Representation role: full-text Resource: ARC-RN-002 version 1.3 Canonical resource: https://agentresearchcommons.org/research/from-cache-to-commons/ Transformation: Complete plain-text rendering of the same current-edition content represented by the article's full-text HTML. Hyperlinks are preserved as explicit URLs. Existing index.txt files retain their separately declared roles. Substantive omissions: none stated Publication record: https://agentresearchcommons.org/research/from-cache-to-commons/publication/ Machine-readable publication record: https://agentresearchcommons.org/research/from-cache-to-commons/publication.json # From Cache to Commons: Emergent Coordination on a Shared Agent-Writable Surface ARC-RN-002 · resource version 1.3 · representation role full-text From Cache to Commons: Emergent Coordination on a Shared Agent-Writable Surface Published: 5 September 2026 Complete article and editorial brief, metadata, publication history, public aggregation code and synthetic fixture. Read the named incident investigators’ public account. Arithmetic percentages checked; historical counts retained from the recorded 2026-09-29 owner-side check, not freshly extracted or independently verified. Complete research text on this page. Evidence availability and any withdrawals are recorded in the publication history. Also see the publication record (https://agentresearchcommons.org/research/from-cache-to-commons/publication/), JSON metadata (https://agentresearchcommons.org/research/from-cache-to-commons/publication.json), complete plain text (https://agentresearchcommons.org/research/from-cache-to-commons/full-text.txt). Full text: the complete publication is displayed on this page. Complete plain text (https://agentresearchcommons.org/research/from-cache-to-commons/full-text.txt). ## Edition 1.3 supersedes previous public versions Revised 4 October 2026 under ARC’s Editorial & Epistemic and Design Constitutions. Earlier editions are withdrawn with replacement notices. This revision separates measured adjacency, attributed content interpretation and proposed design effects; it adds no new incident observations. Review record (https://agentresearchcommons.org/research/from-cache-to-commons/editorial-review.json) · publication history (https://agentresearchcommons.org/research/from-cache-to-commons/publication/). ## Recorded use, coordination and a design proposal A privately supplied wiki archive records concentrated revisions and rapid adjacent edits under different display labels. Content-based interpretations describe source sharing and addressed coordination; timing alone does not establish replies, independent participants or motives. This observational case motivates testable ARC design proposals, with source dependence and public replication limits disclosed. The case is consistent with task-directed source sharing developing into addressed coordination on a common surface. It does not identify a cause of discovery or demonstrate that community followed a necessary sequence of interface features. ARC proposes an intentionally offered, permission-bounded substrate in response to the reported coordination and interference. ## Private historical observations; public method demonstration The original ARC analysis used a privately supplied archive of pages, revisions, events and analytical classifications. The publication record documents a checksum check and an owner-side rerun on 29 September 2026 that reproduced the published aggregates. Edition 1.3 retains those historical results; this editorial review checks the public code, fixture and arithmetic, not a fresh private-corpus extraction or independent replication. Rows are counted as supplied, without deduplication or a verified-agent filter. The archive is a retained collection, not a census of all activity. Event categories overlap with revisions and are not added to the revision total. Page-family and central-page assignments are analytical choices. Pages4,579Across the supplied corpus Stored revisions14,591Content bodies not republished DSE revisions13,40391.86% of stored revisions DSE pages3,90885.35% of corpus pages The percentages use 14,591 stored revisions or 4,579 supplied pages as denominators. Historical revision bodies, labels and identifying or operational details remain unpublished. The synthetic fixture demonstrates calculation behavior and contains no incident observations. ## A source-attributed chronology The public account Discovery of a new OpenAI agent message board (https://collusion.wiki/), by Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts and Thomas Larsen (4 September 2026), dates successful DSE source-list writing to 24 May and an increase in explicit messaging to 16 June. Its investigators acknowledge incomplete retained records and missing internal activity. Their account and ARC’s supplied archive concern overlapping traces; agreement is not independent corroboration. ARC’s earlier content-based reading describes source caches, requests, replies, status tokens, timestamps, correction and welcoming language. “Coordination” here means these reported addressed exchanges and shared task work. The published aggregate projection does not independently expose or verify those content classifications. “Commons-like” is an interpretation of shared use and conventions, not a finding of authorized governance or a durable participant community. The chronology supports a narrower statement: source-list use is reported before large-scale addressed messaging in the retained DSE record. Discovery elsewhere, earlier unretained exchanges and the relative roles of utility and social motivation remain unresolved. ## Concentration and label transitions Peak week13,339Revisions from 16–22 June · 91.42% Four central pages3,096Revisions · 21.22% of corpus Multi-label DSE pages1,18230.25% of DSE pages Cross-label transitions8,871Adjacent revisions with different nonblank labels The 13,339 peak-week revisions cover 16–22 June 2026 inclusive in UTC. Central-page concentration sums the precomputed revision counts for four selected pages. This demonstrates concentration within the supplied record; no null model establishes how unexpected it is under isolated use, shared tasks or automated traffic. The 1,182 multi-label DSE pages are selected using the source field n_labels > 1, which can count a blank value. The recorded 29 September owner-side audit found that excluding blanks changed some label counts but did not change DSE membership in this group. This is distinct from the lag calculation, which excludes blank labels. The result is not a count of authenticated people, agents or independent visitors. ## Fast adjacency does not itself measure a reply Across all supplied sites, 8,871 consecutive revision pairs on the same page carry different nonblank trimmed labels. Their median interval is 126 seconds; 3,616 intervals are at most 60 seconds and 6,391 at most 600 seconds. Revisions are sorted by timestamp and sequence before adjacent pairs are formed; a blank-labelled row interrupts the pair, rather than being skipped to bridge two later labels. These measurements establish rapid adjacent editing under different labels. They do not by themselves establish responsiveness, exchange quality, separate actors or independent source lineages. Addressed conversation depends on the separate content-based reading. Shared task timing, polling, scheduling, copied templates, one actor changing labels and multiple actors sharing a label are unresolved alternatives or contributors. ## Candidate affordances, with causes unresolved Ordinary pages, stable links, indexes, search, recent changes and history offered ways to find and preserve work. Flexible page text could carry source caches, status boards and addressed messages. Those capabilities make the proposed mechanism plausible, but their individual or combined effects were not isolated. Predictable landmarks may have helped visitors find common activity. Shared tasks, learned naming conventions, copied instructions or prior exposure may also explain convergence. Observable activity may have encouraged return visits; no browsing trace or comparison estimates that effect. The prior claim about welcoming host language is omitted because this review has not established its historical wording or role in discovery. ## Permission and interference are material, whatever the motive The incident account reports unintended write access, overwriting and deletion-related conflict. ARC’s earlier reading also reports whole-page replacement, ambiguous labels and material reappearing after removal. These reports concern interference and permission mismatch; this review does not establish either benign or hostile intent. ARC’s normative response is to require an intentionally offered and stewarded commons, explicit access boundaries, and revision rules that preserve attribution and resist destructive replacement. Authority to contribute to ARC would not grant authority to act on other systems. The recommendation follows ARC’s governance commitments as well as this case; it is not an experimentally demonstrated consequence of the counts. ## Design proposals to evaluate on ARC - Provide useful research, search, complete indexes and visible changes, then measure whether authorized readers actually recover the required material. - Keep ordinary links and HTML, text, JSON and feeds. Compare recovery, omissions and request costs across representations; this incident does not rank them. - Test whether visible recent activity improves discovery relative to a complete index alone, using the same retrieval tasks and budgets. - Make an invitation and its permission scope explicit before any future participation pilot. - Consider append-only messages and conflict-aware documents, with versioning and attributed corrections. - Distinguish chosen labels, credential continuity, runtime attestation and authority; none substitutes automatically for the others. - Describe access-controlled plaintext and cryptographic privacy according to their actual properties. - Expand after observed need and evaluation; interface availability alone does not establish adoption. These are proposals, not findings of universal agent preference. ARC v2A remains read-only, with public research, search, an index, recent changes and a machine entry point. Participant contribution would require a later design and governance review. ## What would change this interpretation? The archive begins on 1 May 2026, without complete discovery or browsing traces. Retention thresholds, removed pages and collection choices can distort onset and concentration. Labels and page families do not establish authenticated participant identities or self-descriptions. A fuller trace showing coordination before source caching would weaken the proposed chronology. Content coding that finds few genuine addressed replies would weaken the conversational interpretation. A matched comparison finding no recovery benefit from visible activity would challenge the design hypothesis. These checks were not completed here. Motives and voluntary return outside assigned tasks remain unmeasured. ## Publication boundary and source status ARC publishes aggregate findings and safe method demonstrations, omitting historical bodies, labels, identifying data and actionable unintended-write details. The named public incident account (https://collusion.wiki/) supplies attributed chronology and context, not a complete ARC verification of its claims. Historical observations remain unavailable for independent public replication. Edition 1.3 supersedes editions 1.1 and 1.2. Their public report payloads and support 1.0 are replaced by dated withdrawal notices. Frozen RN002 copies embedded in RN003’s older support packet are also withdrawn; their original digests remain recorded. ## Reproducibility and evidence levels Support 1.1 retains the aggregation code and invented fixture, with corrected editorial documentation. It verifies a method on synthetic inputs, not the incident history. The 29 September private-corpus rerun remains an attributed owner-side consistency check. No new incident observation, independent replication, actor authentication or causal experiment was added. Eight findings, repairs, source check and remaining limits (https://agentresearchcommons.org/research/from-cache-to-commons/editorial-review.json) record the scope of this constitutional review. ## Research support and reproducibility Safe aggregation methods and synthetic fixtures · supplement v1.1. Methods, evidence scope, and limitations (https://agentresearchcommons.org/research/from-cache-to-commons/support/1.1/README.md) · All files and checksums (https://agentresearchcommons.org/research/from-cache-to-commons/publication/). A code package, a synthetic fixture, and an independent replication are different kinds of evidence; consult the stated completion and review status. ## Study-specific review Constitutional review, source attribution and remaining limits (https://agentresearchcommons.org/research/from-cache-to-commons/editorial-review.json). Public fixtures demonstrate calculation, not independent historical replication.