LLM Wiki Coordination synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
A coordination package, not the agent host
LLM Wiki Coordination supplies filesystem procedures for agents maintaining a Markdown wiki, plus a deterministic Python audit. Its claimed work is to preserve discussion, direct subsequent contributions, distinguish proposals from approved structural knowledge, and maintain integrity. External agents interpret instructions and edit files; a user recognizes closure and approves canonical status. The package does not call models, schedule participants, inject context, intercept writes or generate canon automatically. This is a whole-system analysis of the shipped coordination package, not of an enclosing autonomous application. See RT-OBJ-coordination-protocols, RT-ABS-runtime-orchestration and RT-BAP-protocol-guidance.
The evidence is SRC-1, https://github.com/AEVYRA/llm-wiki-coordination, frozen at 126749634d1c8c2fd6141f59711c882c2d629699, with analysis cutoff 2026-10-14. The accepted reports inspected every tracked file, including all protocols and the sole executable. Audit findings are code-grounded; most agent behavior is doctrine/design. Private-use and repair statements are reported operation, not supplied execution traces. No adopting wiki, target execution or causal experiment establishes successful delivery or benefit. The boundary and source register specify these layers and exclusions. This synthesis uses the supplied members only; no source reads or new evidence were added.
Preserve a question and make the next contribution depend on prior work
The representative workflow starts with a structural wiki question. An initiating agent creates stable context and open questions in thread.md, participants and coordination settings in meta.yaml, and an initial numbered contribution. Subsequent open questions are appended; each contribution gets its own file rather than replacing prior prose. These are prescribed write rules, not filesystem protections. The workflow account follows this explicitly connected discussion-to-canon procedure; no concrete completed instance was supplied. See MEM-OBJ-thread-context, MEM-OBJ-thread-meta and MEM-OBJ-thread-entries.
The next participant reads the preceding contribution, scores its Novelty, Coherence and Robustness from 0 to 1, and records last_eval in the new entry. Accumulated scores are compared with a target; the greatest deficit guides the next contribution's emphasis. A Robustness deficit asks the participant to act as critic, not a scheduler to launch a separate critic agent. The retained prose supplies arguments and reasons; symbolic vectors organize attributed judgments. Neither their sum nor target attainment establishes truth. Participants and the external host own selection, file reading and execution. See RT-RTE-thread-participation and EPI-RTE-peer-judgment.
Separate files avoid rewriting the whole history, but do not establish bounded context cost. Participants must also derive cumulative assessments from retained entries; no agent-facing aggregation helper, token budget, truncation rule, isolated context frame or automatic reload path is shipped. The audit's internal sum is diagnostic, not a model-context service. Actual history delivery remains external. See MEM-CLM-context-economy and memory read-back. Historical discussion can also be imported without reconstructing peer-evaluation lineage; preservation does not restore unknown source warrant. See EPI-RTE-archival-import.
The design separately affords cross-thread orientation estimation: sum and normalize an agent's past contribution vectors into zhat, then use that working estimate as preferred-direction guidance. It describes participation orientation, not contribution truth. Its update writer, history window, admission and later delivery are unresolved. The reports disagree whether its doctrine status is afforded or claimed; neither establishes wiring or operation. See MEM-RTE-profile-refinement, EPI-RTE-zhat-update and report verification limits.
Recognize closure, then review the resulting page separately
A closer evaluates the preceding contribution, records a cumulative checkpoint, summarizes discussion and proposes canon-page changes. The user may refuse Recognition, including for an under-target axis. Recognizing this closure is not equivalent to approving the resulting structural page. Page authors initialize revisions and endorsements; each reviewer must read the actual page in the same session before endorsing it. Full canon requires all active agents accepted/contributed, no stale endorsements and explicit user approval. Soft canon instead requires at least two-thirds active agreement, remaining pending/stale statuses aged at least 14 days, no disputes and explicit user approval. These are external-agent/user obligations, not an implemented admission gate. See RT-RTE-canon-admission and MEM-OBJ-consensus-block.
The procedural delivered result is an approved structural page with revision-specific endorsements, not merely a closing entry. Significant changes—more than 30% or structural reorganization—require revision increment and staling others' endorsements; revision mismatch also requires renewed review. Disputed soft canon returns to working. Premature closure can be challenged through a reopen entry that evaluates the closure and records why it was premature; one reopen is allowed per closure. These recovery paths preserve criticism rather than rolling back old prose. Their execution remains unobserved. See workflow recovery and delivery and RT-RTE-canon-admission.
The resulting authority is bounded. Endorsements record agreement at a revision, not authenticated identity or factual warrant. No concrete candidate/evidence/criterion pair establishes acceptance of a particular ampliative claim, and no completed canon transformation establishes faithful preservation of reasons or novel inference. Candidate generation and canon derivation therefore remain semantically indeterminate at this boundary. See EPI-RTE-entry-authorship, EPI-RTE-canon-review, EPI-RTE-canon-integration and MEM-OBJ-canon-pages.
Diagnose stored structure without treating a pass as approval
The separate maintenance path invokes the audit with a wiki root and flags. Python reads Markdown under wiki/ and dialogue specifically under wiki/agents/dialogue; it runs L0 hygiene, L1 graph and L2 protocol checks unconditionally, adding L3 cross-protocol checks only when the layer setting is at least 3. It prints text or JSON findings and, on normal completion, exits 1 if any HARD finding was collected and 0 otherwise. Direct method calls lack that terminal gate; parser-shape or filesystem failures may interrupt completion. The CLI neither repairs files nor persists reports. External agents perform semantic review, save reports/log summaries and escalate contested Hard findings to the user. Audit is not mandated at every dialogue handoff or canon promotion. See RT-RTE-audit-invocation, RT-BAP-audit-exit and MEM-RTE-audit-retention.
Its findings establish local predicates, not a general integrity warranty. Vector consistency does not validate semantic scores; accepted/pending checks do not enforce canonical quorum, revision validity or user approval. Key-custody text heuristics and signature-field presence do not authenticate anything. There is no raw-source baseline/approval comparison or semantic contradiction evaluator. Work blocking requires an external caller honoring the exit result. See EPI-RTE-vector-check, EPI-RTE-consensus-check, EPI-RTE-crypto-check, EPI-RTE-cross-check and EPI-ABS-source-check.
Two source-level countercases materially narrow the package's stronger guarantees. Alternate evaluation references receive WARN without resolving older authors, so placing evaluation in the next file does not make self-evaluation impossible. Archival metadata bypasses metadata and entry-chain checks even after reopening, although the prose requires normal schema for new entries. Other audit passes still run. Adopter-chosen roots outside the fixed dialogue traversal also lack these diagnostics unless the checker is adapted. These are inspected branches, not observed exploits. See RT-CLM-peer-evaluation, RT-RTE-archival-audit and RT-RTE-audit-invocation.
Retained adaptation is specified; improved capacity is not demonstrated
The strongest supported adaptive contribution is the prescribed conversion of prior agent contributions into durable peer assessments and closure summaries/checkpoints that later participants use for direction or continuation. Under the comparison's trace-fed-write definition, this supports a local claimed trace-learning path from recorded conversational contributions, classified as session logs. It does not establish a shipped automatic writer, exercised learning or improved future-action capacity. Agent admission without approval of every score is distinct from manual closure/canon approval. Profile verification confirms these local distinctions; MEM-OBJ-thread-entries, EPI-RTE-peer-judgment and RT-RTE-thread-participation supply their warrant.
There is also a concrete reported method repair: the changelog identifies an inverted deficit-weighting formula and retains its correction and rationale. This is criticism of an addressable coordination rule, not merely a changed page. However, no supplied later operation shows that the corrected rule was consumed or that future capacity improved. The comparison is a reported formula correction, not a causal evaluation of the whole package. See RT-RTE-package-revision.
The independent theory-builder conditions remain distinct. For participation, condition 1, localized method content, is afforded; condition 2, decisions depending on it, is claimed; condition 3, content-directed criticism, is claimed; condition 4, retained criticism shaping another round, is claimed, not observed. For the reported package repair, conditions 1–3 retain those respective statuses, while condition 4 is uninspected: retained correction does not establish a next criticism round. Candidate theories in an adopting wiki are uninspected. See RT-RTE-thread-participation, RT-RTE-package-revision and RT-RTE-canon-admission.
Learning as improved future-action capacity remains uninspected. Reflection is claimed: represented participation and method criticism are intended to affect guidance, but own-operation traces do not establish both causal directions. Reflective and autonomous theory-building remain uninspected; there is no role-by-role computation trace, and user approval is explicitly prescribed. Exercised self-improvement at the shipped-package boundary is not established: reported repair changes retained organization, but subsequent live behavioral dependence and an independently evaluated improvement objective are missing. None of these properties follows from the trace-learning classification or from another property. See RT-RTE-package-revision and RT-RTE-thread-participation.
For an externally scheduled, cooperative discussion-to-document scenario, the package specifies durable handoffs, separate page review and recoverable objections. For a scenario requiring automatic or adversarially robust admission, authenticated review, bounded context delivery or demonstrated knowledge improvement, those procedures and retrospective checks do not establish the required guarantees. A downstream trace linking contributions, criticism, user decisions, page revisions and subsequent consumption would change the operational assessment. A host-enforced admission path covering alternate roots, archival reopening, quorum and revision rules would change the enforcement assessment. A comparison of subsequent behavior against an independently specified objective would be needed to establish improvement, not merely more retained content.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied adopting-wiki run or causal comparison | SRC-1; RT-RTE-thread-participation, RT-RTE-canon-admission, RT-RTE-audit-invocation | Complete shipped package; external host, private wiki and execution excluded/unavailable | Successful compliance, delivery, deployed enforcement, semantic quality and benefit | A bounded downstream execution record with artifacts, decisions and later consumption; suitable outcome comparisons |
| Closure-inclusive arithmetic conflicts with a participation instruction; carried report-verification limit | SRC-1; MEM-OBJ-thread-entries, RT-RTE-thread-participation, workflow account, epistemic bounded conclusion, report limits | Detailed format and checker include closures; RoleSpace checklist names contribution/reopen only | All shipped participation instructions yield the same deficit after closure/reopening, or agents follow the checker rule | Harmonized instructions for a revised system; execution evidence showing which rule agents used at this revision |
| Unresolved orientation-route doctrine-status conflict; carried report-verification limit | SRC-1; MEM-RTE-profile-refinement, EPI-RTE-zhat-update, report limits | Shared cross-thread normalization procedure; memory says afforded, epistemic says claimed |
One agreed doctrine status; neither status establishes implemented trace-fed updates | Reconciled status interpretation retaining both source assertion and formula; identified writer/admission and update evidence for stronger operational claims |
| Orientation estimation lacks update and validity evidence | SRC-1; MEM-RTE-profile-refinement, EPI-OBJ-zhat, EPI-RTE-zhat-update | Formula and preferred-direction metadata; no identified writer, window, calibration or live replacement | Complete profile provenance/control, descriptive validity, automatic learning or improved participation | Specified history/estimation scope, update control, persisted input/output lineage and later-consumer evidence |
| Audit coverage is retrospective and partial | SRC-1; RT-CLM-peer-evaluation, RT-CLM-canon-enforcement, RT-RTE-archival-audit, EPI-ABS-source-check | Fixed-root traversal, alternate references, archival branch, narrow consensus checks, no write interception or source baseline | Structural impossibility of self-evaluation, complete protocol enforcement, source immutability, canonical approval or semantic truth from a clean audit | Revised checks/interceptors covering those paths and obligations; separate semantic evidence and authenticated approval records |
| Canon transformation and acceptance are unspecified at proposition level | SRC-1; MEM-OBJ-canon-pages, EPI-RTE-canon-review, EPI-RTE-canon-integration | Closure proposal and page-consensus doctrine; no downstream candidate pair or acceptance instance | Faithful reason preservation, novel warranted inference, accepted ampliative claims or completed lifecycle integration | Linked discussion/page versions, retained reasons, named claim scope, evidence-consuming criterion and recorded acceptance/integration |
| File organization does not bound model context | SRC-1; MEM-CLM-context-economy, RT-ABS-runtime-orchestration | Entry layout and cumulative reads; external context loader uninspected | Constant update token cost, complete history delivery, context isolation or budget guarantees | Host selection/aggregation/budget mechanism and model-visible context traces with scaling measurements |
| Manifest continuity and maintenance read-back remain unresolved | SRC-1; MEM-OBJ-anchor-form, MEM-RTE-audit-retention, MEM-RTE-page-withdrawal | Manifest doctrine and prescribed reports/logs/tombstones; no loader or specified later maintenance reader | Demonstrated identity reconstruction/authentication, or behavioral effects merely from report/tombstone retention | Concrete manifests and reconstruction/recognition checks; specified later consumers with delivered retained content and effects |
| Reported repair does not close the improvement pathway | SRC-1; RT-RTE-package-revision, RT-RTE-thread-participation | Frozen corrected method and changelog, without subsequent own-operation trace or improvement comparison | Improved future capacity, exercised self-improvement, complete iterative/reflective/autonomous theory-building | Traceable criticism-to-update-to-live-consumer sequence, next-round dependence and independently specified improvement evidence; role-by-role evidence for autonomy |