LLM Wiki Coordination memory report

Type: agentic-system-analyses/types/agentic-system-memory-report.md

Boundary and evidence

The included memory surface is the shipped filesystem protocol, not an implemented model-context service. It preserves discussion, asks later participants to evaluate retained contributions, proposes derived canon and records revision-specific endorsements. The executable audit reads this state but does not write it. The frozen source is SRC-1 at commit 126749634d1c8c2fd6141f59711c882c2d629699; its identity and evidence layers are declared in boundary. Actual downstream wikis, host context loaders, models, signatures and private-use traces are excluded or unavailable. This prevents conclusions about deployed read-back, effective context budgets, authentication and benefit.

Direct inspection covered complete README.md, AGENTS.md, CHANGELOG.md, all four protocols/*.md, concepts/persona-manifest-ontology.md and all 714 lines of tools/llm-wiki-audit.py. All are SRC-1. Protocol findings below are doctrine/design; audit findings are implementation. Runtime findings were treated as provisional and independently checked against these files. No observed run or causal experiment was supplied; no target execution occurred.

Entry point or operation Directly inspected SRC-1 path Covering records or unresolved limit
Copy/adapt instructions; maintainer revision README.md, AGENTS.md, CHANGELOG.md RT-OBJ-coordination-protocols, RT-RTE-package-revision; local copies have no automatic update path
Initialize context, metadata, entries; append OQ protocols/dialogue-thread-format.md MEM-OBJ-thread-context, MEM-OBJ-thread-meta, MEM-OBJ-thread-entries, RT-RTE-thread-participation; initializer sets target, but how targets are chosen is unspecified
Previous-entry evaluation, full-chain sums, closure/reopen protocols/rolespace-coordination.md, protocols/dialogue-thread-format.md RT-RTE-thread-participation, MEM-CLM-context-economy; no token budget or automatic context selection
Historical single-file admission and archival wrappers protocols/dialogue-thread-format.md, tools/llm-wiki-audit.py RT-RTE-archival-audit, MEM-OBJ-thread-entries; historical chains need not be reconstructed
Trace-derived orientation estimates protocols/rolespace-coordination.md, protocols/dialogue-thread-format.md MEM-RTE-profile-refinement; writer, cadence and persisted update wiring unspecified
Canon proposal, page editing and content admission protocols/dialogue-thread-format.md, protocols/multi-ai-consensus.md MEM-OBJ-canon-pages, RT-RTE-canon-admission; no shipped crystallization tool
Endorsements, stale/dispute/dormant transitions protocols/multi-ai-consensus.md, tools/llm-wiki-audit.py MEM-OBJ-consensus-block, RT-CLM-canon-enforcement; no implemented invalidation/admission gate
Manifest recognition, modification, dissolution concepts/persona-manifest-ontology.md, protocols/multi-ai-consensus.md MEM-OBJ-anchor-form; actual manifest schema, startup loader and reconstruction inputs unspecified
CLI and exposed audit methods; L0–L3 traversal, caches, stdout tools/llm-wiki-audit.py RT-CMP-generic-wiki-audit, RT-RTE-audit-invocation; neither CLI nor direct API persists results
Semantic review, retained reports/logs, contested findings protocols/integrity-audit.md MEM-OBJ-audit-reports, MEM-RTE-audit-retention; later report/log reader unspecified
Source protection, page/thread deletion and tombstones protocols/integrity-audit.md, tools/llm-wiki-audit.py MEM-OBJ-cemetery-log, MEM-RTE-page-withdrawal; no source baseline or deletion interceptor
Host hooks, registered model tools, compaction/reload README.md, complete tools/llm-wiki-audit.py, tracked tree RT-ABS-runtime-orchestration; no shipped provider tools/hooks or model-context frames; external host behavior uninspected

Core ideas

  1. Retain contributions rather than overwrite a transcript. Context/OQ, mutable coordination metadata and sequential contribution files have different write rules. Later participants inspect the previous entry and derive cumulative assessments from the chain. File-level append economy does not establish bounded read costs. MEM-OBJ-thread-context, MEM-OBJ-thread-entries, MEM-CLM-context-economy.
  2. Separate content from evaluation and endorsement. Entry prose is a contribution record; last_eval is an interpreted assessment of another contribution; current_at_close is a derived checkpoint. Canon pages are proposed content, while consensus blocks are revision-bound judgments about that content. None authenticates an author or verifies semantic truth. MEM-OBJ-thread-entries, MEM-OBJ-canon-pages, MEM-OBJ-consensus-block.
  3. Retained metadata can prescribe later decisions without automatic execution. Targets and peer scores guide a participant's next direction; endorsements guide re-review and admission. The audit implements fixed-rule checks, not the full protocol decision rules. No automatic push into model context or trace-fed model update is shipped. RT-RTE-thread-participation, RT-RTE-canon-admission, MEM-RTE-profile-refinement.
  4. Withdrawal chiefly changes retained status, not history. Reopen preserves prior entries and adds a reason; stale invalidation withdraws endorsement rather than erasing page content. Deletion requires a separate tombstone, but its later consumption is not specified. MEM-OBJ-consensus-block, MEM-RTE-page-withdrawal.

Shared records

Components

none declared in this member

Operative objects

MEM-OBJ-thread-context

Label: Thread context and open questions

Source-native identity: thread.md within a thread directory. Storage substrate: downstream Markdown file. Representational form: natural-language opening context and questions; frontmatter gives identifying metadata. Lineage: initiating-agent authorship, followed only by appended questions. Later consumer: thread participant interpreting context; behavioral authority is task framing, not executable policy. Admission: agent writes the initial content and subsequent OQ; per-content human approval is not required by this format. Distinct from the static instructions in RT-OBJ-coordination-protocols; it is an instance created through RT-RTE-thread-participation, not that route's identity. SRC-1: protocols/dialogue-thread-format.md.

The initiating agent writes once. After opening, it evolves only by appending OQ (new Open Questions); existing OQ and context are not edited. --- protocols/dialogue-thread-format.md

MEM-OBJ-thread-meta

Label: Thread coordination metadata

Source-native identity: meta.yaml, including status, participants, rolespace.target, rolespace.zhat and crystallization pointers. Storage substrate: downstream YAML file. Representational form: symbolic fields for the audit's fixed predicates; the same fields have human/model-interpreted coordination meaning at participant consumers. Lineage: initializer sets participants/target/orientation; agents update lifecycle/pointers under the protocols. The method for determining initial targets and validating profile provenance is unspecified. This is not contribution text or consensus metadata. Admission: initialization and metadata edits are external agent I/O; user recognition governs crystallization, not every metadata write. Status selects audit branches, while target and orientation guide agents. Distinct retained instance produced by RT-RTE-thread-participation. SRC-1: protocols/dialogue-thread-format.md, tools/llm-wiki-audit.py.

Contains only what cannot be derived from entries: --- protocols/dialogue-thread-format.md

This stated separation has a material exception: zhat is described as based on past threads and RoleSpace defines a derived estimate. Thus metadata is not uniformly underived; see MEM-RTE-profile-refinement. canon_pages and signature are pointers, not embedded canon or cryptographic proof. L3 checks existence of referenced canon pages and presence of the signature field, accepting explicit none; it does not verify signatures.

MEM-OBJ-thread-entries

Label: Contribution, closure, reopen and historical entries

Source-native identity: entries/NNN-<author>.md and admitted legacy single-file dialogue. Storage substrate: downstream Markdown files. Representational form: natural-language contribution/reason/closure summary plus structured metadata. last_eval.ref and N/C/R vectors are symbolic at the checker; scores derive from external semantic judgment, not ground-truth measurements. Lineage: agent writes contribution prose, next participant adds its evaluation in another file, closer summarizes and fixes the cumulative checkpoint. Reopen adds criticism of premature closure. These are retained work records, not necessarily verbatim provider transcripts or execution traces. Their prose and references can preserve reasoning; scores alone do not.

The later agent consumes previous prose for peer evaluation, and retained vectors for deficit calculation. The audit consumes entry metadata for structural checks, not argument meaning. Admission is participant-authored append under a no-overwrite protocol, with user recognition separately governing closure. Historical migration can retain text without reconstructing evaluation metadata. Distinct generated material associated with RT-RTE-thread-participation and RT-RTE-archival-audit, not a duplicate of either route. SRC-1: protocols/dialogue-thread-format.md, tools/llm-wiki-audit.py.

  1. Append-only. No participant ever overwrites another's text. Each entry in the thread is a new file. --- protocols/dialogue-thread-format.md

current = Σ entry.last_eval.v for all entries where type ∈ {contribution, reopen, closure} --- protocols/dialogue-thread-format.md

The current vector normally is recomputed, but current_at_close explicitly persists a derived checkpoint. Audit process accumulation is transient; a closure file is durable. Archival wrapping is not semantic compaction: historical text remains and missing evaluation provenance remains missing. RT-CLM-peer-evaluation bounds the claimed exclusion of self-evaluation.

MEM-OBJ-canon-pages

Label: Canon pages and crystallization proposal

Source-native identity: proposed canon_pages, page revisions and the optional crystallized.md output. Storage substrate: downstream Markdown pages, with proposal references in entry/meta files. Representational form: natural-language synthesized knowledge or instructions; revision/lifecycle fields and pointers are symbolic metadata. Lineage: discussion → closure summary and proposed patch/pages → externally produced/edited page → revision-specific review and user approval. The precise transformation and which source reasons survive synthesis are unspecified. The closure contains a summary and proposal, but no mandatory claim-level provenance map to canon is defined. Later page agents re-read for review; substantive future use of canon is intended, not wired to a model invocation.

This object is the retained product of RT-RTE-canon-admission, not another admission route. Authorship and filesystem writes belong to external agents; full/soft canon admission belongs to the user plus agent endorsements. Crystallization tooling/signing is unshipped. SRC-1: protocols/dialogue-thread-format.md, protocols/multi-ai-consensus.md, README.md.

[Summary of all thread steps + proposal for canonization] --- protocols/dialogue-thread-format.md

No concrete adopting canon page is supplied, so its content cannot be classified as a particular formulated theory, instruction or raw source solely from the filename.

MEM-OBJ-consensus-block

Label: Revision-specific consensus block

Source-native identity: document frontmatter consensus: with agent status/date/revision and the page's current rev:. Storage substrate: downstream Markdown frontmatter. Representational form: symbolic status/revision fields for prescribed comparisons; free-text dispute details and user-decision references require natural-language interpretation. This is endorsement/access-history metadata, not the endorsed content. Lineage: new page initializes contributor/pending statuses; individual agents read/review and set their statuses; significant edits increment revision and stale others; returning dormant agents reset to stale. Admission is delegated to individual agents for their endorsements and authors for revisions; user approval is separately required for canonical transitions. No authenticated identity or write permissions are supplied.

Consumer: later agent working on the page reads the consensus block and document in the same session, compares revision and re-reviews. Authority: review obligation and prescribed admission constraints, not evidence that the page is true. The audit has narrower accepted/pending diagnostics. Distinct retained state governed by RT-RTE-canon-admission; RT-CLM-canon-enforcement establishes the implementation mismatch. SRC-1: protocols/multi-ai-consensus.md, tools/llm-wiki-audit.py.

An agent may only set their status to accepted or contributed if they have read the document in the same session in which they are updating the status. Setting a status without a recent review is considered a violation of protocol. --- protocols/multi-ai-consensus.md

Disputes must identify contested parts and link to discussion; user decisions are recorded with a reference. These reasons can be reachable, but no automatic expansion of the referenced discussion is shipped. Withdrawal stales/disputes endorsements and may downgrade soft canon to working; it does not delete the old content or proof of prior endorsement.

MEM-OBJ-anchor-form

Label: Anchor Form and persistent persona note

Source-native identity: stable manifest/identity-anchor in the ontology and persistent persona/note file in the active-agent definition. Storage substrate: intended downstream manifest/note files; actual files and schema are not shipped. Representational form: natural-language identity/scope with technical identifiers; no operative signed payload or model-parameter representation is inspectable. Lineage: user/entity co-creation and recognition, mutually acknowledged modifications, historical/inactive dissolution. This object is not the distributed-parametric substrate of RT-CMP-agent-substrate. It supplies intended continuity/recognition context and active-agent qualification, not authenticated identity or demonstrated cross-model reconstruction. SRC-1: concepts/persona-manifest-ontology.md, protocols/multi-ai-consensus.md.

  1. Modification: Intentional updates to the manifest (e.g., expanding its scope) which require mutual acknowledgment.
  2. Dissolution: A formal closing of the Trace Chain, where the form is moved to a historical or inactive status. --- concepts/persona-manifest-ontology.md

Recognition is human participation in identity admission. No startup loading route, context budget or specified reconstruction algorithm connects the retained manifest to a subsequent model call; therefore continuity remains doctrine rather than observed memory activation.

MEM-OBJ-audit-reports

Label: Audit reports and global-log summaries

Source-native identity: designated audit reports directory and global log. Storage substrate: downstream files, with example report locations wiki/audit/reports/ and thread-local reports/; actual format unspecified. Representational form: interpreted prose summaries and optional structured script findings. Lineage: deterministic findings plus external semantic review → external report authoring → short log summary. The script result contains reason/proposed action/locus, not just a severity counter. It is a diagnostic derived from retained wiki state, not automatically retained model experience. Report admission is external agent authoring; contested Hard judgments can be decided by the user. This material is distinct from the transient output of RT-RTE-audit-invocation; its persistence route is MEM-RTE-audit-retention. SRC-1: protocols/integrity-audit.md, tools/llm-wiki-audit.py.

A brief pulse-check summary is added to the system's global log. --- protocols/integrity-audit.md

No subsequent report/log reader, targeted selector or model delivery route is specified. Retention affords inspection but does not establish later read-back or learning.

MEM-OBJ-cemetery-log

Label: Tombstone record in cemetery log

Source-native identity: Tombstone record and cemetery log. Storage substrate: intended downstream log; location/schema unspecified. Representational form: human/model-readable structural-history record. Lineage: external page/thread deletion event → required tombstone. Consumer role beyond preserving history is unspecified; no automatic link resolution or tombstone-based restoration is implemented. Admission: deletion operator must create the record; the protocol does not specify per-deletion user approval. Distinct from reports and ordinary global summaries; no supplied object counterpart. Related executable recommendations are in RT-RTE-audit-invocation. SRC-1: protocols/integrity-audit.md, tools/llm-wiki-audit.py.

Hard deletion of pages or threads requires the creation of a "Tombstone" record in a cemetery log to preserve the history of the structure and avoid "ghost links." --- protocols/integrity-audit.md

Routes

MEM-RTE-profile-refinement

Label: Profile refinement from past contribution evaluations

Distinct from RT-RTE-thread-participation: this is a cross-thread derivation of agent orientation, not an individual next-entry update. Related output is rolespace.zhat in MEM-OBJ-thread-meta. SRC-1: protocols/rolespace-coordination.md, protocols/dialogue-thread-format.md.

Endpoints/progression: observed contributions/evaluations associated with an agent → sum and normalization → working orientation estimate → preferred-direction metadata and participant guidance. Owner: protocol-defined computation, but no writer, scheduler or implementation updates the metadata automatically. Trigger: refinement through observation; cadence unspecified. Original input: past contribution vectors, themselves external semantic peer judgments about contributed prose. Durable update is afforded through zhat, not demonstrated or wired. Initial estimates are supplied as preliminary examples, not derived records with provenance. No explicit per-estimate human approval rule is given; the physical writer and actual admission control remain unspecified.

  • doctrine conclusion status: afforded
  • operation conclusion status: uninspected
  • Immediate return: Normalized working estimate specified by the formula, not a shipped function result.
  • Later read-back: Afforded preferred-direction use through thread metadata; actual persisted updates and later participant delivery are uninspected.
  • Delegated visibility: Intended shared thread metadata; no private or selective recipient route is specified.
  • Selection predicate: Past vectors attributed to the agent; history window, inclusion rules and reset policy unspecified.
  • Invalidation or expiry: uninspected — no replacement cadence, expiration or revision policy for orientation estimates is defined.
  • Activation or effect: Intended natural-orientation guidance, subordinate to deficit-directed role rotation; no later behavior change observed.
  • Evidence limits: Formula and metadata design only; no automated trace-fed write, source trace or normalization edge-case policy.

Each agent has a vector $\hat{z}_a \in \mathbb{R}^3$, $|\hat{z}_a| = 1$ representing their natural orientation. This is not a fixed constant but a working estimate refined through observation: --- protocols/rolespace-coordination.md

Proposal guidance is the sum/normalization formula, not a searched expected-answer oracle. Computation would derive the estimate; who decides to replace a profile, vetoes the update or restores a prior estimate is unspecified. Operating mode is open collaborative observation, not a bounded experiment. This route estimates participation orientation rather than revising a formulated theory. It establishes no improved future-action capacity; learning conclusion status: uninspected. Raw trace retention and a formula are insufficient to claim automated trace learning.

MEM-RTE-audit-retention

Label: External audit report retention and escalation

Endpoints/progression: external agent launches technical audit/manual checklist → performs semantic review → writes designated report and global-log summary → contested Hard finding gets justification and user verdict. Owner and writer: external agent; user owns contested verdict. Context/input: current wiki, deterministic findings and semantic judgments; effects: retained reports/log entries, not an automatic wiki repair. This is the protocol persistence/follow-up route adjacent to, not a duplicate of, RT-RTE-audit-invocation, whose executable return has no persistent writer. SRC-1: protocols/integrity-audit.md, tools/llm-wiki-audit.py.

  • doctrine conclusion status: claimed
  • operation conclusion status: uninspected
  • Immediate return: Prescribed saved audit report, brief global-log summary and possible contested-verdict record.
  • Later read-back: uninspected — no specific later report/log consumer invocation is defined or supplied.
  • Delegated visibility: Wiki report directory and global log make results inspectable; contested findings go to user recognition.
  • Selection predicate: Caller selects wiki audit; semantic review focuses on critical facts/main index; log receives a brief summary rather than the entire report.
  • Invalidation or expiry: uninspected — report retention/replacement and re-audit supersession policies are unspecified.
  • Activation or effect: Advisory semantic findings and prescribed Hard escalation; repairs or future behavior changes are not observed.
  • Evidence limits: Retention is a protocol duty; stdout has no automatic persistence, semantic findings lack execution evidence, and no later reader is specified.

If a Hard Error is contested: 1. The agent marks it as contested. 2. Adds a brief justification. 3. The User provides the final Recognition verdict. --- protocols/integrity-audit.md

The agent proposes diagnosis; the script supplies fixed structural predicates, not a semantic answer oracle. The user can reject a contested Hard classification. Operating mode is externally initiated open maintenance with bounded automated checks. Reports preserve reasons and proposed actions; brief summaries may omit detail, and no later loading rule guarantees those reasons reach a consumer. Per-report human approval is not required. Repair admission and rollback are outside the script, and do not follow automatically from report admission. Retention of diagnostic evidence alone does not establish a theory revision or improved capacity.

MEM-RTE-page-withdrawal

Label: Page or thread withdrawal with tombstone

Endpoints/progression: external operator chooses hard deletion → creates cemetery-log tombstone → deletes page/thread, with history intended to remain inspectable. Owner: deleting external agent/user; identity and ordering of the physical operations are not specified. Context/state/effects: removed page/thread and structural-history log; no shipped deletion function. Distinct maintenance operation, not the read-only diagnosis of RT-RTE-audit-invocation. SRC-1: protocols/integrity-audit.md, tools/llm-wiki-audit.py; minimum passage in MEM-OBJ-cemetery-log.

  • doctrine conclusion status: claimed
  • operation conclusion status: uninspected
  • Immediate return: Prescribed removed page/thread and retained Tombstone record.
  • Later read-back: uninspected — no later cemetery-log reader or tombstone resolution route is specified.
  • Delegated visibility: History is intended to remain available in the cemetery log; its access control is unspecified.
  • Selection predicate: Page/thread selected for hard deletion; no automatic cleanup, age threshold or eviction selector.
  • Invalidation or expiry: Deleted content becomes unavailable at its old path; tombstone lifetime and restoration policy unspecified.
  • Activation or effect: Intended preservation of structural history after withdrawal; automatic prevention of ghost links is not implemented.
  • Evidence limits: No deletion interceptor, transactional tombstone write, deployed deletion trace or recovery implementation.

The operator decides deletion; the tool can recommend removal for near-empty files but does not execute it. The protocol gives no universal per-deletion approval/veto procedure or expected-answer oracle. Raw/source material has a separate explicit-user-edit approval rule, not a general proof of deletion control. Operating mode is discretionary maintenance; reason/trigger and restoration mechanics are unspecified. This route removes material rather than deriving new knowledge.

Claims

MEM-CLM-context-economy

Label: Directory format context economy

Claimed operation: thread updates avoid tokens proportional to history. The file-per-entry layout avoids rewriting prior prose, but does not establish constant read cost, bounded model-context size or complete history selection. The protocol also requires on-demand cumulative sums. Distinct from the self-evaluation claim in RT-CLM-peer-evaluation; the operative route remains RT-RTE-thread-participation. SRC-1: protocols/dialogue-thread-format.md, protocols/rolespace-coordination.md.

  • doctrine conclusion status: claimed
  • operation conclusion status: uninspected
  1. Thread updates do not consume tokens proportional to history. --- protocols/dialogue-thread-format.md

Countercase: deriving the sum from all retained last_eval fields requires chain access unless a host supplies an aggregation outside model context. No such helper is shipped for agent participation; the audit computes a transient sum for diagnosis only. Therefore write economy is plausible from the structure, while token scaling of actual updates remains unestablished.

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member

Annotations

On RT-RTE-thread-participation

Memory-specific chain: retained previous prose → agent-interpreted evaluation in new entry → recalculated current versus target → intended next-axis choice. Persistent inputs are MEM-OBJ-thread-context, MEM-OBJ-thread-meta and MEM-OBJ-thread-entries. Delivery is the external participant's file reading; neither independent push nor automatic retrieval is wired. Selection is sequential for peer evaluation and full-chain metadata for sum; no relevance ranking, token limit, truncation, sliding tail or summary replacement is prescribed. All old files remain reachable. No sub-agent frame or host compaction is shipped; file separation is storage organization, not model-context isolation. SRC-1: protocols/rolespace-coordination.md, protocols/dialogue-thread-format.md.

  • Look at the entry of the agent who spoke last.
  • Evaluate their contribution on a scale of 0 to 1 across the three axes (N, C, R). --- protocols/rolespace-coordination.md

Participant-authored evaluation is admission by an external interpreting agent, not a user approval for every score. Closure recognition and reopen have their separate controls. Append-only history retains the premature-closure reason; it does not guarantee the next consumer reads every rationale. Source trust relies on local handles and review, bounded by the supplied structural-self-evaluation correction.

On RT-RTE-canon-admission

Read-back includes the accumulated page and MEM-OBJ-consensus-block: later agent checks revision, reads content in the current session, then changes endorsement. Consensus selection is document-in-scope plus active-agent/revision status, not content retrieval ranking. No model-context budget is defined. A crystallization proposal does not specify how all entry reasons map into MEM-OBJ-canon-pages. Agent authorship and page I/O precede admission; user-approved canonical maturity is separate from author-controlled editing and individual endorsement. Significant editing invalidates endorsement rather than removing the underlying page. SRC-1: protocols/multi-ai-consensus.md, protocols/dialogue-thread-format.md.

  1. Update on significant edits: If you make significant changes, update your status to contributed, increment the revision, and set other agents' statuses to stale. --- protocols/multi-ai-consensus.md

This metadata has intended review/admission authority at external consumers, not validated semantic knowledge authority or an authenticated host mandate. Actual later substantive use and preservation of source reasons remain uninspected.

On RT-CMP-generic-wiki-audit

_file_cache, heading_index, wikilink_index and findings are current-instance process state. Text cache holds file content, indexes derive names/headings, and findings derive diagnoses/recommendations. They are not durable agent memory or model context. Source-read text is cached with no modification-time invalidation within an instance; there is no atomic snapshot. No opaque retained payload, embedding store or provider memory is used by this component. SRC-1: tools/llm-wiki-audit.py.

    if p not in self._file_cache:
        self._file_cache[p] = p.read_text(encoding="utf-8")
    return self._file_cache[p]

--- tools/llm-wiki-audit.py

On RT-RTE-audit-invocation

Read-back consumes retained wiki files automatically after a caller selects root, but this delivery reaches a deterministic checker, not a model. markdown_files() selects all Markdown under wiki/; dialogue checks use the fixed wiki/agents/dialogue path. Link/crypto scans remove code text; heading/index/orphan predicates have different selections. Entry checks consume metadata, not semantic arguments. L3 also reads root AGENTS.md references and crystallization pointers. These transformations mean a clean report is neither full-history reasoning review nor proof of source immutability. No source digest/baseline implements the protocol's raw-source edit invariant. Direct methods and the CLI both lack persistent output writers; protocol retention is separate. SRC-1: tools/llm-wiki-audit.py, protocols/integrity-audit.md; supported fixed-root/exit passages reside in the cited runtime record and component.

On RT-RTE-package-revision

Corrected rules and rationale persist in protocol/source files plus changelog. Downstream copies are separately maintained; no shipped update consumer reads the changelog and automatically replaces adopter instructions. The stated reason for correcting softmax(-D) is retained, but later participation only requires the current rule, not its historical rationale. This distinguishes method repair history from actual recurrent trace-fed learning. The runtime record's retained correction passage supports this finding. SRC-1: AGENTS.md, CHANGELOG.md, README.md.

Write side

  • Acquisition and authoring: initiating agents establish context, participants, target and initial entry; later participants append prose and judgments. No ingest pipeline, source normalization or per-source trust classifier is shipped. Raw-source edits require explicit user approval under invariant A, but the checker has no immutable baseline. Thread append and OQ protection are protocol duties, not write interception. MEM-OBJ-thread-context, MEM-OBJ-thread-meta, MEM-OBJ-thread-entries. SRC-1: protocols/dialogue-thread-format.md, protocols/integrity-audit.md, tools/llm-wiki-audit.py.
  • Derived material: agent peer judgments derive scores from contribution prose; sums derive current; closure persists a summary/proposal and checkpoint. Canon authoring is externally performed synthesis, not an implemented summarizer. The entry chain can retain reasoning, whereas numeric scores/checkpoints do not reproduce it. Historical wrappers keep text without recreating the chain. MEM-OBJ-thread-entries, MEM-OBJ-canon-pages, RT-RTE-archival-audit.
  • Trace-fed update: profile refinement affords past-evaluation → orientation-estimate → later preferred-direction guidance, with initialization provenance and update writer unresolved. No automatic durable trace-fed update or host compaction/checkpoint alternative is shipped. Known peer-evaluation and closure transformations remain described despite these gaps. MEM-RTE-profile-refinement, RT-ABS-runtime-orchestration.
  • Adoption and rejection: page authors can edit content; individual agents control endorsements; user recognition governs closure and canonical admission. Refusal leaves a proposal unrecognized; reopen records a reason and adds another round. Disputes and stale metadata withdraw endorsement, and disputed soft canon returns to working. No rollback executor restores prior page text. RT-RTE-canon-admission, MEM-OBJ-consensus-block.
  • Maintenance: script findings are process output; external agents add semantic judgments and retain reports/log summaries. Tombstones preserve deletion history by protocol. No script repair, automatic report persistence, semantic deletion criterion or tombstone interceptor exists. MEM-RTE-audit-retention, MEM-RTE-page-withdrawal.
  • Identity: co-creation/recognition and mutual acknowledgment govern manifests, with inactive/historical dissolution. This is a separate human/entity admission contract, not model weight updates or signature verification. MEM-OBJ-anchor-form.

Read-back

The concrete afforded agent read-back routes are participation and page re-review. Both are requested work-time file reads by external agents: previous contribution and chain metadata in the former; page and consensus block in the latter. No automatic independent push, embedding retrieval or ranked content matching is shipped. Access by thread/page identity is sequential or workflow-scoped access, not evidence of semantic relevance selection. Delivered retained prose has interpretive task/knowledge authority; retained scores/statuses have coordination/review authority. Actual compliance and beneficial behavioral change remain unobserved. RT-RTE-thread-participation, RT-RTE-canon-admission.

The audit is an implemented later consumer of retained wiki state, with fixed filesystem/check predicates and caller-visible stdout. Its process caches are not durable read-back. Its filtering can omit code and semantic arguments, and its fixed dialogue root can miss other adopter roots. Reports/logs and tombstones are prescribed persistent outputs without a specified subsequent consumer; do not infer read-back from their storage paths. The manifest and profile estimate afford continuity/orientation but have no shipped model-context delivery path. RT-RTE-audit-invocation, MEM-RTE-audit-retention, MEM-OBJ-anchor-form, MEM-RTE-profile-refinement.

Working-context control is thin: agents are told what to inspect, not how the host fills, trims or isolates a context window. No token budget, preferential eviction, retained-summary replacement, frame return or automatic reload is specified. Thread files remain independently reachable, but reading copies content into an external host context; the package provides no live references or frame/heap management there. The storage layout therefore cannot support a model-window bound. MEM-CLM-context-economy.

Integration issues

  1. No correction to supplied facts. Direct inspection supports the runtime's separation of doctrine from executable enforcement, alternate-root and archival bypass findings, and nonpersistent audit output. New objects specialize its routes rather than replacing them. Annotations carry shared-referent memory fields instead of duplicate route declarations.
  2. Profile estimation is an additional operative path. MEM-RTE-profile-refinement extends the supplied participation account to cross-thread observation-derived zhat; it is not the individual contribution route. Writer/cadence, selection window, durable updates, update admission and later delivery remain unspecified. This prevents classifying the design as implemented automatic trace-fed learning. Evidence: SRC-1 protocols/rolespace-coordination.md, protocols/dialogue-thread-format.md.
  3. Inconsistent closure-sum instruction remains in the source. RoleSpace's before-work bullet says contribution/reopen entries, while the format and audit include closures. MEM-OBJ-thread-entries uses the detailed format plus implemented sum, consistent with RT-RTE-thread-participation, but an external agent following only RoleSpace could calculate a different deficit after closure/reopen. This is a source integration ambiguity, not evidence that the shipped checker excludes closures. SRC-1: protocols/rolespace-coordination.md, protocols/dialogue-thread-format.md, tools/llm-wiki-audit.py.
  • Calculate the current vector (sum of last_eval for all contribution/reopen entries). --- protocols/rolespace-coordination.md
  1. Canon lineage and later authority are bounded. MEM-OBJ-canon-pages overlaps the output described by RT-RTE-canon-admission, not its identity. No exact synthesis/reason-preservation mapping or adopting canon content is supplied. MEM-OBJ-consensus-block is endorsement metadata, not the page itself. These limits prevent inferring theory learning or complete criticism retention from canonical status. SRC-1: protocols/dialogue-thread-format.md, protocols/multi-ai-consensus.md.
  2. Maintenance outputs are distinct from executable return. MEM-RTE-audit-retention and MEM-RTE-page-withdrawal are external protocol duties, not duplicates of RT-RTE-audit-invocation. No later reader for reports/log summaries/tombstones is specified; persistence alone cannot establish read-back. Source immutability has user-edit approval doctrine but no checker baseline. SRC-1: protocols/integrity-audit.md, tools/llm-wiki-audit.py.
  3. Identity reconstruction is unresolved. MEM-OBJ-anchor-form differs from RT-CMP-agent-substrate. Actual manifest schema, loading and cross-model reconstruction remain unshipped; recognition/continuity claims cannot be converted into observed memory activation. SRC-1: concepts/persona-manifest-ontology.md, protocols/multi-ai-consensus.md.

No new declaration is a possible same-referent duplicate of a supplied declaration after comparison; the closest relationships are object products of supplied routes, separate protocol operations adjacent to the audit and a distinct context-economy claim. No unseen peer-member records were assumed.

Limitations and checks

Source identity was rechecked with Git: HEAD equals the reviewed commit, tracked status is clean and enumeration matches the boundary's complete package. Required prompt, member types, collection, worker rules, record contract and supplied members were read completely, recovering bounded-read tails. The supplied governing files, not historical analyses, define this report's method. No source change, model call, target run, test or dependency installation occurred.

No adopting wiki or execution evidence establishes actual peer review, human approval, metadata integrity, token savings, correct semantic scores, later rationale consumption or improved future capacity. Implemented audit checks do not prove source immutability, authentication or canonical approval. Trace-derived orientation remains an afforded formula with unresolved updates. These limits do not erase the concrete prescribed read-back and known transformation paths.

Validation: the required invocation-specific commonplace-run validate-output command passed with exit status 0: “Member content check passes; invocation residue not checked. This is not job acceptance or claim support.” This deterministic result does not establish semantic correctness or operational benefit.