78 Adjudicated Persistence and the Adaptive Commit Boundary

78.1 Chapter status

Field	Value
Chapter ID	`adjudicated-persistence-and-the-adaptive-commit-boundary`
Part	Part IV - Evidence, Implementation, and the Living Book
Status	conceptual
Manuscript maturity	source-complete integrated argument
Last updated	2026-08-11
Primary source	Corben Sorenson’s version 1.0 Adjudicated Persistence paper and complete package
Claim label	Design rationale
Evidence level	argument
Source loading state	The complete paper package has been passage-reviewed and mined. Its external citations have not been inherited as evidence.
Test state	Five proof targets and five experimental suites are planned; none has been implemented or run.

78.2 Drafting guardrail

This chapter proposes a control boundary. It does not report that an Adaptive Commit Boundary, placement compiler, qualification service, decompiler, or LocusBench implementation exists. The source paper supplies a formal vocabulary, bounded propositions and conjectures, architecture, algorithms, and benchmark design. Those are useful specifications, not empirical results.

The chapter also does not claim that every experience should become a lesson, that every lesson should persist, or that a central controller can identify a globally optimal storage location. Its narrower claim is that durable adaptation is consequential enough to require an explicit, challengeable decision whose evidence, authority, realization, qualification, monitoring, and revocation remain visible.

78.3 Human Reading Path

Concrete lens. Two tax-support agents can reach the same correct answer after failures that require different durable repairs.

Imagine two support agents each apply an outdated tax rule, catch the error before replying, and eventually produce a correct answer. Their visible outcomes match: a near miss followed by recovery.

But one agent relied on stale jurisdictional memory. Its repair may be a dated knowledge update with a freshness rule. The other had correct facts but passed them through a tool with the wrong field mapping. Its repair may be a test, typed wrapper, and blocked tool version.

Updating both models’ weights would be broad and hard to inspect. Writing the same memory to both would leave the tool defect untouched. Making every near miss an organization-wide rule would bury people under process.

The decisive question is not only what the system learned, but what durable thing, if any, should change. A lesson might enter context, memory, a test, tool, router, parameter, evaluator, workflow, environment, or institution. Each locus changes reach, cost, authority, reversibility, and failure.

The Adaptive Commit Boundary stands before that choice becomes causal. It preserves the event and proposed lesson, compares possible loci, asks whether evidence justifies their commitment, and records qualification, monitoring, invalidation, and recovery. The aim is the least-committing durable intervention that can solve the problem without concealing uncertainty or making reversal impossible.

78.4 Problem

An adaptive system changes because experience reaches into its future. That reach may be obvious, as in fine-tuning a model, or quiet, as in writing a memory, generating a reusable tool, changing a router, retaining a benchmark, editing an evaluator, or teaching an organization a new routine. All of these actions can alter later behavior. They are persistence decisions even when no gradient is computed.

The surfaces are not interchangeable.

Persistence surface	Typical strength	Characteristic benefit	Characteristic risk
Active context	local and short-lived	cheap, inspectable experiment	disappears or contaminates the current task
Durable memory	retrievable and revisable	explicit knowledge with provenance	stale or overgeneralized retrieval
Procedure or tool	repeatable and binding	deterministic reuse and testing	brittle automation or authority laundering
Semantic model	shared conceptual structure	consistent interpretation across tasks	ontology lock-in and hidden coupling
Orchestration or routing	cross-component behavior	system-level coordination	widespread effects from a local mistake
Parametric update	broad latent influence	compact generalization	opaque scope, forgetting, and difficult rollback
Evaluator or benchmark	changes what is rewarded	durable selection pressure	Goodhart pressure and self-confirmation
Environment or institution	changes available action	robust external constraint	legitimacy, rights, and reversal burdens

The same lesson may need one surface, several surfaces, or none. A temporary context guard can be enough while evidence is weak. A high-consequence defect may require a test, tool repair, monitor, and institutional review together. A successful incident response may produce useful observations but no stable causal account, making durable change premature.

Without an explicit boundary, implementation convenience decides. The module that notices a problem writes to the state it can reach. A memory subsystem writes memory. A trainer updates weights. A coding agent creates a test or tool. A manager creates a policy. This is not neutral. Access to a persistence surface becomes implicit authority to shape future behavior.

78.5 Why existing approaches are insufficient

Most learning pipelines answer a downstream question: how to update a chosen object. Reinforcement learning changes a policy. Continual learning changes a model or replay distribution. Memory systems change retrieval state. Compilers turn a specification into an artifact. Cognitive loop closure turns repeated work into procedures and tools. Organizational learning changes routines and responsibilities.

Each can be the right mechanism after placement has been decided. None alone answers the prior cross-surface question.

Three substitutions are especially dangerous.

First, outcome is not placement. The same observed success or failure can arise from different latent defects, and the same latent defect can admit different repairs under different authority and recovery constraints. The tax example above is a simple collision: equal outcome, unequal cause, unequal proper persistence.

Second, causal attribution is not authorization. Even a well-supported cause does not determine who may alter weights, memory, tools, evaluators, or institutional rules. Nor does it show that the alteration has passed the tests required for use.

Third, construction is not qualification. A system can successfully write a memory, compile a tool, or train a checkpoint that is unfit for its intended scope. Treating artifact existence as readiness collapses engineering and governance into a single irreversible step.

78.6 Core Claim

[adjudicated-persistence-and-the-adaptive-commit-boundary.core, label: Design rationale, support: argument] Every transition from experience to durable causal influence should cross an Adaptive Commit Boundary as an authority-bearing adaptation transaction that keeps the experience record, lesson hypothesis, persistence disposition, concrete realization, qualification lease, and authority grant distinct; selects the least-commitment admissible locus portfolio under evidence, authority, observability, recovery, cost, and descendant obligations; and preserves denial, uncertainty, deoptimization, invalidation, and revocation paths.

This claim adds a layer between learning signals and update mechanisms. It does not replace either. Regret, incidents, evaluations, and human feedback can propose lessons. Memory, training, compilation, and institutional processes can realize them. Readiness and operations can qualify and monitor them. The Adaptive Commit Boundary owns the decision that joins those stages without allowing one stage to impersonate another.

78.7 Mechanism

78.7.1 Six objects that must not collapse

The transaction uses six records.

Experience record. What happened, under which state and authority, with what provenance, outcome, uncertainty, affected parties, and alternatives.
Lesson hypothesis. A falsifiable statement about what should differ in future behavior and why. It is not yet a storage instruction.
Persistence disposition. ADMIT, DENY, or UNKNOWN, plus a selected locus portfolio, commitment profile, evidence obligations, and residuals.
Realization. The exact memory, procedure, tool, route, model delta, evaluator change, environmental constraint, or institutional artifact.
Qualification lease. The scope, tests, monitor conditions, expiry, invalidators, and evidence under which that realization may be relied upon.
Authority grant. Who may create, activate, broaden, suspend, revoke, or replace the realization and under what contestability and remedy rules.

These objects may reference one another, but they cannot share identity. A lesson can be sound while one realization is defective. A realization can be technically correct but unauthorized. A qualified tool can later become stale when its dependency changes. An authorized experiment can remain unqualified for production. Separate records make these states expressible.

78.7.2 Reusable knowledge needs a separate admission round

A valid candidate abstraction is not yet a beneficial persistence decision. The candidate may compress prior solutions, pass independent semantic checks, and appear in later traces while adding no marginal value. It can also remain a useful archived artifact even when a removal or placebo intervention shows that the default library should not expose it. Validation status and transfer disposition therefore remain separate fields.

The test round must consume a frozen knowledge version. Extraction, development tuning, and independent validation happen before that freeze; held-out tasks then compare presence, absence, placebos, removal, and any library-search interaction. An accepted candidate can enter only a subsequent version. Promoting it into the evaluated version would let test information alter the intervention and turn the benchmark into part of the training loop.

This is the Adaptive Commit Boundary applied to reusable symbolic knowledge: the lesson hypothesis concerns future task benefit, the realization is the exact library entry and retrieval surface, the disposition records the causal test, and qualification names its distribution, search, verifier, budget, expiry, and removal conditions. From Compression to Forward Transfer proposes this lifecycle but reports no admission transaction or measured effect.

flowchart LR
  E["Experience record<br/>provenance, context, outcome"] --> L{"Learning eligible?"}
  L -- "no" --> R["Retain evidence or close with reason"]
  L -- "unknown" --> S["Slow path and additional evidence"]
  L -- "yes" --> H["Lesson hypothesis<br/>future behavior, cause, uncertainty"]
  H --> P["Candidate locus portfolios<br/>memory, tool, policy, weights, institution"]
  P --> C["Commitment profiles<br/>scope, binding, persistence, coupling, authority, reversibility, descendants"]
  C --> A{"Evidence and authority sufficient?"}
  A -- "deny" --> D["Deny with residuals"]
  A -- "unknown" --> S
  A -- "admit" --> T["Transactional realization"]
  T --> Q{"Independent qualification"}
  Q -- "fail" --> X["Rollback or deoptimize"]
  Q -- "pass" --> U["Scoped use under lease"]
  U --> M["Monitor material change and descendants"]
  M -->|"expiry, contradiction, drift"| X

How to read the adaptive commit boundary: Evidence moves forward only through typed decisions. Failure and uncertainty do not disappear; they route to denial, additional inquiry, rollback, or residual custody. The final use state remains leased rather than becoming timeless truth.

78.7.3 A worked two-case transaction

Return to the tax near misses. The purpose of the example is not to claim that the boundary can already diagnose either case. It is to show which distinctions an implementation must preserve before a diagnosis can justify durable change.

In Case A, the agent’s durable memory says that a filing threshold is $10,000. The rule changed on a known effective date, but the retrieval packet does not carry temporal validity and therefore presents the old value as current. The agent initially uses it, then a fresh authoritative lookup causes the answer to be corrected. In Case B, the agent retrieves the current threshold correctly. A tool adapter maps gross_amount into a field that expects taxable_amount, producing the same erroneous intermediate answer and the same final correction after manual inspection.

An outcome-only record makes the cases look equivalent:

Field	Case A	Case B
Initial answer	wrong	wrong
Error caught before external effect	yes	yes
Final answer	correct	correct
Operator-visible consequence	delay and rework	delay and rework

The complete experience records do not. Case A names the retrieved semantic object, its source, original and effective dates, memory snapshot, query, consumer, conflicting observation, and correction path. Case B names the tool schema, adapter version, argument binding, execution trace, output, reviewer, and corrected mapping. The lesson hypotheses are correspondingly different:

A: a time-sensitive rule must not be admitted into decision context without a freshness comparison against its effective interval;
B: this adapter must not execute unless source and destination field semantics match the reviewed tool contract.

The candidate portfolios expose the placement decision. For Case A, candidates include a transient reminder, a versioned memory correction, a retrieval-time freshness guard, a semantic-model migration, or some combination. A broad parameter update is a possible candidate only if the system can state what it would solve that those narrower interventions cannot. For Case B, candidates include a prompt warning, a regression test, a typed adapter repair, a tool version quarantine, or a routing rule that forces review. Writing a memory that merely says “be careful with tax fields” is cheap but does not bind the failing interface.

Suppose the adjudicator selects a versioned memory correction plus freshness guard for A, and a typed wrapper plus regression test for B. Each disposition records why rejected alternatives were insufficient or overcommitting. The A transaction binds the exact memory object and retrieval policy, while the B transaction binds the tool schema, adapter, test, and quarantined predecessor. Neither transaction is complete merely because its files were written.

Qualification then checks different obligations. The memory repair must retrieve the current rule under the relevant jurisdiction and date, preserve the historical rule where it remains valid, deny or escalate an undated query, and invalidate derived summaries that presented the old value as timeless. The tool repair must reject the old argument binding, accept reviewed valid bindings, preserve unrelated calls, expose the invoked schema version, and keep the predecessor unavailable except through an explicit recovery route.

The two leases also have different invalidators. A new rule, source withdrawal, jurisdiction change, or failed freshness lookup can invalidate A. A tool-schema change, adapter change, failed regression, or newly discovered affected call can invalidate B. If a downstream procedure copied the old threshold or a generated tool embedded the bad field mapping, revoking only the immediate artifact is incomplete. Those descendants must be repaired, deoptimized, compensated, requalified, or recorded as unresolved residuals.

This example illustrates the chapter’s information requirement. The boundary does not need omniscience before it can refuse a bad shortcut. It needs enough identity, provenance, causal alternatives, authority, and lifecycle state to distinguish materially different candidates. When that information is absent, UNKNOWN is the correct disposition. The system may preserve the incident, run bounded probes, or install a temporary guard, but it may not relabel uncertainty as permission for a durable broad update.

78.7.4 Cross-surface adaptation assignment

The central decision can be stated as a constrained assignment problem. Given a lesson hypothesis, the system proposes candidate portfolios across available persistence surfaces. It estimates whether each portfolio can achieve the declared behavioral requirement and records the obligations created by doing so.

A portfolio is not judged only by expected task quality. It also carries a commitment profile:

scope: which users, tasks, systems, environments, and times it can affect;
binding strength: advisory, default, mandatory, or physically enforced;
persistence: transient, leased, durable, or inherited by descendants;
coupling: how many components and assumptions must change together;
authority: which powers are needed to create and use it;
irreversibility: how completely prior state and external effects can be restored; and
descendant reach: which derived artifacts, behaviors, and institutions can inherit it.

The profile is multidimensional because a single scalar hides tradeoffs. A memory entry may be long-lived but easy to delete. A short-lived tool execution may have irreversible external effects. An institutional policy may contain no code yet shape thousands of decisions. A parameter update may be compact in bytes while broad in behavioral scope.

78.7.4.1 Evidence-Commitment Matching

Stronger commitments need stronger evidence. This is not a claim that one universal numerical formula can price every adaptation. It is a monotonic governance rule: if candidate A is at least as broad, binding, persistent, coupled, authority-bearing, irreversible, and descendant-reaching as candidate B, A cannot receive a weaker evidence obligation merely because A is familiar or easy to deploy.

Evidence obligations can include independent causal review, affected-party input, stronger holdouts, adversarial tests, longer monitoring, formal invariants, rollback exercises, or higher approval authority. Rights and legitimacy constraints are not reducible to performance confidence.

78.7.4.2 Minimum Sufficient Persistence

The preferred portfolio is the least-commitment admissible one that can meet the declared behavior and reliability requirement. This Minimum Sufficient Persistence rule resists a common bias: making every discovered improvement as general and automatic as possible.

Minimal does not mean weak. A narrow memory patch that cannot prevent a high-consequence recurring tool error is insufficient. Conversely, broad fine-tuning is not justified when a typed wrapper and test can address the failure more transparently. Portfolio selection is conditional on current evidence and can change as that evidence changes.

78.7.5 Transactional realization and qualification

An admitted disposition compiles into one or more locus-specific artifacts. The transaction binds the exact experience and lesson identities, compiler and policy versions, target loci, pre-state, proposed delta, authority, tests, monitor, rollback, expiry, descendants, and residuals. Partial success is not silently committed. A multi-surface repair that updates a tool but fails to install its monitor either rolls back or records a narrower explicitly re-adjudicated result.

Qualification follows realization. It asks whether the concrete artifact is fit for a named use under current conditions. Its lease can be invalidated by new evidence, environment change, dependency drift, implementation change, authority change, threat-model change, or expiry. Requalification is a new decision, not a timestamp edit.

This distinction matters for learned updates. A trained checkpoint is a realization. A successful loss curve is part of its construction record. The checkpoint is not qualified until target evaluations, retained-capability tests, reward-hacking probes, authority checks, rollback state, and monitor conditions pass. The same logic applies to memory and institutional policy, even when their tests differ.

78.7.6 Guarded compilation, deoptimization, and deliberation reserve

Some adaptations should begin as guards around a slower trusted path. A compiled tool can handle familiar cases while novelty, contradiction, low confidence, lease expiry, or missing authority routes back to deliberation. This makes UNKNOWN operational rather than rhetorical.

The slower path must remain reachable and sufficiently maintained. Otherwise optimization destroys the system’s ability to discover that the optimization is wrong. The chapter calls this retained capacity deliberation reserve. Its records include rejected alternatives, counterfactual branches, source evidence, and enough executable or human procedure to re-examine the lesson.

Deoptimization reverses compilation without pretending that time ran backward. It can disable a tool, unmount a memory, restore a route, quarantine a checkpoint, or suspend a policy while preserving the event, downstream use, and external effects. If a descendant cannot be removed, compensated, or requalified, it remains an owned residual.

78.7.7 Carrying cost and adaptation debt

Persistence creates a continuing obligation. Storage is often the smallest part. Durable artifacts need validation, monitoring, security review, dependency maintenance, requalification, rollback capacity, human attention, and eventual retirement. They can also narrow future search by making one interpretation or procedure feel canonical.

Adaptation debt is the accumulated burden of commitments whose current benefit, evidence, ownership, or reversibility no longer justifies their maintenance. The concept is intentionally broader than technical debt. A high-performing rule can still be debt if no one can explain its scope, test its invalidators, or remove its descendants. Resource Economics owns the measurement of those costs; the commit boundary requires them to be visible in the choice.

78.7.8 Governing the governor

A placement compiler may eventually learn which surfaces work well. That creates a reflexive risk. If the compiler can generate its own evidence, judge its own evaluator, authorize its own update, and broaden its own reach, the boundary becomes a ceremonial wrapper around self-ratification.

Meta-compilation therefore requires independent-enough evidence, evaluation, and promotion authority. The exact separation may vary by consequence and institution, but no candidate should control all three. Changes to the compiler’s objective, locus vocabulary, commitment estimator, admissibility rule, or exception path are constitutional changes to the adaptation system, not ordinary placement decisions.

78.7.9 Review, appeal, and supersession

Persistence decisions can affect users who never observed the triggering experience. A routing rule may change which specialist serves them; a memory entry may alter later recommendations; an institutional procedure may shift work or liability. The transaction therefore records affected scope and the route by which a disposition can be challenged. The required standing and review process depend on the locus and consequence. A code-owner review may be enough for a reversible internal test, while a policy affecting employment, privacy, or access requires the relevant institutional and affected-party process.

Appeal does not edit history. It appends a competing interpretation, evidence, and requested disposition. An upheld appeal can suspend use, narrow scope, replace the realization, demand requalification, or revoke the commitment. Rejected appeals remain attached to the decision because later evidence may make their objection material. This preserves dissent without pretending that all interpretations are equally supported.

Supersession is similarly explicit. A new realization does not silently erase the old one or inherit its lease. It names the predecessor, states what changed, replays the applicable evidence and authority obligations, and identifies which descendants migrate. Historical reconstruction remains possible, but current consumers receive only a specifically qualified version.

78.8 Interfaces

The chapter coordinates existing owners without replacing them.

Evidence States decides what a claim is supported by; the commit boundary cannot promote evidence by choosing a locus.
Regret, incident response, evaluation, and Assurance-Shift Learning can emit eligible lesson candidates; they do not directly authorize updates.
Memory, Cognitive Compilation, Policy Optimization, Data Engines, and organizational governance realize accepted dispositions in their own media.
Artifact Graphs preserves lineage and descendants.
Readiness qualifies realizations for use, and Governed Operations monitors them after admission.
Resource Economics accounts for carrying cost and opportunity cost.
Human and institutional governance retains legitimacy, rights, appeal, and remedy authority where people are affected.

This division is what makes the new layer useful. It closes a missing decision without creating a universal adaptation department or transferring another owner’s evidence, support, or authority.

78.9 Invariants

The commit boundary must preserve distinctions that convenience continually pressures an implementation to erase. An experience is not a lesson; a lesson is not a persistence instruction; a disposition is not its realization; a realized artifact is not qualified for use; and qualification does not create authority. UNKNOWN remains different from ADMIT, and temporary containment remains different from durable learning. Every accepted realization retains the source experience, lesson identity, chosen locus or portfolio, rejected alternatives, evidence class, authority basis, qualification lease, monitor, invalidators, descendants, carrying cost, and recovery route.

Those distinctions must survive composition. A downstream memory, tool, router, model, evaluator, environment, or institution cannot inherit broader support or authority merely because it consumes an admitted realization. Revocation and material change invalidate dependent leases according to their recorded scope; rollback accounts for known descendants and residuals rather than equating deletion of the parent with complete reversal. The placement policy cannot approve an expansion of its own authority, suppress an appeal, or count its own successful write as evidence that the placement was correct.

78.10 Failure modes

The characteristic failures are structural because each one collapses a decision that must remain inspectable. They may produce an apparently useful update while concealing whether the lesson was supported, the chosen locus was the least committing option, affected authority was present, the artifact was qualified, or recovery remains possible. A successful task therefore does not clear these failures by itself:

one-event overlearning: a salient incident becomes a broad rule before recurrence, alternatives, or affected contexts are understood;
premature locus selection: a subsystem chooses the surface it owns before comparing alternatives;
commitment laundering: a memory write, checkpoint, tool, or policy is described as “learning” to avoid its evidence and authority obligations;
unknown-to-admit collapse: missing evidence is treated as permission;
qualification collapse: successful construction is treated as fitness for use;
hidden descendants: derived artifacts or institutional reliance survive apparent rollback;
observability loss: compression removes the traces needed to diagnose a bad adaptation;
adaptation debt: accumulated rules and artifacts consume capacity and constrain future learning without active justification;
self-ratifying compilation: the placement mechanism approves changes that expand its own power; and
institutional persistence without legitimacy: technically effective rules affect people without contestability, remedy, or rightful authority.

78.11 Strongest Alternatives and Objections

The first alternative is to let each subsystem govern its own updates. That is often simpler and may be sufficient in a small, single-owner system. The objection becomes decisive when a lesson can plausibly belong in more than one surface or when its realization creates cross-owner descendants. Local update quality does not answer whether another locus would have been narrower, more observable, or easier to revoke. LocusBench must therefore include competent fixed-locus policies; if they match the boundary at lower total cost, the cross-surface compiler has not earned its complexity.

The second alternative is to centralize all durable state in an explicit memory or knowledge graph and keep the model fixed. This improves visibility, but not every behavior is reducible to retrieval. Tool contracts, evaluator definitions, routing policies, motor skills, and institutional authorities have different semantics and failure modes. The chapter does not assume that all surfaces are always needed. It requires the proposal set to include the simpler memory-only design and to justify every additional locus.

The third objection is epistemic: reliable causal attribution and cross-surface comparison may be infeasible in open systems. The chapter accepts that possibility. UNKNOWN, temporary guards, bounded probes, and retained deliberation are first-class outcomes. The architecture is useful only if it can abstain without converting every ambiguity into permanent process. An implementation that routinely produces elaborate packets but cannot improve placement, recovery, or refusal decisions should be retired.

The fourth objection is political. A technical boundary cannot determine legitimate institutional authority, affected-party standing, or acceptable tradeoffs. Correct. The control plane can preserve identities, evidence, obligations, and enforcement hooks; it cannot manufacture legitimacy. Its role is to prevent a technical write path from silently substituting for the human or legal decision and to make unresolved authority block the relevant commit.

The final objection is reflexive overhead. Monitoring every descendant and maintaining deoptimization paths may cost more than the adaptation is worth. That is why carrying cost belongs in admission and why Minimum Sufficient Persistence permits denial. The boundary is not justified by the number of records it creates. It is justified only when those records prevent or expose material placement, authority, qualification, or recovery failures at an acceptable whole-lifecycle cost.

78.12 Minimum Viable Implementation

The first implementation should be a sandboxed software-maintenance system, not a self-modifying foundation model. Exact repository snapshots provide experience records. Seeded defects create matched outcomes caused by different latent failures. Candidate realizations include a temporary context patch, memory entry, procedural checklist, generated test, typed tool wrapper, routing rule, and bounded policy change.

Before results are visible, the campaign freezes:

lesson-eligibility and causal-review rules;
candidate locus portfolios and commitment-profile fields;
ADMIT, DENY, and UNKNOWN policies;
strong fixed-locus and outcome-only baselines;
held-out defect families and placement-regret measures;
mutation, stale-lease, guard-bypass, and descendant-escape tests;
useful task quality, latency, compute, human burden, and carrying-cost denominators; and
rollback, deoptimization, and residual-closure acceptance rules.

This is the first tier of the proposed LocusBench program. A useful result would show not merely that the compiler chooses different loci, but that its choices improve held-out outcomes or recovery relative to simpler baselines without unacceptable governance cost. Null and negative results remain valuable: they can show that the boundary adds overhead without better placement, that commitment estimates are unstable, or that a simpler fixed policy is sufficient.

78.13 Mature Research Target

The mature target is a distributed persistence control plane spanning model parameters, memory, tools, orchestration, evaluators, environments, multi-agent populations, and institutions. It would estimate whole-lifecycle carrying cost and adaptation debt; preserve rights and authority constraints; select portfolios under uncertainty; maintain deliberation reserve; re-expand compiled structures when conditions change; and audit descendants and external reliance across organizations.

Its placement policy could itself improve, but only inside the non-self- ratifying loop above. Strong evidence would require prospective natural workloads, matched per-locus and cross-locus baselines, adversarial manipulation, complete attempt and cost denominators, independent evaluators, independent reproduction, and heterogeneous transfer. No current artifact in this repository meets that bar.

At that scale, the control plane would need to reason over interactions rather than isolated writes. A narrow tool repair might reduce immediate failures yet increase dependence on one vendor; a parameter update might improve average performance while making deletion, explanation, or jurisdictional separation harder; an institutional rule might improve consistency while creating adaptation debt and excluding legitimate exceptions. Mature adjudication would compare these portfolios prospectively, retain dissent and rejected alternatives, and revisit the decision when costs, rights, environments, or downstream reliance materially change. Its success criterion would not be the number of lessons committed. It would be better held-out outcomes and recovery with lower unjustified commitment than competent fixed-locus and human-reviewed alternatives.

78.14 Codex test plan

Test	Purpose	Status
`lean:persistence.outcome_only_placement_nonidentifiability`	Show in a finite model that equal observed outcomes can correspond to different admissible persistence loci.	planned; not implemented
`lean:persistence.commitment_dominance_evidence_monotonicity`	Preserve the rule that a commitment-dominating candidate cannot receive a weaker evidence class under a monotone policy.	planned; not implemented
`lean:persistence.transaction_requires_distinct_qualification_authority`	Reject realized artifacts that lack either qualification or authority.	planned; not implemented
`lean:persistence.meta_compiler_non_self_ratification`	Reject a compiler change when the candidate controls its sole evidence, evaluation, and promotion authority.	planned; not implemented
`lean:persistence.descendant_invalidation_residual_closure`	Require every modeled descendant to be removed, deoptimized, compensated, or retained as an explicit residual.	planned; not implemented
Adaptive transaction mutation suite	Mutate identity, provenance, disposition, locus, commitment, qualification, authority, expiry, rollback, descendants, and residuals.	planned; not run
LocusBench Tier-1 matched-collision suite	Compare cross-surface assignment with outcome-only and fixed-locus baselines.	planned; not run
Deliberation-reserve and deoptimization test	Check that novelty and invalidation can still reach a maintained slower path.	planned; not run

Finite Lean results would prove only consequences of the encoded predicates. They would not prove that the inventory is complete, that a lesson is true, that a placement improves behavior, that a monitor detects real drift, or that revocation reverses external effects.

78.15 Source crosswalk

Source	Contribution	Boundary
`adjudicated_persistence`	Primary source for the Adaptive Commit Boundary, six-object separation, Cross-Surface Adaptation Assignment, commitment profiles, Evidence-Commitment Matching, Minimum Sufficient Persistence, guarded compilation, qualification leases, adaptation debt, non-self-ratifying meta-compilation, and LocusBench.	Conceptual author paper and experimental specification; no implementation, benchmark result, checked proof, safety result, or support transition.
`assurance_shift_learning`	Learning eligibility, boundary evidence, evaluator-first repair, least-invasive repair, and fast-versus-slow adaptation distinctions.	No implemented allocator, natural campaign, or empirical placement result.
`regret_engine`	Typed experience residuals, causal alternatives, and learning-eligibility signals.	A regret packet can propose a lesson but cannot establish placement, qualification, authority, or support.
`cognitive_loop_closure`	Repeated cognition becoming verified procedures and tools as one persistence family.	Procedural compilation is one candidate locus, not evidence that a lesson belongs there.
`deterministic_capability_compilation`	Immutable input identity, compilation receipts, artifact lineage, fallback, revocation, and rollback concepts.	Deterministic compilation does not establish lesson validity, placement quality, qualification, or authority.
`platonic_world_model`	Explicit state, model-versus-reality residuals, semantic versions, and counterfactual distinctions.	World-model framing does not identify the correct locus or validate causal attribution.
`viea`	Typed commands, roles, verification, failure behavior, and escalation for adaptation transactions.	Command structure alone does not validate persistence decisions or outcomes.
`talos`	Bounded execution authority, audit, intervention, and policy-controlled realization interfaces.	Architecture lineage does not establish placement, qualification, revocation efficacy, or deployed safety.
`ext_bourtoule_machine_unlearning_2021`	External comparator for training-time sharding, isolation, slicing, aggregation, and bounded retraining paths after deletion requests.	Source-reported SISA results do not validate cross-surface placement, verified removal, complete descendant erasure, privacy, organizational revocation, or the Adaptive Commit Boundary.

The source package includes 42 references, including recent work. This intake does not import those references as independent support. Each external result must receive its own source note and chapter-specific mapping before it can move a claim.

78.15.1 Manifest source assignment reconciliation

These rows keep Adjudicated Persistence and the Adaptive Commit Boundary’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
`forward_transfer_program_synthesis`	Passage-reviewed comparator: From Compression to Forward Transfer: Evaluating Reusable Knowledge in Program Synthesis. Adds a separate later-version admission step after independent validation and frozen evaluation, preserving valid-but-transfer-negative artifacts without routing them as beneficial defaults.	No admission service, knowledge-version transition, removal test, benchmark result, or support movement was implemented. No local implementation, reproduction, performance, safety, deployment, support-state, or ASI result is established by this reconciliation row.

78.16 Current Evidence and Non-Claims

The complete manuscript, DOCX, README, five figures, and package manifest were reviewed. All eight package hashes match. This establishes source custody and supports faithful exposition of the author’s framework.

It does not establish that:

the Adaptive Commit Boundary or LocusBench exists as running software;
cross-surface assignment outperforms fixed-locus, human, or simpler rules;
commitment profiles are complete, measurable, or comparable across every technical and institutional surface;
Minimum Sufficient Persistence is globally optimal;
guarded compilation preserves enough deliberation reserve in natural use;
descendant-aware revocation is complete or externally effective;
meta-compilation is safe, aligned, legitimate, or manipulation-resistant;
the architecture improves useful throughput or lowers total adaptation cost; or
the book has proved safety, SOTA performance, AGI, or ASI.

The chapter remains at argument until an accepted evidence transition says otherwise.

78.17 Summary

Adaptive systems do more than update weights. They turn experience into memories, tools, procedures, routes, evaluators, environments, and institutions. The placement decision determines how far a lesson reaches, how strongly it binds, who may invoke it, what it costs to maintain, and whether it can be undone.

The Adaptive Commit Boundary makes that decision explicit. It separates the event from the lesson, the lesson from the persistence disposition, the disposition from the realization, the realization from qualification, and qualification from authority. It prefers the least-commitment admissible portfolio, preserves uncertainty and a slower path, and requires invalidation, deoptimization, descendant closure, and independent governance of the placement mechanism itself.

The architecture is currently a disciplined proposal. Its next honest step is a small prospective LocusBench implementation with strong simple baselines, complete denominators, and negative-result retention.

78.18 Handoff

Governed Operations supplies incidents, residuals, recovery evidence, and bounded learning proposals after emergency authority has expired. This chapter decides whether such an experience is eligible to become durable influence and which persistence surfaces may carry it.

Policy Optimization and Learning from Feedback receives only dispositions that select a policy or parametric locus, with their evidence boundary, commitment profile, authority ceiling, qualification obligations, monitor conditions, rollback state, descendants, expiry, and non-claims intact. A reward signal or eligible lesson is not permission to train, and a trained checkpoint is not a qualified policy.