution`, `game_direction`, `experience_diagnosis`, `source_curation`, `content_decision`, `platform_fact`, `high_risk_action`, or `ai_branch_management`. The adapter defines valid evidence, weak evidence, default probe, and domain stop rule. It must not replace the Decision Object. ```text approx_net_voi = P(action_switch) × decision_delta × reuse_or_scale × reversibility_factor - total_information_cost ``` This is a triage heuristic. If all plausible signals lead to the same action, stop or classify the work as model learning or information consumption. ## 7. UL (Uncertainty Ladder) VOI selects the uncertainty worth reducing. UL controls how much uncertainty the next probe exposes. See `references/uncertainty-ladder-protocol.en.md`. ```text build model -> isolate actions -> controlled composition -> expose failure -> diagnose bottleneck -> targeted practice -> increase complexity -> test transfer -> update model ``` The rungs are `UL-L0`, `UL-L1`, `UL-L2`, `UL-L3`, `UL-L4`, and `UL-L5`. This is not a scalar difficulty score or a fixed waterfall. Track input novelty, context ambiguity, tool/environment variability, coordination, authority/consequence, and evaluation ambiguity separately. Release one major variable per round by default. Record `held_constant`, `scaffolds_present`, `consequence_budget`, and `fallback_rung`. If ablation, controls, or counterfactuals cannot distinguish leading explanations, mark the failure `confounded` and return to a more controlled environment. Do not build permanent rules on confounded evidence. Controlled success supports only the current rung. Remove scaffolds progressively and require transfer plus negative-transfer evidence before claiming robustness. Authority, publishing, money, long-term memory, and real-user impact never auto-promote with the rung; RJR-AI and Human Gate still govern them. ## 8. Orient-First OODA Observe goals, evidence, surprise, failures, corrections, costs, latency, and triggered Obstacles. Orient around the current frame, obsolete narratives, decision boundary, user/domain/operating models, and whether the information is decision, model, or consumption value. Decide one action or probe and explicitly reject low-VOI branches. Act through an artifact, tool call, or bounded test. Evaluate the chain: ```text prior -> signal -> posterior -> action_before -> action_after -> stop reason ``` ## 9. Model Compression Gate Check whether the model is too short to expose mediators, too long to execute, or merely moving complexity. Every new rule should name its mediator, control point, observation, validation, and effect on future decision cost. ```text total_description_cost = core_model_length + routing_rule_length + state_injection_length + validation_observation_length + exception_patch_length + failure_recovery_length ``` ## 10. Task and Meta Loops ```text Decision Object -> VOI Gate -> UL -> smallest probe or direct action -> result -> Human Gate / stop ``` ```text trace -> repeated/high-impact failure -> bottleneck attribution -> mutation candidate -> behavior eval -> approval -> promotion/rollback ``` Never promote a permanent rule from one attractive case. More complete prose is not evidence of lower decision error or lower attention cost. ## 11. Candidate Mutation Gate A mutation enters the queue only when it is repeated or high-impact, can change a system decision or reduce major risk, is repairable, reusable, evidenced, reversible, cheaper than its expected benefit, tested against a counterexample, and has a stop rule. Its current rung, exposed variables, and attribution must be explicit; controlled success must not be presented as transfer proof. ## 12. AI Fatigue and Anti-Generic Output ```text open branches -> map to decisions -> keep action-changing branches -> archive model learning -> close consumption -> choose one probe ``` Preserve local facts, negative feedback, failure details, provenance, constraints, and action impact. Added headings and abstract structure without added decision boundaries, signals, costs, or actions are high-structure/low-VOI output. ## 13. Permission Ladder | Level | Example | Default | | --- | --- | --- | | A0 | analysis, draft, VOI audit | automatic | | A1 | read-only research, local inspection, log sampling | automatic with decision/cost trace | | A2 | docs, templates, candidate skill files, reversible experiment | automatic with backup, validation, and stop rule | | A3 | long-term memory, global skill install, production policy | Human Gate | | A4 | deletion, publishing, money, real-user impact | explicit approval | ## 14. Public Skill Checks The package must keep names aligned, keep `SKILL.md` lightweight, route to one-hop references and templates, include decision/default-action/signal-to-action/cost/stop fields, cover VOI and uncertainty-ladder failure cases in behavior evals, preserve public/private boundaries, and retain rollback.