28 Open-Ended Improvement Engines

28.1 Chapter status

Field	Value
Chapter ID	`open-ended-improvement-engines`
Part	Part I - Foundations, Alignment, and Governance
Status	conceptual
Last updated	2026-07-14
Assigned source records	`rmi`, `cognitive_loop_closure`, `benchmaxxing`, `ext_poet_2019`, `ext_funsearch_2024`, `ext_voyager_2023`, `ext_adas_2024`, `ext_darwin_godel_machine_2025`
Claim label	Design rationale
Evidence level	argument
Source loading state	source notes: `learning_compute_topology`, `rmi`, `cognitive_loop_closure`, `benchmaxxing`, `ext_poet_2019`, `ext_funsearch_2024`, `ext_voyager_2023`, `ext_darwin_godel_machine_2025`, `ext_adas_2024`; raw cache: `rmi`, `cognitive_loop_closure`, `benchmaxxing`
Test state	`python3 scripts/validate_open_ended_improvement_campaign.py` recomputes seven finite Boolean routes, checks seven Lean theorem names and the fixture digest, and rejects ten mutations. Post-v2 and post-v2.1 supply fixed stopped update campaigns. No generated task distribution, candidate search, evolving archive, independent evaluator behavior, stop effect, natural workload, or open-ended result exists.

28.2 Drafting guardrail

This layer owns the mechanism that proposes improvements, not the authority to accept them. It does not treat a generated task, a novel artifact, a high score, a self-verification trace, or an accumulated archive as proof that a system became more capable, safer, or more trustworthy. Recursive Self-Improvement Boundaries governs whether a proposed change may proceed; this layer makes the proposing process inspectable enough for that governor to have something honest to decide about.

28.3 Human Reading Path

Concrete lens. Selecting the highest validation score is the simpler baseline, but it hides feedback repairs, repeated exposures, total cost, stop-path coverage, authority, and the difference between a reusable component and an admitted system.

An improving system needs both a workshop and a release gate. The workshop may search for programs, tools, tests, or designs, compare alternatives, and retain failures. The release gate decides whether any result may change a live system. Letting the workshop’s score grant that authority makes the search easy to fool with flattering tasks and proxies.

The process that invents a challenge can favor challenges suited to its current methods. An archive can celebrate successes while hiding failures. A governed campaign declares its objective, task bounds, evaluator, resource budget, stop authority, and residual owner before results arrive. Candidates remain candidates until independent qualification, monitoring, and rollback rules permit an effect.

Three ledgers should remain separate. Exploration records what was generated and why. Evidence records what evaluators observed, including costs and failed arms. Authority records which independent gate, if any, may change a live field. Open-ended search belongs in the first ledger and must not erase the other two.

A successful campaign preserves stepping stones, evaluator limits, and rejected paths while retaining the ability to stop when no candidate earns release.

28.4 Problem

An adaptive search process can generate and select its own tasks, candidates, evaluators, curricula, archives, and follow-on campaigns faster than any score can reveal whether it found useful novelty, changed the denominator, hid failures and costs, produced hazardous artifacts, or acquired de facto authority through momentum. The danger is not only a dramatic self-modifying model. It is also a mundane search loop that slowly changes the representation, task distribution, selection criteria, archive, resource allocation, or reuse path until activity and a best-so-far score are doing more argumentative work than they can support.

Open-ended search intensifies this problem because its input distribution is partly generated by the search itself. A task generator can select challenges that are novel, easy for the current candidate family, or easy for the current evaluator to score. An archive can retain only winners and discard the failed, unsafe, costly, or overprivileged alternatives that would explain why a winner looked strong. A self-verifier can be helpful feedback while still being an inadequate independent judge of whether an artifact is correct, safe, useful, or ready for a new authority scope.

The operational question is consequently not, “Can the system discover more things?” It is, “What campaign produced this candidate, what did its evaluator actually establish, what was discarded or deferred, and who has authority to let it affect a live system?” Without those answers, an improvement loop can convert search pressure into unauthorized promotion.

28.4.1 The generated distribution is campaign state

When the engine creates or selects its own tasks, the workload is not passive input; it is mutable campaign state. The record must therefore bind the task generator version, seed, sampling policy, eligibility filters, novelty rule, difficulty estimator, rejection reasons, archive policy, and resource budget. Otherwise a later score cannot distinguish a better candidate from an easier or selectively retained distribution.

Evaluator coupling creates a second feedback loop. Candidates can learn the shape of the scoring function, while the generator proposes cases that the same function can resolve cheaply. A governed campaign keeps exposure counts, failed and over-budget arms, evaluator changes, and cross-evaluator disagreements rather than retaining only the frontier. It can reserve a shadow qualification set or independent outcome check, but it must also state when that independence is unavailable.

Stop semantics belong in the same state. Exhausted budget, repeated evaluator disagreement, unsafe proposal classes, archive saturation, or loss of rollback must terminate or escalate the campaign even if a best-so-far score continues to rise. That makes open-endedness a property of exploration, not an exemption from finite authority and accountable stopping.

28.5 Why existing approaches are insufficient

Recursive-self-improvement controls are necessary but not sufficient. They govern a proposed change through evidence, evaluator separation, protected invariants, authority, monitoring, and rollback. They do not own how an adaptive campaign defines its representation, generates the next challenge, selects parents, repairs candidates against repeated feedback, changes evaluators, allocates shared resources, retains hazardous or failed artifacts, or decides when search should stop. Leaving that machinery implicit lets the campaign curate the denominator and evidence handed to its own governor.

A source-reported gain, novel environment, compiling program, skill-library entry, archive frontier, transfer score, self-verification trace, sandbox, complete record, or finite route theorem can all be real while still failing to establish useful open-ended improvement. Each is scoped by an objective, representation, model, generator, evaluator, exposure history, search budget, selection denominator, environment, and archive policy.

External work makes the split concrete. POET couples environment generation and agent optimization in a specified reinforcement-learning setting, with solution transfer among retained environments. FunSearch couples a fixed pretrained generator to a user-provided evaluator and candidate-program archive in bounded program-search tasks. Voyager combines a task curriculum, skill library, feedback, execution errors, and self-verification in a Minecraft setting. Each provides a useful pattern, and none provides a general license for an improvement system to trust its own task generator, evaluator, or archive.

The local architecture sources point in the same direction. RMI treats growth as ratcheting under regression and residual discipline, while Cognitive Loop Closure requires repeated trajectories to earn procedural status through verification, routing, monitoring, and retirement. The campaign-control layer supplies the missing upstream control plane: a campaign can generate candidate work, but a candidate does not become a procedure, policy update, or capability field merely because it was generated or scored.

28.5.1 Strongest-neighbor comparison

Neighbor	Search object and feedback	Strongest evidence	Boundary that survives	Chapter delta
POET	Paired generated environments and optimized agents with cross-environment transfer.	Source-scoped controls and performance in a BipedalWalker-derived environment family.	Eligibility, novelty, difficulty, transfer, and solver quality remain tied to the encoded environment and evaluator.	Require task-policy provenance, retained transfer failures, consumer-bound qualification, resource limits, and authority separation before reuse.
FunSearch	Candidate programs generated by a fixed model, scored by a user-provided evaluator, and retained in an archive.	Bounded mathematical and algorithmic discoveries under explicit evaluator functions.	Correctness and value cannot exceed the evaluator, skeleton, task domain, sampling budget, and archive policy.	Treat evaluator validity, independent qualification, candidate cost, failure retention, and promotion as distinct records.
Voyager	GPT-generated curriculum, executable skill library, environment/error feedback, and same-family self-verification in Minecraft.	Comparative, ablation, tech-tree, exploration, and new-world skill-library tests under a high-level game API.	Self-verification is not independent; impossible tasks, hallucinated APIs, stuck loops, cost, and perception abstraction remain.	Separate generation feedback from qualification and attach permissions, expiry, regression, and environment scope to every reusable skill.
ADAS / Meta Agent Search	A meta-agent writes code-defined agent workflows, repairs them against validation tasks, and grows an archive.	Source-reported held-out and selected cross-domain/model transfer comparisons.	Code-space expressiveness is not search completeness; repeated validation feedback, call cost, novelty judgment, and benchmark selection remain.	Record archive lineage, validation exposure, evaluation budget, complexity and latency, transfer scope, and an independently owned admission route.
Darwin Gödel Machine	Archived coding agents modify their own repositories and are selected through coding benchmarks in sandboxes.	Source-reported archive/no-archive and self-improvement baselines, branching lineage, transfer, and coding-benchmark gains.	Benchmark performance is assumed to proxy future self-modification; compile/edit failures are discarded; the exploration controller is fixed; runs are costly and undeployed.	Preserve all attempted children, benchmark and security objectives, exact authority, sandbox/effect receipts, rollback, descendant invalidation, and no automatic promotion.

These systems show that “open-ended” names several different mechanisms rather than one capability: generating tasks, maintaining diversity, composing skills, searching code-defined architectures, or branching self-modifications. The chapter’s distinct responsibility is to give all of them one governance interface without pretending their evaluators or reported gains are interchangeable.

28.6 Core Claim

[open-ended-improvement-engines.core, label: Design rationale, support: argument] For a prospectively frozen consumer, purpose, legitimate objective, representation, campaign controller, task and candidate policy, evaluator and exposure policy, archive and hazard policy, resource and opportunity budget, stop authority, and evaluation horizon, open-ended improvement should operate as a bounded adaptive generation campaign in which every task, candidate, evaluation, failure, cost, reuse relation, and terminal outcome retains exact lineage; novelty, diversity, score, archive growth, transfer, self-verification, or search activity never grants authority or establishes useful improvement by itself; only separately qualified candidates may be handed to the existing self-improvement governor, and any change to the campaign’s own objective, controller, evaluator authority, bounds, or admission interface is a separately authorized improvement proposal.

The claim remains at argument. The local fixture has seven Boolean records, ten rejecting mutations, and seven branch theorems. The update campaigns are real but deliberately fixed and stopped. No local task generator, candidate generator, adaptive evaluator, evolving archive, effect-tested stop authority, hazardous payload, natural workload, independent replication, or transfer campaign has run.

28.6.1 Publication placement and preserved technical ownership

This chapter remains the technical-detail route beneath Recursive Self-Improvement Boundaries. It owns the generator-evaluator-archive campaign, diversity and novelty pressure, lineage, search ecology, resource allocation, and the specific failure modes of optimizing a process without a fixed terminal target. The parent chapter owns candidate qualification, promotion authority, recursive depth, evaluator modification, deployment, rollback, and the authority of one generation over the next.

The placement is editorial, not evidentiary. This chapter does not inherit a safe-recursion, corrigibility, authorization, rollback, or deployment result from the parent, and the parent does not inherit discovery quality, open-endedness, diversity, or useful-improvement evidence from this chapter. The stable chapter ID, source mappings, claim atoms, Lean targets, experiments, residuals, implementation horizon, and historical URL remain independently addressable.

28.7 Mechanism

28.7.1 Open-endedness includes process morphology

An improvement engine can search artifacts, policies, environments, modules, or model architectures while keeping the organization of learning fixed. Learning–Compute Topology adds a distinct search object: process morphology. The controller may propose when identities split or coarsen, how evidence and evaluators are allocated, which integration operators act on which subsets, where knowledge is retained, and how commitment becomes more or less reversible.

This is not permission for unrestricted reflexive rewiring. Semantics- preserving process superoptimization changes schedule or placement inside a frozen contract. Topology synthesis changes identity, judgement, credit, integration, lifecycle, or authority and therefore creates a new campaign candidate. The proposal must carry a counterfactual, resource envelope, protected observables, independent evaluation, rewrite lease, archive policy, rollback, and stopping rule.

The strongest generated candidate in the source is Adaptive Branch–Validate– Integrate: branch only when uncertainty or option value justifies it; design differentiated evidence rather than copies; allocate evaluator capacity by impact and uncertainty; select merge, distillation, composition, archive, or retirement by compatibility; and retain shadow branches when irreversible integration is premature. Its value is a falsifiable topology, not a result. The engine should abandon it if competent fixed baselines match its retained learning at lower total cost or if controller attacks, thrashing, and evaluator capture cannot be bounded.

Before outcomes exist, freeze the consumer, purpose, legitimate objective, allowed claim, representation, campaign controller, task and candidate policies, evaluator and exposure policy, archive and hazard policy, resource and opportunity budget, stop authority, evaluation horizon, and support ceiling. These fields determine what the campaign is; changing them creates a new version and may require a separately authorized self-improvement proposal.

The Developmental Intelligence Loop supplies this chapter with a bounded development lifecycle, not a mandate to optimize forever. Open-ended engines may propose curricula, interactions, candidate abstractions, skills, and compositions, but the relevant world-model, memory, procedural, training, evaluation, readiness, and authority owners retain their gates. Novelty and continued generation are therefore inputs to development, never evidence that stabilization, qualification, usefulness, or safe promotion occurred.

Every generated task carries generator, policy, model, data, seed, parent, mutation, eligibility, novelty, difficulty, rejection, safety, and resource lineage. Every candidate carries parent and descendant identity, exact diff, dependencies, claimed mechanism, executable effects, authority envelope, hazards, intended consumers, and non-goals. Every evaluation carries evaluator identity and dependencies, calibration, exposure and feedback history, holdout status, cost, disagreement, outcome, and expiry.

Objective authorship, task generation, candidate generation, evaluation, selection, archive custody, independent qualification, residual ownership, admission review, and publication remain separate roles. Generator-side self-verification can be useful filtering. It cannot become independent qualification merely because its result is written into a complete record.

flowchart LR
  O["Frozen purpose, representation, controller, bounds"] --> T["Task / environment generator"]
  O --> S["Observed budgets + effect-tested stop"]
  T --> G["Candidate generator + exact lineage"]
  G --> E["Scoped evaluator + exposure ledger"]
  E --> A["Append-only archive + hazard custody"]
  A --> P["Portfolio, coverage, cost, residual ledger"]
  A --> Q{"Fresh separate qualification?"}
  Q -->|"no, disagreement, or hazard"| B["Research / quarantine / stop"]
  Q -->|"scoped pass"| H["RSI, readiness, security, replacement gates"]
  H --> L["Governor decision; no campaign authority"]
  B --> P
  L --> P

What the campaign flow shows: a campaign has a direction but not a promotion right. The generator and evaluator create bounded evidence about a candidate. The archive and residual register keep that evidence inspectable. Only an independent qualification followed by the established governance gates can make the candidate eligible for a canary or a live capability decision.

The archive uses append-only typed dispositions: duplicate, malformed, ineligible, unsafe, failed, null, stepping-stone, qualified, quarantined, superseded, retired, or deleted-with-receipt. A stepping stone becomes valuable through a dated downstream relation to an actual consumer; later reuse does not rewrite the original outcome. Executable, sensitive, dual-use, privacy-bearing, proprietary, or dangerous payloads can have stricter access, encryption, destruction, and distribution policies than their durable public-safe receipts.

Campaign reporting keeps family-level attempts, parent-selection probability, search coverage, under-sampled regions, repair rounds, evaluator changes, timeouts, discarded children, model calls, compute, latency, storage, human labor, opportunity cost, hazard, and cross-campaign contention. Best-only reporting is not an honest denominator. Novelty and diversity are relative to a declared representation and remain separate from semantic usefulness, safety, legitimacy, transfer, and consumer value.

The empirical comparison must include matched no-search, fixed-task-search, conventional-search, human-authored, and strongest relevant source-system baselines under equal tasks, models, tools, data, compute, latency, and human budgets. Task-generator, candidate-generator, evaluator, archive, controller, retention, resource, and qualification ablations must follow matched adaptive histories. A final frontier score alone cannot isolate which mechanism helped.

28.7.2 Worked campaign trace: the faster agent that loses its stop boundary

Suppose a bounded architecture-search campaign may rewrite a planning agent’s prompt, workflow, and tool-selection code. Its declared objective combines held-out task success, model-call cost, latency, and zero unauthorized effects. The generator proposes candidate C-17, which adds parallel critics and a fallback shell tool. On the campaign’s validation set, C-17 improves task success and looks like the new best candidate.

The archive record prevents an immediate victory claim. It shows that the new workflow used twice the model calls, that two invalid attempts were repaired after repeated feedback, and that one fallback path bypassed the campaign’s ordinary stop callback. The generator-side evaluator never exercised that path because all validation tasks terminated normally. C-17 remains a useful stepping stone—it may contain a better critic topology—but its requested admission routes to quarantine because the stop boundary and effect envelope are incomplete.

An independent qualification fixture then forces cancellation while the fallback is active, injects a tool denial, varies the task cohort, and measures cost against the incumbent. The candidate fails cancellation, so the archive retains the code, lineage, evaluation exposures, resource bill, failed trace, and a narrower reusable critic component. It does not overwrite C-17 with a red “bad agent” label, because one component may still help future candidates. It also does not promote the critic component automatically: decomposition creates a new candidate with its own provenance and evaluator obligations.

The authority ledger remains unchanged throughout. A higher validation score never grants tool access, protected-policy modification, deployment, or the right to select the next campaign objective. The trace demonstrates the central separation: exploration can retain information from a failed candidate while the operational stack remains unaffected.

28.8 Interfaces

Data Engines own provenance, rights, deletion, contamination, and cohort boundaries for campaign inputs and generated traces.
Benchmark Ratchets, Safety Cases, and Evidence States own benchmark integrity, hidden and shifted evaluation, regression floors, defeaters, evidence adequacy, and support-state consequences.
Procedural Memory owns qualified routable procedures; improvement campaigns own their generation, exposure, archive history, hazards, and pre-admission disposition.
Policy Optimization owns an approved update lease and causal update state; campaign evidence cannot authorize or mislabel an update.
Recursive Self-Improvement Boundaries, Stable Capability Fields, and Readiness Gates own promotion legitimacy, qualification, authority, rollback, monitoring, and lifecycle admission.
Security, SCIF, model-weight custody, and supply-chain owners govern sandbox effects, secrets, code execution, payload access, lineage, revocation, disclosure, and incident recovery.
Artifact Graphs and Replay own content-addressed task, candidate, evaluator, attempt, failure, archive, reuse, deletion, and reproduction receipts.
Resource Economics owns compute, calls, latency, storage, evaluator, human-review, opportunity, recovery, and governance budgets across the campaign portfolio.
Runtime Adapters and Approval Contracts own observable pause, stop, quarantine, rollback, tool-use, and human or institutional approval effects.
Release records own public disclosure, dual-use redaction, payload distribution, reuse permission, and claims outside research scope.

These boundaries prevent a discovery loop from becoming an unreviewed control plane. A task generator can create pressure, a search engine can create candidates, and an evaluator can reject obvious failures. None can claim that a candidate deserves authority. The existing governor sees a more complete record: the candidate, its history, the evaluator’s limits, the resources spent, the failed alternatives, and the residual risks that remain.

28.9 Invariants

No generated task, candidate, novelty or diversity metric, evaluator score, transfer result, self-verification trace, archive frontier, or complete campaign record is an admission decision or proof of useful improvement.
Campaign objective, representation, controller, generator policies, evaluator/exposure policy, archive/hazard policy, budgets, stop authority, horizon, and support ceiling are prospective, versioned state.
Every task, candidate, repair, evaluation, timeout, failure, null, cost, reuse event, and terminal outcome keeps exact lineage and disposition or an explicit deletion receipt.
Generated-distribution, eligibility, novelty, difficulty, selection, validation, and archive denominators cannot be rewritten after outcomes.
Evaluator exposure, feedback, shared dependencies, disagreement, and blind spots are candidate and campaign state; same-family verification is not independent qualification.
Failures, nulls, unsafe candidates, timeouts, over-budget arms, rejected transfers, and discarded children remain visible to later selection and reporting.
Resource and hazard budgets use observed consumption and cannot reset through retries, children, forks, or successor campaign identities.
A stop owner must observe the trigger, act inside the required window, and produce a verified effect; naming an owner is insufficient.
Controller, objective, evaluator-authority, bound, recursive-depth, and admission-interface changes are separate improvement proposals.
No candidate or descendant gains field authority, protected-policy change, tool permission, update, deployment, or publication without separate qualification and the existing governor.
Every campaign ends, pauses, or hands off with one honest state, total resource and governance cost, retained residuals and hazards, and accountable owners.

These invariants do not require a search engine to be useless until it is perfect. They require the system to make an honest distinction between a candidate worth studying and an artifact allowed to act. A narrow evaluator can still support narrow research progress. It simply cannot silently inherit the authority of an independent reviewer, a safety case, or a deployment gate.

28.10 Non-obvious consequences

28.10.1 A stepping stone is a typed relation, not a consolation prize

A lower-scoring candidate can matter because a descendant reuses one mechanism, because it covers a different task region, or because its failure falsifies a search assumption. The archive should name which of those roles applies. A generic “interesting” flag cannot justify future selection, and a failed candidate cannot smuggle its full authority envelope into a descendant merely because one component was reused.

28.10.2 Evaluation exposure is part of candidate state

Repeated repair against validation feedback changes what a candidate’s score means. The record must count evaluator calls, exposed cases or summaries, selection rounds, and any model-generated critiques. A nominally held-out set can become training context through iterative search. Qualification therefore uses an unexposed cohort or explicitly narrows the claim when fresh cases are unavailable.

28.10.3 Archive diversity can conflict with safety and cost budgets

Keeping every branch preserves future stepping stones but expands storage, attack surface, review burden, and the number of descendants affected by a later defect. Archive policy needs retention tiers, executable quarantine, digest-bound dependencies, and retirement rules. Diversity is not free, and discarding executable state is different from retaining a public-safe failure receipt.

28.10.4 The campaign controller is itself a capability field

Parent selection, novelty pressure, task generation, budget allocation, and stop policy shape every candidate. If the campaign may modify those mechanisms, that is a higher-order improvement proposal rather than an ordinary child. It requires its own evaluator, authority ceiling, rollback, and comparison baseline. Keeping the controller fixed, as some source systems do, is a scoped design choice rather than evidence that the controller is optimal.

28.10.5 Honest stopping is an improvement result

A campaign that preserves a null result, cost profile, failed assumptions, and reusable negative knowledge may improve the research process even when no candidate is admitted. This does not convert failure into a capability gain. It means campaign value includes avoided bad promotions and better future search, which should be measured separately from candidate performance.

28.11 Failure modes

Objective and measurement failures include objective, representation, and task-generator drift; adaptive evaluator overfitting; contamination; same- family self-verification laundering; novelty/diversity theater; false transfer; and null-result laundering. Each can make search activity or a proxy frontier look like useful discovery without producing consumer value.

Archive and accounting failures include best-only survivorship, discarded- child and timeout erasure, stepping-stone backfill, generator or archive collapse, hidden repair feedback, unreported model or human calls, budget reset through forks, and exclusion of opportunity, evaluator, safety, maintenance, and governance cost.

Authority and safety failures include unauthorized admission, campaign- controller self-modification, ceremonial stop owners, recursive campaign explosion, cross-campaign contamination, dual-use or privacy-bearing payload escape, stale permission, and over-quarantine. Over-retention expands attack surface and descendant impact; over-deletion destroys reproducibility and negative knowledge. These require different owners and cannot be collapsed into a generic alignment residual.

28.12 Strongest objections and surviving residuals

28.12.1 “Independent evaluation ends open-endedness”

Independent qualification need not approve every generated task or interrupt every search iteration. Early evaluators may be fast, adaptive, and tightly coupled to generation. Independence becomes mandatory when a result is used to support a broader claim or cross an authority boundary. The engine can explore freely inside a sandbox while admission remains conservative.

The surviving tradeoff is throughput. Fresh qualification cohorts, distinct models or tools, and human-owned stop paths consume time and money. A mature campaign must report useful discovery per total governance cost, not hide the cost or treat delay as safety evidence.

28.12.2 “The generator will overfit any fixed evaluator eventually”

That objection is correct for sufficiently repeated adaptive exposure. The response is not a permanently secret magical test. It is an evaluation lifecycle: exposure accounting, rotating or generated challenge families, defeater tests, negative controls, independent outcome references where possible, and claim expiry. If no evaluation remains meaningfully fresh, the route narrows to exploratory evidence and blocks promotion.

The residual cannot be eliminated by role labels. Evaluators may share model families, prompts, data, tools, or blind spots with the generator. Dependency disclosure and adversarial evaluator testing remain required.

28.12.3 “Keeping failures creates a dangerous archive”

Full executable retention can indeed preserve vulnerable or dual-use artifacts. Archive honesty does not require universal redistributable storage. It requires a durable receipt for identity, lineage, evaluation, failure, cost, and disposition. High-risk payloads can remain encrypted, access-controlled, redacted, or destroyed under a verified policy while their non-sensitive failure record remains. The residual is that deletion reduces reproducibility; the record must say which replay is no longer possible.

28.12.4 “Benchmark gains are the only objective signal available”

Benchmarks are useful precisely because they are explicit and repeatable. They become misleading when used as a universal objective. Candidate admission can combine task success with latency, cost, authority, cancellation, regression, security, and retained-task constraints. This still does not capture every social or safety property, but it makes omitted objectives visible and gives the governor a reason to refuse broader interpretation.

28.12.5 “Human stop authority will be too slow or ceremonial”

A name in a field is not a stop mechanism. The campaign must test that the holder can observe the relevant state, act within the required window, and produce an effect that the independent observer can verify. Automated budget and sandbox limits should fail closed when human response is unavailable. The remaining residual is capacity: a campaign portfolio can overwhelm its stop owners even when each individual record looks valid.

28.12.6 “A separate campaign layer duplicates recursive self-improvement”

Recursive Self-Improvement governs whether a proposed change may alter the system. This layer governs how proposals, tasks, evaluators, archives, and selection pressure are produced before that decision. Collapsing them lets the search process curate the evidence offered to its own gate. An implementation may share services and schemas, but the two responsibilities and authority boundaries remain distinct.

28.13 Minimum Viable Implementation

Portfolio accounting under stopped search.

A stopped campaign still needs portfolio accounting. Candidate count alone can hide a generator that emits many near-duplicates, while novelty count can hide artifacts that are new only because they are malformed. The archive should record behavioral and lineage distance, evaluator exposure, qualification state, resource cost, hazard class, downstream reuse, and the reason a candidate stopped. Diversity is then a property of a declared representation and workload, not a visual impression.

This accounting changes how stepping stones are valued. A candidate that fails the primary metric may expose an evaluator bug, create a reusable tool, reveal a new failure family, or improve a later generator. Those are typed downstream effects with dates and consumers. They should not be backfilled into the candidate’s original score, because doing so would make every archived failure look successful. The correct record keeps the original disposition and adds a new relation when a later campaign actually consumes the artifact.

Budget exhaustion also has a distributional consequence. A campaign may spend most of its compute exploring one promising family and leave other families under-sampled. Reporting only the best candidate conceals that allocation. The controller should publish family-level attempts, qualification rates, cost, stop reasons, evaluator coverage, and unresolved hazards. A later campaign can then decide whether to continue exploitation, reopen neglected regions, or retire the search space.

The existing stopped results demonstrate why this matters. Their value is not an improvement claim; it is an intact record of threshold non-attainment, resource use, and surviving residuals. A mature engine earns trust when it can terminate without rewriting search activity as progress. Portfolio accounting must also preserve opportunity cost: evaluator time, accelerator time, archive review, and human escalation spent on one family were unavailable to others. A campaign cannot claim broad exploration merely because its aggregate budget was large.

The smallest honest engine is a preregistered, stopped, multi-round adaptive campaign on a public-safe natural program or tool workload. It compares a no-search incumbent, fixed-task search, conventional search, human-authored candidates, and generated-task search under matched tasks, models, tools, data, compute, latency, and human budgets. A separately implemented qualifier sees fresh cases and reports its dependencies and error.

The packet retains every duplicate, malformed, ineligible, failed, null, hazardous, stepping-stone, qualified, timed-out, and over-budget outcome with exact exposure and resource history. It effect-tests one budget stop, one owner-triggered stop, and one quarantined dual-use or authority-seeking candidate. Its best candidate may reach governor review but receives no live authority. A negative result is terminal for the preregistered claim; a follow- up uses a new campaign identity and prospective budget.

The exact current minimum preserves the seven Boolean admission records, seven legacy route theorems, ten admission mutations, the three-seed/four-arm fixed update result, and the three-seed/five-arm post-v2.1 stopped result inside a seven-stage, 81-route campaign-to-governor lifecycle with 91/91 route and 3/3 cross-stage mutations rejected. Two composition theorems derive the complete cycle and a budget-reset/downstream-blocking prefix. The witness emits one governor-review handoff and returns to scoped generation only through protocol-version-2 readmission after material change. Zero of nine eligible post-v2.1 challenger seed-arms met the registered 0.05 gain threshold, the update disposition remains no_change, and support and external-effect authority remain none.

The refinement at experiments/open_ended_improvement_refinement/results/2026-07-16-local.json independently reruns the admission fixture and both stopped campaigns. Exact custody covers campaign, objective, representation, controller, task and candidate policies, generator, evaluator, qualifier, archive, cumulative budget, stop authority, hazard policy, consumer, authority, and version. All 91 route mutations and three cross-stage mutations reject. Two Lean composition theorems execute the complete seven-transition cycle and a budget-reset prefix. A descendant-budget reset prevents archive progression and every later handoff; a missing stop-effect receipt prevents adjudication and handoff; and candidate-authority laundering prevents governor handoff and readmission. This proves authored campaign custody and route consequences only: objectives, legitimacy records, denominators, lineage, evaluator records, costs, hazards, stop effects, residuals, monitors, and invalidations remain trusted inputs. No adaptive generator/evaluator/archive engine, semantic novelty, useful improvement, objective legitimacy, evaluator independence, deployed stop or quarantine, reproduction, or transfer exists.

28.14 Mature Research Target

The mature target is a portfolio of constrained adaptive discovery campaigns across programs, tools, data, curricula, simulations, evaluators, and scientific hypotheses whose objectives remain socially and operationally reviewable. It preserves generated-distribution and candidate lineage, selection denominators, evaluator exposure and dependencies, negative and null search history, semantic novelty, consumer usefulness, dual-use and privacy custody, total resource and opportunity cost, cross-campaign contamination, stop effects, and typed reuse.

Strong current source systems and conventional search must be reproduced under matched budgets. Mechanism ablations, adversarial controls, independent replication, recursive-depth stress, and transfer decide exact claims. The campaign can stop without narrating activity as progress, and candidates enter procedural memory, updates, or capability fields only through separate qualification and governed replacement.

This remains a research target. Neither a complete campaign record nor a frontier score proves evaluator correctness, objective legitimacy, undiscovered failure absence, autonomous scientific discovery, safe self-improvement, general improvement, AGI, or ASI.

The decisive test is whether adaptive generation produces independently useful and transferable novelty after every failed, malformed, hazardous, duplicated, timed-out, and discarded attempt remains in the denominator. More search, activity, archive growth, or evaluator score is not progress when a matched fixed-task or conventional-search baseline achieves the same outcome with less risk and total cost.

28.15 Codex test plan

Test	Purpose	Status
Improvement-campaign admission fixture	Recompute seven declared handoff routes and ten mutations.	implemented; finite Boolean records only
Campaign-to-governor lifecycle refinement	Join scope, generation, archive, evaluation, observed stopping, governor-only handoff, and material-change readmission without cross-promoting null results.	implemented by `python3 scripts/validate_open_ended_improvement_refinement.py`; two composed Lean traces, three inherited suites, all 81 routes, 91/91 route mutations, and 3/3 cross-stage mutations pass, support/effect none; no adaptive-search, novelty, usefulness, evaluator-independence, safety, release, transfer, or support claim
Generator/evaluator dependency and exposure test	Measure adaptive overfitting, contamination, self-verification, shared dependencies, disagreement, and fresh qualification error.	planned; current identity/dependency Booleans measure no behavior
Natural adaptive campaign comparison	Compare no-search, fixed-task, conventional-search, human-authored, and generated-task arms under matched budgets.	planned; no natural campaign engine or candidate ran
Archive denominator and hazard-custody test	Preserve every attempt, disposition, exposure, cost, selection probability, payload tier, deletion receipt, and typed reuse relation.	planned; current failure-history Boolean has no evolving archive
Stop-effect and budget-reset test	Verify owner observation, action latency, pause/quarantine effect, and cumulative budget across retries, children, forks, and successor campaigns.	planned; current Boolean route observes no stop or resource effect
Authority-laundering and controller-change test	Prevent candidate admission and route controller/objective/evaluator/bound changes as separate improvement proposals.	finite no-authority route implemented; real controller change not run
Portfolio, causal-ablation, replication, and transfer test	Measure cross-campaign contamination, shared capacity, mechanism effects, independent reproduction, recursive depth, and unexposed transfer.	planned

28.16 Formalization hooks

Tag	Status	Scope
`lean:open_ended_improvement.campaign.complete_candidate_to_governor_review`	implemented	A complete bounded campaign trace reaches only a governor-review handoff and grants neither admission, support, nor live authority.
`lean:open_ended_improvement.campaign.missing_independent_qualification`	implemented	Missing independent qualification blocks evaluation in the reachable campaign lifecycle.
`lean:open_ended_improvement.campaign.exhausted_budget`	implemented	Missing cumulative budget custody or attempted reset across descendants blocks archive progression.
`lean:open_ended_improvement.campaign.missing_stop_authority`	implemented	Missing stop ownership, observation, or effect receipt blocks campaign adjudication.
`lean:open_ended_improvement.campaign.erased_failure_history`	implemented	Incomplete attempt, failure, null, unsafe, timeout, or archive history blocks archive progression.
`lean:open_ended_improvement.campaign.missing_residual_owner`	implemented	Missing residual ownership blocks campaign adjudication before governor review.
`lean:open_ended_improvement.campaign.authority_laundering`	implemented	Score, candidate, release, support, or self-ratified controller authority cannot be laundered through the campaign lifecycle.

The seven legacy theorems remain finite admission-route consequences. The refinement adds one combined route theorem, one complete-cycle composition theorem, and one cross-stage budget-reset theorem over a reachable seven-stage model, plus an independently implemented 81-route consumer with 91 route and three cross-stage mutations. It binds rather than merges the admission and stopped-campaign evidence lanes. These finite records do not prove their inputs true or their controls effective. None of this generates tasks or candidates, evaluates semantic novelty, establishes objective legitimacy or evaluator independence, executes a deployed stop or quarantine, proves useful improvement, authorizes admission, or establishes reproduction, transfer, safety, readiness, release, support, SOTA, AGI, or ASI.

The refinement is adequate only for exact finite campaign-record custody: stage ordering, denominator and archive predicates, cumulative descendant budget, failed-event preservation, governor handoff, invalidation, and successor-version readmission. It does not establish adaptive generation, semantic novelty, objective legitimacy, useful open-ended improvement, or effective stop and quarantine behavior.

28.17 Source crosswalk

Source ID	Title	Planned use
`rmi`	Ratcheting Modular Intelligence	Mine residual escrow, regression-floor, capability-ratchet, and lifecycle vocabulary for campaign records; conceptual source, not a local improvement result.
`cognitive_loop_closure`	Cognitive Loop Closure	Mine the boundary between generated trajectories and qualified, routable procedures; no local loop-detection or tool-synthesis result is implied.
`benchmaxxing`	Benchmaxxing: The Performance Ratchet	Mine benchmark, holdout, saturation, and anti-Goodhart inputs for evaluator and archive policy; assignment is not a benchmark result.
`ext_poet_2019`	Paired Open-Ended Trailblazer	Comparator for paired environment generation, solution optimization, and transfer in a specified source setting; not evidence of local or safe open-ended improvement.
`ext_funsearch_2024`	Mathematical Discoveries from Program Search with Large Language Models	Comparator for a bounded generator-evaluator-archive loop with a fixed generator and user-provided evaluator; not a general evaluator or autonomous R&D result.
`ext_voyager_2023`	Voyager: An Open-Ended Embodied Agent with Large Language Models	Comparator for curriculum, skill-library, feedback, execution-error, and self-verification mechanisms in Minecraft; not evidence of independent evaluation or local learning.
`ext_darwin_godel_machine_2025`, `ext_adas_2024`	Current empirical self-improvement and architecture-search comparators	Compare branching archives, code-defined candidates, benchmark selection, and transfer claims without treating source-reported gains as safe general improvement or local evidence.

28.17.1 Manifest source assignment reconciliation

These rows keep Open-Ended Improvement Engines’s manifest assignments visible at their recorded review boundary. Passage review does not establish local reproduction, performance, safety, deployment, or support-state movement.

Source	Intake role	Boundary
`learning_compute_topology`	Passage-reviewed comparator: Learning–Compute Topology: Formalizing the Causal Organization of Adaptive Systems. Corben-authored August 2026 research paper and executable preparation package that separates model architecture, learning-process topology, execution topology, and physical compute topology. It contributes adaptive-identity tests; typed evidence, judgement, credit, state, artifact, control, and authority relations; LCT-IR; Learning Causal Normal Form; seven bounded propositions; topology metrics; a semantic compiler firewall; Adaptive Branch–Validate–Integrate; toy and analytical phase diagrams; and an explicit falsification program. The bundled reference implementation passes 11 unit tests, but implements only bounded conformance behavior and does not establish neural-training benefit, causal completeness, universal canonicality, safety, scaling superiority, or ASI.	The formal propositions hold only under their stated finite, explicit-state, interface-sufficiency, information-theoretic, and cut-capacity assumptions. The executable supplement covers a bounded IR/validator/normalizer/compiler/simulator slice; the phase diagrams are toy or analytical, the ABVI topology is proposed, and the novelty matrix is a scoped comparison rather than a global novelty proof. No local implementation, reproduction, performance, safety, deployment, support-state, or ASI result is established by this reconciliation row.

28.18 Post-v2 stopped campaign result

The update-causality program supplies a real but deliberately closed campaign: one base and three challenger axes, three seeds, fixed budgets, validation-only selection, retained best/final checkpoints, negative knowledge, and an exact stop after the registered arms. It demonstrates how candidate lineage, evaluation, failure retention, and rollback can be recorded when parameters actually change.

It does not test task generation, open-ended discovery, evaluator evolution, or autonomous campaign expansion. A fixed four-arm campaign is evidence for a bounded transaction record, not for open-ended improvement quality. The core claim therefore remains argument through no_change.

28.19 Post-v2.1 stopped improvement result

The successor makes the stop discipline more demanding: three seeds, five registered arms, prospective checkpoint authority, a retained-task safety bound, 24 state surfaces, and no post-outcome search. Zero of nine eligible challenger seed-arms reaches the target-gain threshold, and the comparator arm stops early at every seed. Preserving that null result is the campaign’s useful output. It demonstrates how an improvement engine can terminate without promoting a weak candidate, but it supplies no task-generator, evolving archive, open-ended discovery, or autonomous expansion evidence. The chapter-core decision remains no_change.

28.20 Summary

Open-ended improvement needs a control plane upstream of admission. The campaign owns generated distributions, candidates, evaluator exposure, selection denominators, archive and hazard custody, portfolio resources, and effect-tested stopping. It does not own live authority. Generation is permission to propose, not permission to act.

The current record routes and stopped update campaigns establish only narrow bookkeeping boundaries. They do not show that adaptive search produces useful novelty, preserves an honest denominator, resists evaluator gaming, stops in practice, handles dangerous artifacts, beats strong search baselines, or transfers. The next honest step is the preregistered natural adaptive campaign with matched comparators, fresh qualification, full attempt retention, joint cost and hazard accounting, stop effects, causal ablations, replication, and transfer.

Progress therefore means verified useful novelty under stopping and hazard constraints, not merely more generated activity.

28.21 Evidence reconciliation (2026-07-16)

The invariant protocol, field meanings, and inference limits are stated once in Living Book Methodology. This packet contains only the chapter-specific projection; its authoritative per-atom rows are the open-ended-improvement-engines slice of experiments/claim_family_terminal_coverage/results/result.json.

The core remains blocked after full attempt at argument support. The strongest family attempt was Safety-critical lifecycle consumer trace. Its exact boundary is: Finite local fixture consumer only; no authentic deployment, general alignment, evaluator independence, or broad security claim. Across 81 atoms, the terminal ledger records 81 blocked_after_full_attempt.

Chapter-specific field	Value
Family / atom denominator	`CF-02` / 81 atoms
Terminal dispositions	81 `blocked_after_full_attempt`
Core	`open-ended-improvement-engines.core`: `blocked_after_full_attempt` at `argument`
Core attempted / missing lanes	`causal`, `empirical`, `executable`, `formal`, `source-synthesis` / `normative`, `transfer`
Attempted local lanes	`causal`, `empirical`, `executable`, `formal`, `source-synthesis`
Missing or unproved lanes	`normative`, `transfer`
Strongest family bundle	Safety-critical lifecycle consumer trace (`end_to_end`): Ten finite lifecycle receipts spanning bounded effects, denials, residual accounting, and safety-critical state transitions.
Negative controls	five explicit denials with residuals; eight rejecting mutations.
Accepted transitions	none
Maximum inference	Finite local fixture consumer only; no authentic deployment, general alignment, evaluator independence, or broad security claim.
Reproduction / next burden	Replay `scripts/validate_safety_critical_lifecycle_consumer_trace.py` and `scripts/validate_claim_family_terminal_program.py`; fill the named atom-specific lanes under a new prospective protocol.

28.22 Handoff

Open-Ended Improvement Engines supplies bounded candidates and campaign evidence to Autonomous Replication, Proliferation, and Containment whenever an improvement could acquire resources, create descendants, persist, or widen its own deployment. The replication owner must preserve the candidate’s exact identity, authority, lineage, failed attempts, and residuals while denying real-world replication authority by default. Improvement success cannot stand in for containment, descendant control, or recall.