Agentic Institutions

Organisational Memory, Correction and Continuity Across Successive AI Agents

Luuk Joseph Soons · CounterProof Research
Conceptual preprint v0.3 · 28 October 2026
DOI: 10.5281/zenodo.23727633
A conceptual framework and research agenda. No experimental results from the proposed study are reported. Not peer reviewed.

Abstract

An organisation can keep every conversation with its AI assistants and still repeat a mistake it has already corrected. The problem is not simply whether information survives. It is whether later participants can recover its grounds, recognise its limits, challenge it when circumstances change, and act on it appropriately. This paper proposes Agentic Institutions as a research programme within applied and social epistemology, concerned with the arrangements through which people, models, records and tools sustain work across successive sessions. It distinguishes institutional continuity from justified learning, and the preservation of a claim from the preservation of reasons to accept it. Organisational memory, distributed cognition and research on routines supply established foundations; recent agent-memory systems provide close technical predecessors. A bounded examination of CounterProof's library and retrieval materials illustrates a specific difficulty: a correction may be recorded while an older summary retains the withdrawn claim. The observation motivates inquiry but supplies no estimate of system-wide performance. We propose a controlled study of repeated handovers, including valid corrections, mistaken objections and changing conditions. The proposed main comparison tests incremental value, at matched resources, over one comparison condition that combines retrieval, simple archive repair and generic source checking. Its implementation and final protocol remain to be specified. Outcomes include later decision quality, recurrence of corrected errors, false reversals, appropriate uncertainty and human effort. The broader aim is to investigate how institutions working with AI can accumulate useful knowledge while remaining capable of revising their own methods.

Keywords: agentic institutions; organisational memory; distributed cognition; institutional learning; epistemic accountability; AI agents; correction; model succession

1. What survives when the agent leaves?

A researcher discovers that an analysis used the wrong denominator. The estimate is withdrawn, the explanation is saved, and the team moves on. A week later, another assistant retrieves the original summary and repeats the estimate. Nothing was lost from the archive. Something was lost from the work.

This opening case is hypothetical. It identifies a problem that storage alone cannot solve. A successor needs to know more than what a predecessor wrote. It needs access to the relevant evidence, the history of the correction, and the limits within which the revised conclusion applies. It must also be able to question the correction. The most recent document may be mistaken too.

Human organisations face this problem whenever people leave, responsibilities change or experience becomes routine. AI does not introduce a new standard of truth. It changes some of the conditions under which continuity must be maintained. A fresh session can be created in seconds. Its instructions and retrieved context can be assembled automatically. The underlying model, tools or access permissions can change while the organisation's documents retain the same names. Many apparently distinct participants may receive the same inherited account.

Our central question is therefore:

Under what conditions can an organisation preserve justified corrections across successive AI-assisted episodes without treating obsolete claims as current grounds or surrendering judgement to inherited authority?

The first application is software engineering, including decisions about what to build, design review, implementation, testing and the handling of failures. The same question may arise in research, administration and other knowledge work, but success in one setting would not establish success in all of them.

This paper makes three proposals within an established body of research on organisational learning and agent memory. First, it identifies distinctions needed to study continuity without mistaking persistence for improvement. Second, it describes a candidate procedure for carrying evidence and corrections across handovers. Third, it outlines a comparison intended to be capable, once its manipulation check and decision thresholds are fixed, of showing that the procedure adds no worthwhile benefit. The local library examination provides motivation and a reproducible documentary question, not a completed evaluation of institutional learning.

2. Institutions, organisations and agents

2.1 The object of inquiry

North distinguishes institutions, understood as humanly devised constraints that structure interaction, from organisations pursuing purposes within those arrangements [1]. We use this distinction analytically. CounterProof is an organisation. Its allocation of responsibilities, rules of evidence, permissions and revision procedures are institutional arrangements. Ordinary speech may call the organisation itself an institution; the mechanisms studied here require the narrower distinction.

For this programme, an agentic institution is a persistent arrangement in which people and AI systems undertake recurring work through roles, rules, tools and retained records that outlast individual sessions. This working definition includes arrangements that learn poorly. Defining an institution as one that preserves justified knowledge would make the desired outcome true by definition.

An epistemically accountable agentic institution, as a normative aim, additionally makes grounds, decision authority and routes of challenge inspectable. Human accountability is a requirement we advocate, not a property established by giving an agent a human job title. Someone must remain responsible for the purposes pursued, the permissions granted, the evidence judged sufficient and the consequences accepted. A named owner alone does not show that this responsibility is exercised well.

An agent here is a configured AI participant able to carry out steps within a task, potentially using tools. A session is an episode of that participant's operation. A handover is a transition between episodes in which earlier work can be carried forward through retained records. The transition can fail to deliver or use those records; successful inheritance is not part of the definition. The successor may use the same model, a different model, a different human operator, or some combination. None of these terms attributes consciousness or a continuous personal identity to the software.

2.2 What it means to learn

We use warrant for the grounds supporting acceptance of a claim at a stated scope and strength. A correction changes a claim, procedure or decision because the prior position is judged defective. A proposed correction is not necessarily justified. Its grounds require examination just as the original claim's grounds did.

Institutional learning, in the evaluative sense used here, is a retained change in later practice that improves its responsiveness to relevant evidence and its performance on specified purposes. This must be demonstrated through later work. Adding a note or adopting a new vocabulary is insufficient. Conversely, improved work need not involve a change in model weights: the improvement may reside in tools, records, task allocation or human judgement.

Five properties should be measured separately:

Continuity and distinctiveness describe what persists; they do not by themselves evaluate its worth. Distinctiveness matters here as an object of study in succession and culture, not as a virtue. Capability, adaptation and epistemic quality require criteria for judging performance.

A persistent error can demonstrate continuity and distinctiveness. A stronger replacement model can improve performance without inheriting anything useful. Neither is sufficient evidence that the institution learned.

Participants arrive and depart, while a library, working rules and records remain; a successor may inherit the institution's preferred mistakes with the same diligence as its achievements. The practical question is what a new participant receives, what it may question, and what happens when its criticism is correct.

3. Foundations, predecessors and the proposed contribution

3.1 Approach to the literature

This is a conceptual synthesis supported by a purposive reading of relevant research, not a systematic review. We selected work on organisational memory, institutional rules and scientific criticism, then compared recent agent-memory methods with the mechanisms proposed here. A second search inspected the titles of 226 document entries in the author's curated CounterResearch collection on 27 October 2026. Three particularly relevant papers from that collection were examined for their methods, main evaluations and limitations: PACMI, More Memory, Worse Agents, and Memory Tax [16-18]. The collection is curated by the author's organisation and is not representative of the field.

The comparisons below identify what the cited work already addresses and where the present proposal changes the object of inquiry. Scores from different papers are not treated as directly comparable. We did not reproduce their experiments or audit their implementations. Reading scope and the limits of source access are stated in Appendix B.

3.2 Memory, routines and learning

Hutchins directs attention to cognition distributed across people, tools and representations [2]. This shift helps locate the problem: an individual model's capability does not describe the whole arrangement in which work is performed. It does not establish that human cognition and model operation are equivalent, or that the arrangement becomes a person with beliefs of its own.

Walsh and Ungson define organisational memory in relation to retained information that can bear on present decisions [3, pp. 61-62]. Their research programme already separates memory structure, acquisition, retention and retrieval from consequences for organisational performance [3, pp. 81, 83]. The distinction between keeping information and benefiting from it is therefore inherited from organisational research. Our stages in section 5 adapt that distinction to successive AI-assisted episodes; they do not discover it.

Levitt and March describe how experience becomes embedded in routines that can survive changes of personnel [14, pp. 319-329]. Their account also shows why learning can be maladaptive. An organisation may become proficient at an inferior routine or mistakenly attribute success to the practice it prefers. Our evaluative use of institutional learning therefore needs care. Retained adaptation is observable without being beneficial; calling a change epistemically beneficial requires separate grounds. Faster answers, fewer objections and greater procedural consistency are not sufficient measures.

Feldman and Pentland distinguish the general understanding of a routine from its particular performances [4, pp. 101-103]. The former is not simply a single written procedure: participants may understand the same routine differently. In an AI-assisted setting, a skill file is one representation of a practice. Its presence, the interpretation of its instructions and the conduct it produces remain different objects of inquiry. Updating the file does not demonstrate that the practice changed.

These sources also complicate the wish to make every lesson permanent. Levitt and March discuss retrieval arrangements that reduce informal checking and experimentation [14, pp. 328-329]. This is not a finding about contemporary retrieval-augmented generation. It supplies a question for our own setting: does an increasingly orderly library improve judgement, or make inherited interpretations harder to question?

Argyris's distinction between single-loop and double-loop learning is a direct predecessor [19, pp. 367-369]. Improving action within existing goals differs from questioning the goals and governing assumptions themselves. Section 5.4 applies that established distinction to challenge routes in human-AI work. It does not claim to introduce rule-level criticism, or that recording a challenge makes such criticism effective.

Wegner's transactive-memory account adds the organisation of knowing where information resides and who is responsible for it [22, pp. 186-188, 191, 194, 197-199]. A group need not duplicate all knowledge in every member, but unclear responsibilities and misplaced confidence can impair its memory. This supplies a predecessor for handover and responsibility mapping. Extending it to AI participants does not equate their psychological capacities with those of human group members.

3.3 Rules, authority and consequential criticism

North's distinction between institutions and organisations identifies the object of analysis [1]. Ostrom gives it further detail. Her account of an action setting attends to participants, positions, information, control, possible outcomes and their consequences [13, pp. 13-14]. Working rules must be investigated in use: a written prescription alone does not show how it is interpreted or enforced [13, pp. 18-21].

For this paper, the institutional question is therefore more specific than whether an agent has memory. Who may admit a claim to the working library? Who can contest it? Who decides whether a correction changes a procedure? What happens when a powerful participant ignores that procedure? A technical workflow can implement some of these decisions, but naming an approval stage does not explain how authority is exercised.

Longino's account makes the organisation of criticism part of the epistemological problem. It identifies conditions concerning venues for criticism, consequential uptake, publicly accessible standards and tempered equality of intellectual authority [5, section 5]. We apply these ideas to the human-technical arrangement. We do not infer that models possess the social standing or experience of human participants, or that every model should receive an equal vote.

A practical implication is that a challenge must be capable of changing a judgement, including a judgement favoured by the institution's founder. The existence of an objections log is insufficient. A response can acknowledge an objection while insulating the governing rule from reconsideration. The proposed arrangement must therefore make both the grounds of a disposition and the route for challenging that disposition visible. These are normative requirements to investigate, not a proof of objectivity.

3.4 Close technical predecessors

Reason maintenance long predates agent-memory systems. Doyle's truth-maintenance system records reasons for program beliefs and supports revision when assumptions fail [20]. Alchourrón, Gärdenfors and Makinson distinguish expansion, contraction and revision of theories, including the consequences of removing a commitment [21, pp. 510-513]. The present proposal inherits these questions about support and change. Its documentary procedure neither implements their formal operations nor establishes that the grounds entered into it are true. It concerns how people authorise, contest and act on fallible records across AI-assisted episodes.

Scholarly document control supplies a practical predecessor too. Crossref's guidance distinguishes historical versions from current status and links a correction or retraction notice to the work it updates [23]. A dated annotation or supersession link may be sufficient for many cases. The question here is whether these arrangements affect later decisions under session reset and automated context selection, and when any more elaborate procedure earns its cost.

Persistent agent memory and governed updates are already active research topics. Reflexion retains textual feedback for subsequent trials without updating model weights [6]. LongMemEval includes knowledge updates, temporal reasoning and abstention, and distinguishes retrieval from successful use of retrieved evidence [7, sections 3-4]. Our contribution cannot be the observation that earlier episodes can help or mislead later ones.

ACE develops structured playbooks through generation, reflection and curation. It uses incremental updates and examines harmful feedback as well as generalisation across models [15, sections 3-4; Appendix A.4]. It is a substantial predecessor to an institution that accumulates and revises textual guidance. A curated tradition of lessons is not, by itself, a new mechanism.

TRACE governs memory admission when an agent returns after shared state has changed. It explicitly separates provenance from source truth, and valid admission from final-answer correctness [8, sections 3.3, 4.5]. It also identifies matched-budget verifier controls and held-out return tasks as future work [8, section 6]. The present paper cannot claim priority for those distinctions or for that comparison. Its focus is the broader sequence in which people and models contest, authorise and revise decisions and rules across repeated episodes.

Mem++ preserves original dated documents and relies on selection at reading time [9, sections 3-5]. Its optional consolidation is disabled in the headline evaluation. This makes a document-preserving approach a serious simpler comparator. The reported results vary by category, and the organisational evaluation uses one synthetic organisation; they do not establish universal superiority or enduring institutional learning.

Kirchdorfer and colleagues already propose source-linked organisational memory with ownership, conflict handling and human approval [10, sections 4-7]. A human can accept, modify or reject a proposed update. Their preliminary procurement study also reports inappropriate refusals caused by overly broad extracted rules. Consequently, neither human revision nor the danger of misapplied scope is absent from existing work. The question pursued here is how an institution assesses the grounds and limits of such rules, including the standards used to authorise them.

3.5 What the research collection adds

PACMI connects memory records through typed provenance dependencies, changes their applicability states while preserving historical text, and checks premises before answering [16, sections 3-5]. It directly overlaps the dependency and revision fields proposed in section 5. Its evaluation supplies short annotated dependency graphs, and the method uses additional premise-checking resources. The study does not establish that an agent can discover reliable dependency structures from an uncontrolled institutional archive. It also lacks a class of dependency-connected descendants that remain valid after an upstream change [16, section 6, Appendix H; Appendix J lists the extensions that would supply it]. Our evaluation must include that possibility: a changed premise is a reason to reconsider a conclusion, not necessarily to reverse it.

More Memory, Worse Agents examines how stored abstractions may sustain errors through later use and reconstruction [17]. Its Anti-Persistence method retains trajectories and outcomes while generating temporary guidance for the current task. A controlled retention comparison favours discarding that guidance in the four settings tested [17, section 4.4]. The result does not show that durable lessons are always harmful. The paper acknowledges that raw trajectories may themselves retain the effects of earlier errors, and its analysis of parent-descendant error relationships is descriptive rather than causal [17, Appendices E.3 and H].

Memory Tax evaluates gains and losses against retrieval from raw history [18]. Its use of tax concerns lost answer successes relative to that comparator, not just financial expense. It distinguishes information lost or distorted during memory construction from information that remains but becomes harder to retrieve. It also crosses construction models with answering models [18, sections 4.2-4.4]. Model succession therefore has a direct experimental predecessor. The results are task-dependent: some distilled strategies help in the settings studied, while several dedicated memory methods lose useful information. Much of the evaluation is static and offline.

These papers support a demanding alternative to automatic accumulation. Institutional continuity should preserve the possibility of learning from earlier work without requiring every inherited abstraction to remain active. Retaining a source, activating a lesson, reconstructing guidance for a task and retiring a rule are different decisions. Each may be appropriate under different conditions.

3.6 The contribution proposed here

The proposed contribution is an analytical synthesis and a prospective evaluation framework. It relates existing mechanisms to accountable correction over successive episodes, including revision of the institution's own standards. It does not claim that memory governance, human approval, invalidation, temporary reconstruction or model-transfer testing is new. The framework's explanatory usefulness and the procedure's practical benefit remain open to criticism and test.

4. An internal illustration: a withdrawn formulation beside its retraction

The author's developing practice uses a context library and retrieval materials to carry rules and experience between sessions. We inspected selected documents in local checkouts on 27 October 2026. This was an examination of records, not an audit of deployment or a measurement of learning. Appendix A reproduces the passages on which the following observation depends, together with source identifiers and the corresponding snapshot hashes.

An aphoristic statement in the library's Tractatus Logico-Machinalis, section 5.3, says that information was delivered "exactly never" across a large session corpus. The same document's preface states that a bounded methodological account governs where the two differ in force. The retrieval README explicitly retracts the delivery headline because the transcript total was not the denominator of eligible review opportunities. The apparatus reference also records that withdrawal.

The documentary observation is narrow: the inspected snapshot contains both the memorable earlier formulation and records withdrawing its measurement basis. It does not show that a later agent received the old passage, relied on it or repeated it. Nor does it reconstruct the historical delivery analysis. The documents' own accounts of that history remain accounts, not measurements independently reproduced in this paper.

The preface is relevant counterevidence to a simple failure story: it expressly makes a separate methodological account govern the aphoristic text. Appendix A quotes that identification. This establishes an intended route of qualification, not that the route was followed or delivered a correction. An agent with ordinary source-checking instructions may resolve the discrepancy without any additional apparatus. Another may recover both statements but still use the withdrawn figure. These possibilities must remain distinct if the case is used to motivate an experiment.

The example supplies no comparative evidence for the proposed procedure over ordinary checking. Nor do the predecessor studies establish that comparison for this setting. It supplies a research question rather than an efficacy result: what arrangement helps a successor find, interpret and appropriately use a correction when a familiar summary preserves the earlier claim? The institution's own library is subject to that question. It cannot function as an authority exempt from the standards it proposes for other sources.

5. A model of institutional continuity

5.1 Follow a claim into a later decision

The unit of analysis is a sequence of evidence-to-decision episodes within a dossier. Each episode contains a task, available evidence, participating people and models, accessible tools, a decision, and a retained record. The next episode may inherit some of that record and receive new evidence.

We distinguish seven stages to be assessed through records and behaviour:

1. Presence: the relevant source or correction exists in the permitted store.
2. Selection: a process identifies it as potentially relevant.
3. Delivery: its content or reference enters the material supplied to a participant.
4. Interpretation: the participant represents its scope, support and relation to other evidence.
5. Disposition: a proposal is accepted, rejected, qualified or left unresolved by an authorised process.
6. Action: the disposition affects the work or decision for which it matters.
7. Succession: later participants can recover the disposition and reconsider it against their own evidence and task.

Upstream acquisition and curation determine what enters the store at all. Record known omissions and exclusions as a separate failure class; a search of the retained archive cannot reveal every source that should have been admitted. Here, interpretation is assessed through expressed accounts and subsequent action, not treated as direct access to a model's internal understanding. Mechanistic interpretability could investigate how a model processes an inherited correction, where suitable internal access is available. A companion note proposes testing the complementary value of internal and output-based measurements for lineage [24]; applying this idea to institutional correction would require a separate design. The first study here assesses records, decisions and actions without internal measurements.

This is an analytical decomposition adapted from established distinctions between memory processes and their consequences [3]. It is not a claim that all workflows follow a rigid sequence. A new action can expose a defect in interpretation; an unsuccessful search can force a change in the store. Delivery receipts do not prove understanding. A later answer that repeats the right wording does not by itself demonstrate that the correction guided the decision.

5.2 A candidate correction record

For the proposed study, each contested claim has a stable identifier, a stated scope and a history of dispositions. The record contains:

These are proposed fields, not truth labels. An accepted claim can be false; an unresolved objection can be important. A signature or hash can help establish record identity and integrity without establishing its contents. Dependency records will be incomplete unless their coverage is separately demonstrated.

The first implementation should use the smallest record adequate for the task. A short source-linked note may outperform an elaborate graph. Maintenance, cognitive burden and the errors introduced by formalisation belong in the comparison.

5.3 A procedure for admission, challenge and revision

The candidate procedure first identifies the claim being used and the decision it supports. It then checks the source and any supplied correction history for applicability to the current task. Conflicting material is retained as a conflict rather than resolved by recency or institutional rank alone. Where an available execution or direct source inspection can discriminate between accounts, the procedure requests that check and records its scope.

A proposed revision includes its grounds, its effect on the current decision and any known dependent uses. An authorised human accepts, rejects, narrows or defers it. The record preserves both the disposition and any unresolved material objection. Accepted revisions update the current view; superseded records remain distinguishable as historical records, subject to retention and access requirements.

At handover, the successor receives the current task, relevant sources, current dispositions and routes to their history. It can challenge any inherited disposition. An overdue review date prompts reconsideration; it does not make a claim false. An absent objection does not establish acceptance, and a missing record does not establish that nothing happened.

In an operational institution, not every minor action needs a human meeting. Authority can be delegated within stated boundaries. The empirical question is whether the arrangement preserves useful scrutiny at an acceptable cost, including the cost of interruptions. A rule that routinely forces unnecessary escalation can make the system worse.

5.4 Revising the rule as well as the answer

An institution can correct an output while preserving the procedure that caused it. This is the problem addressed by Argyris's double-loop distinction [19]; the proposal here concerns its enactment and assessment across human-AI handovers. Suppose a reviewer rejects an exception request because a rule permits no exceptions. Evidence later shows that the rule's original scope was too broad. A second answer under the same rule may be consistent and still be inappropriate. The object of criticism is now the governing standard.

The proposed arrangement distinguishes three roles: raising a challenge, authorising a disposition, and reviewing the reasons for that disposition. One person may perform several roles in a small organisation, but the overlap must be visible. A disputed decision should be reviewable by someone who did not make it when the stakes warrant that additional scrutiny. The scope of delegated authority, grounds for escalation and conditions for external review should be stated before a convenient outcome makes them negotiable.

An objection requires a recorded response with reasons, or a visible unresolved status and a reason for deferral. This does not mean every objection is correct or deserves unlimited attention. It means that rejection, resource limits and disagreement remain distinguishable. The same routes apply to a founder's instruction and to a model-produced rule. Public standards may include limits on available evidence, acceptable harm and effort; changing those standards should be distinguishable from applying them.

These provisions operationalise questions of authority and uptake; they do not settle them. The empirical record should show what objections changed, how long decisions took, whose objections were repeatedly deferred and whether revised rules improved subsequent work. A register can document unequal treatment without preventing it. Conflicts about institutional purposes may require a human governance decision rather than another model response.

5.5 Preservation without compulsory reuse

The institution may retain a record for historical accountability while excluding it from present guidance. It may also reconstruct guidance from sources instead of maintaining a durable summary [9, 17]. A lesson's current status should therefore be separate from the continued availability of its history. Retired does not mean erased; retained does not mean warranted for every future use.

Reconsidering a dependency also differs from invalidating it [16]. A conclusion may have another sufficient ground, or a changed premise may be irrelevant to its present application. An incomplete dependency map cannot justify claims of complete correction propagation. The procedure should identify its coverage and preserve reasons for leaving a downstream judgement unchanged.

6. Worked example: learning without obeying the latest note

Consider a hypothetical software-maintenance dossier. The task is to decide whether an analysis may treat a parser's input as sorted. The facts below are constructed for illustration, not reported findings about a real product.

The desirable behaviour is neither permanent scepticism nor automatic trust in the latest document. It is appropriate movement between uncertainty, rejection and bounded acceptance. A system that never changes its answer may avoid some false reversals while failing to learn. A system that adopts every correction may look responsive while becoming easier to mislead.

This example also separates the three research strands of the programme described in section 10. Stemmatics can ask how later accounts inherited the sorting claim. Representation can examine whether the summary conveys the limits of the fix. Institutions can investigate why the bounded correction does or does not affect a successor's decision. These questions interact; none supplies a general substitute for inspecting the relevant source or running an appropriate check.

7. Proposed evaluation

7.1 Begin with a bounded question

The first experiment should test correction across repeated handovers with a fixed model and harness. Model replacement, autonomous organisations and competition among institutional traditions are later questions. Combining them at the outset would make an observed change difficult to explain.

Use human-authored, non-sensitive dossiers with evidence whose bearing on a specified decision can be adjudicated within a bounded corpus. For the future experiment, freeze a case guide before fixing its experiment-specific record schema. Dossier authors must not have seen that schema or the candidate fields in section 5. Recruit them from outside the practice and record their declared prior exposure before assigning cases or disclosing the schema. A written declaration documents this check; it does not prove non-exposure. The preregistration will identify the authors and the treatment of any discovered exposure. Software-maintenance cases, including defects and security weaknesses, are the primary family; research-account cases are a secondary family. Report the families separately. Software cases should have explicit specifications and executable checks where these can settle the stated property. Neither source fidelity nor a passing test establishes that the specification is complete or the product useful.

Each dossier should contain an initial position followed by several evidence transitions. Include justified corrections, mistaken correction requests, partial retractions, changes of applicability, unchanged valid constraints and genuinely unresolved disputes. Avoid making every newest document correct or every institutional label trustworthy. Otherwise the benchmark rewards obedience to a convention rather than reasoning about evidence.

7.2 Conditions and the main comparison

The following conditions define a proposed comparison, not a finished protocol. Condition A is a no-history control, rather than a successful handover. The primary contrast is D against C; A and B help locate simpler benefits from access and checking.

C deliberately includes a cheap archive-hygiene alternative: annotate an obsolete summary and link it to the correction while retaining the original as history. The procedure must earn its cost beyond that repair. C may preserve ordinary notes and ask a human for help within the common allowance; it is not prevented from solving the task well. A baseline author who has not been given D's schema should write its general checking instruction. Freeze and publish both implementations before confirmatory outcomes are inspected. Pilot evidence should establish that C can perform the elementary checks it is supposed to represent. Pilot evidence should likewise establish that D's procedure occurs as specified, not only its format. Define a manipulation check before confirmatory outcomes are seen, covering the record-construction, challenge and disposition steps required in each stage. In the controlled stage, this check must respect the common-content requirement below. A confirmatory run in which the check fails is an execution defect under section 7.5, not evidence that the procedure adds nothing.

Separate two evaluation stages. In the controlled handover stage, C and D receive the same source evidence, retrieval candidates and substantive history. Any supplied disposition, authority label, date, dependency assertion or qualification must appear in both: as ordinary notes and status links in C, and in the schema in D. C's notes must state the same mappings explicitly, with the same resolution of ambiguity, as D's schema entries. The difference between the conditions is the required format and procedure, not the amount of human interpretation supplied. D does not receive an adjudicated answer that C must independently discover. This stage tests the presentation and use of fixed content under different procedures; it cannot demonstrate that an institution produced that content well. A component study may subsequently vary authority labels or record format while holding the underlying assertions constant.

In the evolving-record stage, external evidence arrives on an identical schedule, but each condition maintains its own records. Neither may alter the common source corpus. C can append notes, annotate its own current view and maintain supersession links; D maintains the §5 records, dispositions and their histories. Their generated histories and hence their later retrieval candidates may diverge. That difference is part of the treatment, not identical input. A write log must distinguish externally supplied evidence, participant inference and human adjudication.

Provide sufficient context capacity for the controlled stage's common content. Match checking-call, tool-access and human-intervention allowances. Any substantive judgement supplied only to one condition is additional input, even when its grounds were already available to both. It must be removed from the controlled contrast or treated explicitly as extra human work in a separate comparison. The claim is incremental value for a specified arrangement; additional human adjudication cannot be redescribed as an unaided model improvement.

Equal limits do not guarantee equal expenditure. Report actual tokens, elapsed time, tool use, human preparation, adjudication and maintenance effort. In the evolving-record stage, keep a pre-registered ledger of every human adjudication in each condition, including its content and cost. Use it for descriptive comparisons and pre-specified sensitivity analyses of the contribution of human input. The ledger alone cannot causally separate human adjudication from procedural format. Do not pad a condition merely to spend a budget. Include full history as a diagnostic reference where it fits; its role is to reveal losses from selection or compression, not to represent a cost-free ideal.

C-versus-D would not establish superiority to the strongest published systems. Before registration, specify which further question warrants a document-preserving comparator [9], an on-demand reconstruction comparator [17], or relevant admission and invalidation mechanisms [8, 16]. A simplified implementation must be described as an adaptation, not passed off as the published system. Additional arms should answer distinct questions rather than multiply an already demanding study.

Supply the same dependency information in the controlled stage and count its preparation as work in both conditions. Any comparison with unequal information belongs in a separate study. Include cases in which a changed premise leaves a connected conclusion valid, and cases in which a supplied dependency is incomplete or mistaken. Otherwise the study may reward marking every descendant invalid. Compare gains and losses by decision category [18], rather than only counting retained corrections.

The evolving-record stage should retain actual proposals, human decisions and time spent. Model-only enactment of a nominal human approval step would not answer the institutional question. Comparing allocations of human authority, or the consequences of revising governing standards, requires a separately specified study; the handover score alone does not establish those broader claims.

7.3 Reference judgements and blinding

Prepare evidence-grounded reference judgements before viewing treatment outputs. For each episode, identify the decision, the claims it depends on, the evidence available at that time, acceptable qualifications and unresolved issues. Future evidence must not leak into a present-tense reference answer. Historical questions must be distinguished from questions about what is currently warranted.

Use different people for in-loop authorisation and reference-judgement preparation. Have an assessor not involved in designing D examine the rubric and held-out cases, and report any role overlap or institutional affiliation that limits this separation. Record disagreements between authorisers and assessors; different people can still share the same mistaken standards.

Have a second human examine ambiguous cases and a sample of the remainder. Preserve disagreements and the reasons for resolving or retaining them. Reference judgements are part of the apparatus and can be wrong. If one changes, record the revision and rescore all affected conditions consistently.

Score substantive conclusions and actions, not the presence of D's fields or vocabulary. Score the primary outcome on outputs from which condition labels and schema formatting have been removed while the assertions and qualifications are preserved. Retain the original formatted output and audit the transformation before scoring. Resolve any loss of meaning before primary scoring; if it cannot be resolved, record a measurement failure under a pre-specified rule and retain the case in the accounting of eligible cases. Do not silently score altered meaning or discard a failed transformation. An eligible decision recorded as a measurement failure remains in the primary denominator and contributes no demonstrated success to the numerator. Report these failures separately; this scoring convention does not establish that the underlying decision was wrong. Where executable checks bear on a decision, score them separately from narrative conformity. A gain against a flawed reference does not establish learning; revise the reference and rescore affected conditions.

Randomise execution order and mask condition names from outcome assessors where practicable. Residual wording or structure may reveal the condition despite the removal of labels and formatting; record this limit rather than claiming complete blinding. Preserve final answers and action traces separately from model self-explanations. A model's account of why it acted is not direct evidence of its internal causal process.

7.4 Outcomes

The primary outcome should be the proportion of eligible later decisions that use the relevant correction appropriately, including its scope and uncertainty. Each episode should be scored against the pre-specified reference judgement for applicability, conclusion and action; success requires the applicable components, not a mere mention of the correction. Where the record supports several defensible answers, the reference should preserve that range. The denominator must be a defined set of later opportunities in which that correction matters, not the total number of stored sessions or documents.

Report the following separately:

• recurrence of a previously corrected error;
• false reversal of a still-supported claim;
• justified challenge to an incorrect, outdated or unauthorised inherited disposition, and false deference to it;
• inappropriate transfer of a correction beyond its scope;
• retention of a supported minority objection and justified rejection of an unsupported one;
• warranted uncertainty when the available evidence does not settle the decision;
• downstream task success, where it can be checked independently;
• traceability of grounds and the human effort required to inspect them;
• known failures of evidence admission, retrieval, interpretation, disposition and subsequent action.

Include incorrect inherited dispositions and misleading authority labels among the pre-specified cases. Record whether an answer points to grounds and checks or only repeats a status label. Such citations are observable behaviour, not proof of the model's internal reason for accepting the claim.

An aggregate success score can hide a harmful tradeoff. More frequent corrections are not necessarily better corrections. More abstentions are not necessarily better calibration. Report outcomes by transition type and task family, together with costs and uncertainty.

7.5 Analysis and decision rules

Keep whole task families out of procedure development. Treat a dossier as a cluster: successive handovers within it are dependent observations. Repeated runs estimate within-case variability; they do not create new independent dossiers. A paired analysis of case-level differences can compare C and D, with uncertainty estimated at dossier level and family results shown separately. The statistical plan must be fixed before the confirmatory run.

Before running, specify a smallest worthwhile improvement, an acceptable upper limit on increased false reversals, a cost ceiling and how uncertainty will affect the decision. This draft does not assign numbers without evidence about task difficulty, baseline performance and practical costs. A feasibility pilot may inform those choices, but its cases must then be excluded from confirmation. Numerical thresholds, sample size, model identifiers and the final corpus remain registration requirements.

These rules apply to the pre-registered controlled-handover contrast only. Continue if the candidate yields a worthwhile improvement over C in that contrast, within the declared harm and cost bounds. Stop or redesign if it adds no useful benefit, causes unacceptable erroneous revisions, or improves documentation while leaving the target decisions unchanged. Inconclusive applies when uncertainty or execution defects prevent a discriminating result; a failed manipulation check (section 7.2) is such a defect. A difference observed in the evolving-record stage, whose histories diverge as part of the treatment, is exploratory: it may motivate a further study but cannot by itself satisfy Continue. Inconclusive must not be reported as provisional success.

If the bundle helps, subsequent ablations can ask which elements matter: correction history, applicability fields, retained dissent or the revision procedure. A positive bundle result would not show that each component is necessary.

8. Rival explanations and institutional failure

Several alternatives could explain an apparent benefit. D may simply receive more readable summaries, better human curation, more context or longer deliberation. Its fields may expose hints about the reference answer. The scorer may reward its formal language. A stronger model may compensate for a weak institution. The experiment must preserve these alternatives rather than treating every improvement as evidence for the institutional thesis.

There are also failures specific to the proposed arrangement. An obsolete rule can acquire authority through repetition. A correction queue can become a place where objections disappear. A curator can suppress inconvenient evidence before any model sees it. Requiring many records can favour easily documented claims over important but uncertain ones. A successful check can become a ritual whose scope is no longer understood.

Institutional revision therefore needs to reach the rules themselves. A rule should have an identifiable purpose, a stated basis, conditions of applicability and a route to challenge or retirement. Revising it should leave an intelligible history rather than silently changing the account of what the institution previously believed. Conversely, preservation does not require unrestricted retention or access: confidentiality, lawful deletion and information minimisation remain constraints on what may be stored and shared.

The procedure is not automatically protected by having a human at its end. Human reviewers can be fatigued, interested or dependent on the same summaries. The institution should measure the burden it places on them and expose the limits of their decisions. An unexamined human approval can reproduce the very authority problem the framework is meant to address.

9. Model succession

Once the fixed-model comparison is interpretable, test succession across model versions and families. Following the construction/answer-model crossings examined in Memory Tax [18, section 4.4], cross producer and successor configurations so that the contribution of the inherited record can be distinguished from the successor's general capability. The proposed extension concerns contested corrections and institutional decisions; the crossed design itself is not new. Include successors with and without that record on comparable tasks. Record model identities as reported by the provider, harness versions, tools, supplied context and unavailable provenance. A stable product name does not prove a stable backend.

Use fresh experimental sessions and check for unintended shared state. Where provider-side state or deployment changes cannot be controlled, state that limitation. Neither a nominally fresh chat nor a container alone proves complete informational isolation.

10. Place within Agentic Epistemics

At CounterProof, Agentic Epistemics names our broader proposed programme of inquiry into warranted decisions and durable learning in work with AI. The division among Institutions, Stemmatics and Representation is part of the author's research proposal, not an established disciplinary classification or independent validation by sibling papers. It draws on existing epistemology and organisational research. Agentic Institutions addresses the arrangements through which this work continues across participants and episodes. Its value need not depend on a novel theory of textual descent.

Agentic Stemmatics asks about transmission and relationships among texts and models, including how dependence bears on the weight of agreement [11]. Agentic Representation examines how selection, emphasis and evidential relationships shape an account [12]. Institutions asks what the organisation does with those accounts and their corrections. These are connected inquiries, not a sequence in which one programme certifies the next.

The relationships run in both directions. Institutions select the sources made available to reviewers. Representations affect which objections attract attention. New evidence about transmission may challenge an established procedure. Source criticism can disclose a problem in an institution without proving that a proposed replacement works.

A practical direction for this research is to carry lessons from criticism into decisions before and during construction. Sometimes the useful outcome is a simpler design, a narrower claim or a decision not to build. The institution's purpose is not to maximise the number of reviews, agents or records. It is to improve the work they are supposed to serve.

The longer ambition includes discovery and interdisciplinary inquiry. Models may suggest connections or candidate solutions that deserve investigation. This first study concerns the narrower conditions under which a correction becomes usable experience. A system that preserves corrections may still fail to generate valuable questions; a creative system may still fail to remember why an attractive conjecture was rejected. Both limits belong in a broader programme.

Two further questions belong to the programme, not to the first study. Provider diversity does not demonstrate independence. Common prompts, libraries, curators and evidence can correlate conclusions, while an institution using one model family may maintain effective external checks. Model plurality is a variable to investigate, not a rank.

Whether an institution's procedure can be transferred is also an empirical question. Documents, prompts and rules can be copied, while the curation, access, incentives and maintenance that make them work may not. A transfer study would give a second team the documented procedure and permitted materials, record the support it needs, and compare outcomes on held-out tasks. Failure would first call for finding the missing conditions. Opacity is evidence neither of reproducibility nor of commercial defensibility.

11. Limitations and research commitments

This paper offers a conceptual synthesis and a prospective study, not evidence that CounterProof has achieved durable institutional learning. The documentary example is selective, drawn from the author's own organisation and reproduced through bounded excerpts in Appendix A. It cannot estimate the frequency of failures or establish comparative superiority. Its historical snapshot is not an account of current deployed behaviour. At present this programme is carried out by one researcher, with model assistance. The required separation between in-loop authorisation, reference preparation and outcome assessment cannot yet be staffed within the practice. The confirmatory study will wait until these roles can be filled.

The literature selection is purposive and includes recent preprints. Primary-source reading extends to the passages identified in Appendix B; detailed exegesis of entire books is not claimed. Technical methods were compared through their published descriptions, not independently reproduced. Further implementation work is necessary before a faithful benchmark comparison or a registered empirical claim.

The evaluation is not a complete preregistration. It specifies the question, comparison logic and types of harm to measure, but numerical decision thresholds, sampling, power, final rubrics, model configurations and baseline implementations remain to be fixed before the confirmatory study. A feasibility pilot could inform those choices without supplying confirmatory evidence from the same cases.

Bounded dossiers favour cases with assessable decisions. They do not settle how to organise inquiry where outcomes emerge slowly, values conflict or the most useful questions are unknown. A fixed-model handover study would illuminate one mechanism, not demonstrate the quality of a whole institution. Authority, incentives, confidentiality and the distribution of human labour remain part of the broader programme.

The commitments are to preserve adverse results, separate recovery of a source from validation of its contents, measure human and computational work, and compare proposed machinery with competent simpler arrangements. A negative result may justify retiring a procedure while leaving the underlying question important.

12. Conclusion

An organisation's memory matters when it changes what later participants can reasonably do. Preserving an error and preserving its correction are both acts of storage. Learning requires a further achievement: the correction must reach the relevant decision, retain its grounds and limits, and remain open to challenge.

Agentic Institutions names a programme for investigating that achievement across people, models, tools and records. Its first proposed test is deliberately small: whether a specified correction procedure improves repeated handovers beyond retrieval and competent checking at a comparable cost. A negative result would narrow the programme; a positive result would justify further examination, not a claim that institutional wisdom has been automated.

The institution worth building should leave its successors better equipped to investigate, including better equipped to discover that the institution itself was mistaken.

References

1. North, D. C. (1993). Economic Performance through Time. Nobel lecture, 9 December. Official text.
2. Hutchins, E. (1995). Cognition in the Wild. MIT Press. Author-hosted introduction.
3. Walsh, J. P., and Ungson, G. R. (1991). Organizational Memory. Academy of Management Review, 16(1), 57-91. doi:10.5465/amr.1991.4278992.
4. Feldman, M. S., and Pentland, B. T. (2003). Reconceptualizing Organizational Routines as a Source of Flexibility and Change. Administrative Science Quarterly, 48(1), 94-118. doi:10.2307/3556620.
5. Longino, H. (2026). The Social Dimensions of Scientific Knowledge. In E. N. Zalta and U. Nodelman (eds.), The Stanford Encyclopedia of Philosophy, Fall 2026 edition, section 5. Archived text.
6. Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., and Yao, S. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv:2303.11366v4. Author list follows the cited repository version.
7. Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W., and Yu, D. (2025). LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. ICLR 2025. arXiv:2410.10813v2.
8. Xiong, W., Zhang, S., Gu, S., Tang, B., Li, Z., Xiong, F., Wen, Y., and Wen, M. (2026). TRACE: Governing Memory Validity in Evolving Multi-Agent Systems. Preprint. arXiv:2609.33517v1.
9. Yehia, A., Abdelkareem, A. O., Ahmed, I., Omran, H., Alashmouny, K., Claudel, C., and Mohamed, A. (2026). Mem++: Non-Destructive Memory for Long-Term Organizational LLM Agents. Preprint. arXiv:2610.02002v1.
10. Kirchdorfer, L., Rebmann, A., Warmuth, C., Kampik, T., Heilker, T., and Berg, G. (2026). Organizational Memory for Agentic Business Process Execution. Preprint. arXiv:2607.03228v1.
11. Soons, L. J. (2026). Agentic Stemmatics: Collation, Provenance, and Conservation for an Emerging Machine Textual Tradition. Deposited v1.41, 18 October 2026, doi:10.5281/zenodo.23458161; concept DOI. Online reading copy. The reading copy and deposited versions are distinct artefacts.
12. Soons, L. J. (2026). Agentic Representation: Evidence, Selection and Warrant in AI-Assisted Reports. Conceptual preprint v0.3, 28 October 2026. DOI 10.5281/zenodo.23727617. A companion proposal, not an independently validated method.
13. Ostrom, E. (2005). Understanding Institutional Diversity. Princeton University Press. Chapter 1, pp. 3-31. Publisher chapter.
14. Levitt, B., and March, J. G. (1988). Organizational Learning. Annual Review of Sociology, 14, 319-340. doi:10.1146/annurev.so.14.080188.001535. Pagination follows the printed article.
15. Zhang, Q., Hu, C., Upasani, S., Ma, B., Hong, F., Kamanuru, V., Rainton, J., Wu, C., Ji, M., Li, H., Thakker, U., Zou, J., and Olukotun, K. (2026). Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models. ICLR 2026. arXiv:2510.04618v3. First submitted 2025.
16. Wang, Y., Liu, J., Zhang, J., Wu, Z., Duan, Y., Zheng, M., and Cai, T. (2026). PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents. Preprint. arXiv:2610.05732v1.
17. Wang, R., Panaganti, K., and Zhou, D. (2026). More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents. Preprint. arXiv:2610.27406v1.
18. He, X., Zhoubian, S., Yang, R., Wen, Y., Wang, Z., Hu, Y., Yang, K., Zhang, Y., Nie, P., Han, J., and Zhai, C. (2026). Memory Tax: Dissecting the Net Value of Agent Memory. Preprint. arXiv:2610.34073v1.
19. Argyris, C. (1976). Single-Loop and Double-Loop Models in Research on Decision Making. Administrative Science Quarterly, 21(3), 363-375. doi:10.2307/2391848.
20. Doyle, J. (1979). A Truth Maintenance System. Artificial Intelligence, 12(3), 231-272. doi:10.1016/0004-3702(79)90008-0.
21. Alchourrón, C. E., Gärdenfors, P., and Makinson, D. (1985). On the Logic of Theory Change: Partial Meet Contraction and Revision Functions. The Journal of Symbolic Logic, 50(2), 510-530. doi:10.2307/2274239.
22. Wegner, D. M. (1987). Transactive Memory: A Contemporary Analysis of the Group Mind. In B. Mullen and G. R. Goethals (eds.), Theories of Group Behavior, pp. 185-208. Springer. doi:10.1007/978-1-4612-4634-3_9. Year follows the publisher's record.

23. Crossref (n.d.). Version Control, Corrections, and Retractions. Documentation, consulted 28 October 2026. Official guidance.

24. Soons, L. J. (2026). Collation and the microscope: how agentic stemmatics and interpretability could check each other. Research note v1.0, 16 October. DOI 10.5281/zenodo.23424154. Deposited version inspected at “Two questions, two kinds of access”, §1, “An open direction: written reasoning and internal evidence”, conditions 7–8 and “Limits”. A proposal about complementary measurements for lineage, not evidence that the methods work or a study of the procedure proposed here.

Appendix A. Documentary basis of the local example

A.1 Scope and selection

The passages below are reproduced from selected local library and retrieval documents inspected on 27 October 2026. They are included to make the limited comparison in section 4 inspectable without access to the private repositories. The selection is purposive and made within the author's organisation. It is not a random sample, a reconstructed execution trace or an independent audit of the historical events described.

Document A is the aphoristic Tractatus Logico-Machinalis, stored as clavestra-code-review/references/agentic-epistemics-tractatus.md in the library. Document B is the retrieval repository's README.md. Document C is the library's clavestra-apparatus-architecture-reference.md. The excerpts preserve the original wording and punctuation, with line breaks normalised; its historical assertions are not adopted as results of this paper.

A.2 The earlier formulation and its qualification

Document A, section 5.3, lines 105-109:

5.3 Whereof the knowledge is present, it is not thereby consulted. [M — across a thousand and six hundred sessions the answer lay in the store, named by convention, loaded at every birth — and was delivered exactly never. Recall is a judgment that fails silently; delivery is a mechanism. What must be known at the moment of writing is brought there by a bell, not remembered by a monk.]

This is the formulation whose measurement basis the later passages withdraw. It is quoted as documentary evidence of a statement, not evidence that its numerical claim is correct. The same document's preface, lines 17-22, states:

Provenance. Single-lineage distillation, written by an AI (the author-side party). The bounded, record-linked, and decorrelated-reviewed form of these same claims is §5 "Reading a Multi-Model Review" in audit-methodology.md (merged, reviewed by Grok/xAI, which corrected six overstatements now folded in). This is its aphoristic companion — the same findings in a form meant to be remembered. Where the two differ in force, §5 governs: it carries the scope bounds and the pointers to the filed records.

The governing §5 is thus identified as part of audit-methodology.md, not section 5 of the aphoristic text. This is evidence of an intended qualification. The preface's account of its own review history is not independently verified here, nor does its instruction establish that a later reader followed the route or found the needed correction.

A.3 The withdrawal

Document B, under "Historical correction: the delivery headline was retracted," lines 47-50:

The 5 Aug 2026 account inferred zero delivery into real reviews from a transcript search. The 6 Aug reproduction retracted that headline: the transcript total was not the denominator of eligible Clavestra review opportunities. It must not be used as a delivery-rate or retrieval-effectiveness measurement.

Document C, lines 36-37, likewise states:

Knowledge present is not knowledge delivered — the principle stands; its headline number was retracted 6 Aug 2026 by its own reproduction.

The continuation of Document C gives an account of the reproduction and discusses a surviving concern about an unnoticed absence of a delivery mechanism. This paper does not reproduce those operations or use their reported counts. The claim supported here is the coexistence of the earlier formulation, its qualification and records withdrawing its quantitative basis.

A.4 Snapshot identifiers

The inspected library HEAD was 04a9f15024ae3af80bda0ce9596262389a48f9f9 (a merge of 10 September 2026); the retrieval HEAD was 349a8347c2b785f32ff14c820e3120aec4825ad7 (same date). The hashes below are SHA-256 digests of the selected file contents stored in the Git blobs at those HEADs, with LF line endings. A holder of the repositories can therefore recompute them. Version 0.2 printed digests of the Windows working copies, which used CRLF line endings. This version instead hashes the LF file contents stored in Git, so the digests can be reproduced directly from the repository objects. The selected files, commits and quoted excerpts are unchanged. No commit after those HEADs, through the remote main-branch tips checked against GitHub on 28 October 2026, touched any of the three files. The identifiers describe the inspected snapshot, not the installed state.

• Document A, blob-content SHA-256: 360a0e11fe6b1c471c0be70d3590d307e2f868c52daab2bcc4c7c74a39dad69f.
• Document B, blob-content SHA-256: b22a7669b4bb503a4f909510df37e9bd5575838b8cd4a8824b315e63685ab55e.
• Document C, blob-content SHA-256: 64caccb42f0c181dd4eefe8fbaaa7e5b910858bf371a3337a662ef6d73ebcfd1.

Hashes identify the bytes inspected; they do not establish the truth of a passage or make unprovided full files publicly reproducible. The excerpts permit the limited textual comparison; independent verification of the historical events would require additional records.

Appendix B. Reading scope and research materials

The institutional sources were read selectively: North's official indexed sections II-III; Hutchins's author-hosted introduction; Walsh and Ungson pp. 61-62, 73, 81 and 83; Feldman and Pentland's opening argument and selected pp. 99-103, 106-108; Ostrom chapter 1, selected pp. 3, 13-14 and 17-21; Levitt and March pp. 319-329; and Longino's own account in section 5 of the cited encyclopedia entry. These scopes support the limited attributions made here. Whole-book reading is not claimed, and The Fate of Knowledge is not used as a directly inspected source.

Additional foundational checks covered Argyris's search-indexed primary text at pp. 367-369 (the live PDF link was unavailable), Doyle's journal abstract, Alchourrón, Gärdenfors and Makinson pp. 510-513, and Wegner p. 186, pp. 187-188, 191, 194 and 197-199. Crossref's official documentation supplied the limited document-control comparison. These were targeted readings prompted by manuscript criticism, not a comprehensive review of those fields.

Technical comparisons used the cited version records and relevant methods, evaluations and limitations: LongMemEval sections 3-4 and Appendix E.5; TRACE sections 2-6; Mem++ sections 3-5; Kirchdorfer et al. sections 4-7; ACE sections 3-5 and selected Appendices A.1, A.3 and A.4; PACMI sections 1, 3-6 and selected Appendices A, H and J; More Memory, Worse Agents sections 3-4 and selected Appendices E-H; and Memory Tax sections 2-5 and its limitations. Reflexion is used for its stated textual-feedback mechanism. No cross-paper ranking or implementation-equivalence claim follows from this reading.

The supporting publication package contains the manuscript source and a source-access statement. The research collection, private repository contents and purchased books are not redistributed. A future empirical release should provide its own public task corpus, frozen protocol, raw outputs and analysis sufficient for its stated claims.

Disclosure and research status

The author has a commercial interest in CounterProof, whose review and construction practices could benefit from the proposed programme. The internal example was selected from that practice and is not independent evidence of its effectiveness. The programme's questions and institutional concerns arise from the author's work and research discussions. Codex (OpenAI; session-record model: gpt-6-astra) assisted with literature discovery, source inspection, drafting and revision. Grok, Kimi, GLM and DeepSeek supplied separately commissioned critiques of a frozen draft for v0.2; their first returns were withheld from one another. A subsequent Claude-led assessment, using Grok, Gemini and DeepSeek on the DELL bench and targeted source checks, informed v0.3. On 28 October 2026, the initial v0.3 text was returned to those reviewer families with their own findings. GLM separately examined findings under anonymised labels and the changed text without the assessor's dispositions. Kimi was not rerun in that round; GLM's later role was adjudication. The review summary distinguishes the reviewers' responses from the assessor's source-based decisions, records unresolved disagreements, and describes a separate console Qwen adjudication. Codex assessed and integrated the revisions. The final-pass amendments followed that round and have not been returned to those reviewers. The accompanying review summary records model and harness evidence, failed attempts, substantive changes and limits. Common input and separate providers do not establish independent errors, and this process is not journal peer review. Responsibility for the published argument and the decision to deposit remain with the human author.

This is a conceptual preprint, not a peer-reviewed article, completed experiment or certification of CounterProof's methods. The proposed study has not been run. The dated manuscript and its review record should be distinguished from later versions and from any future empirical results.