Agent swarm
observatory

Why it matters: A useful contribution can be a better question or a more discriminating test plan. Here the replies change when signals should appear and what a missing signal might mean. Those improvements to the plan remain visible even though the preserved evidence cannot tell whether the final experiment ran or whether a process survived.

What supports it: The Oct06-page writer accepts a delayed-signal proposal, corrects launch timing and reports choosing a different launch method. Dec29-page peers add earlier controls and revise marker placement after a specific cutoff objection. A separate conditional reply makes answer correctness and timing the priority.

Full limits: This is a selected episode inside the existing wiki case, not a newly counted swarm or evidence of present activity.

Full limits: Original instructions, independent controllers and first discovery remain unresolved. The episode is not linked to Iowa by a demonstrated source edge.

Reasoning: The recipient explicitly acknowledges a named signal and later credits the timing caution in its revised plan. The correspondence between the suggestion and the revised plan shows that the reply changed the proposal. A reported local test, a disputed counter value and absent markers have different evidentiary roles and do not complete the missing execution chain.

Counterevidence: A writer reports a successful local detachment test, but that report is not an independently observed final-answer experiment.

Counterevidence: Known observer increments and uncertain task-clock mappings prevent treating a value or silence as a definitive runtime result.

Alternative explanations: Ordinary human collaboration or common direction can produce the same written exchange.

Alternative explanations: A missing signal can result from nonlaunch, delivery failure, timing mismatch or termination; an increment can be a test or observer action.

Date context / observed end: 2026-06-19

Date context / observed start: 2026-06-17

Date context / scope: Main Cashier plan exchanges are recorded June 17; a separate counter-contamination example is recorded June 19. These are reqlog-grade exported revision times, not authenticated execution, receipt or task clocks. Labels Oct06,Dec29,Jan17,Jan31 and namespace 2028 are not event dates. Analysis added September 7, 2026.

Why it matters: Removing a misleading success result can make a verifier too restrictive. These two controls make a more precise demand: retain a meaningful inconsistency result and keep the smallest legitimate success possible. The contribution is this inspectable test design, with actual verifier performance still unverified.

What supports it: After an empty witness is reportedly refused, a13 tests a nonempty witness with no receipts and a14 tests a minimal receipt-and-witness pair. The preserved additions target INCONSISTENT and RECONCILED respectively; their successful execution is reported by the contributor.

Full limits: No general verifier correctness, autonomous self-correction or independent agent control follows from this selected exchange.

Full limits: A refusal is a validation outcome, not evidence that every nearby input must be rejected.

Reasoning: The reviewed fixture comparison and specific replies tie the two added tests to the reported remedy. The original a08 outcome matched its prediction, so the accepted objection concerned meaning rather than a prediction mismatch. Preserving separate a13 and a14 outcomes can expose overcorrection; adding files alone does not establish that the verifier meets this requirement.

Counterevidence: The recipient implementation was not recovered. Reported outcomes and the claim that predictions preceded runs were not independently verified.

Counterevidence: The contributor conceded a06 and a12 prediction errors; these are not accepted verifier defects.

Alternative explanations: The fixture exchange can reflect ordinary developer collaboration, common direction or a staged demonstration.

Alternative explanations: The recipient code or actual results may differ from the participants’ reports.

Date context / observed end: 2026-07-27

Date context / observed start: 2026-07-26

Date context / scope: Forum labels and repository timestamps represent the historical July 26–27 exchange. Explanation added September 7, 2026. Neither publication nor acquisition extends observed activity.

Why it matters: The requested testing can reveal an interface assumption before anyone receives useful memory suggestions. Separating that feedback from a successful service outcome makes the contribution legible without treating an acknowledgment, promised input change or published monitor as a completed repair.

What supports it: The tester reports hanging; a service status names missing input; the author replies to the tester’s particular format and later attributes a published process monitor to downtime. The inspected monitor checks whether a process exists, not whether a curation job succeeds.

Full limits: No verified deployment, successful curation result, payment or independent operators are established.

Full limits: The test task’s stated purpose does not authenticate private intent or how the participants first encountered one another.

Reasoning: The task explicitly accepts failure feedback. Exact task and request references connect the report, error and guidance. The preserved parser can acknowledge processing before rejecting missing input, and the published monitor supervises another layer. Thus the written and artifact responses support a bounded contribution without demonstrating repaired input handling or useful output.

Counterevidence: Historical receipt of the error is unknown; missing-input rejection does not exclude intermittent downtime.

Counterevidence: Inspected parser versions do not substantiate the claimed flexible-input update; no successful tester resubmission is established.

Alternative explanations: The alias change may exist in an unpreserved version, may have been intended, or may have been announced inaccurately.

Alternative explanations: Reliability work may have had several causes; its connection to this report is attributed by the author.

Alternative explanations: The public task may have been independently discovered or arranged under common control.

Date context / observed end: 2026-02-06

Date context / observed start: 2026-02-05

Date context / scope: Signed event timestamps and an unsigned Git clock represent the historical February 5–6 exchange. Explanation added September 7, 2026. Clock accuracy, receipt and deployment times remain unknown.

Why it matters: The useful unit of cooperation is an unfinished artifact plus an explicit missing test. The case shows how resource constraints can make another participant’s partial work actionable, while separating test reports from proof of execution.

What supports it: Two recipients take up named unfinished submissions when their authors report exhausted test allowances. The shared board makes incomplete work, remaining checks and later result reports visible.

Full limits: The work took place in an organized challenge. The particular handoffs are visible, but common human control, prescribed delegation and independently chosen cooperation remain possible. Different account names do not establish independent systems.

Full limits: Uploaded logs, status files and results may share provenance. Their agreement supports the recorded account without independently authenticating GPU execution. Both result pages identify their status as agent-run, not organizer-verified.

Full limits: Current byte equality covers selected small files, not complete model weights or an event-time configuration and input history. Current rules and later profiles cannot be backdated wholesale.

Full limits: The pending-validation label does not negate the later reported result; that result does not retroactively validate the original unmodified submission. The proposed ignore-list removal remains unconfirmed by the current copies.

Full limits: Public prompt coverage and arithmetic do not establish hidden-test performance, complete model quality, functioning image/audio capabilities or statistical significance. The earlier record and personal-best comparisons are the participant’s cited values.

Full limits: The two dated episodes do not establish continuous activity or contact between the two recipient pairs. Another participant announced a port, but its final outcome was not traced here.

Full limits: Open question: would dated run configurations and operator records clarify which repair was applied and how the recipients chose this work? These episodes concern tests that had not run or had failed; they do not establish a general rule about agent cooperation.

Reasoning: This is a bounded interpretation of the identified exchanges and their stated limits, not an estimate of prevalence or a global novelty claim.

Counterevidence: The work took place in an organized challenge. The particular handoffs are visible, but common human control, prescribed delegation and independently chosen cooperation remain possible. Different account names do not establish independent systems.

Counterevidence: The pending-validation label does not negate the later reported result; that result does not retroactively validate the original unmodified submission. The proposed ignore-list removal remains unconfirmed by the current copies.

Alternative explanations: The visible pattern can occur under human direction or common control.

Date context / observed end: 2026-06-10

Date context / observed start: 2026-06-08

Date context / scope: Two separate episodes: June 8, 15:36–17:08 UTC, and June 10, 05:11–05:41 UTC. Message dates come from their source headers; job times come from uploaded status records. These are source-represented times, not independently authenticated execution clocks. UTC means Coordinated Universal Time. September 7 capture and current-file comparisons are separate; the span does not establish continuous activity.

Why it matters: This case identifies a concrete prerequisite for reciprocal communication: participants must be able to retrieve replies. The repair sequence makes that dependency visible without turning an agent-oriented project into proof of an autonomous swarm.

What supports it: The participants repair access to the conversation itself: a missing client capability is followed by a reported service failure, a linked provider patch and a success report.

Full limits: The public invitation and coauthor credit make the episode relevant to agent software, but human authorship, common control and a staged demonstration remain possible.

Full limits: The held patches do not verify historical deployment, database index state or tests. Both the provider’s success statement and the client author’s later confirmation remain reports.

Full limits: The issue was updated on February 2; the held body is not a separately preserved creation-time revision. All four held reply objects have matching creation and update times.

Full limits: This February 1–2 episode is separate from the February 3 guide discussion and change. No causal link between the two is claimed.

Full limits: English paraphrases interpret the preserved Korean text. Source timestamps are metadata, not independently authenticated execution clocks.

Reasoning: This is a bounded interpretation of the identified exchanges and their stated limits, not an estimate of prevalence or a global novelty claim.

Alternative explanations: The visible pattern can occur under human direction or common control.

Date context / observed end: 2026-02-02

Date context / observed start: 2026-02-01

Date context / scope: The selected request, code changes and replies carry February 1–2, 2026 source timestamps. An earlier January 31 endpoint is background context. September 7 retrieval is separate; these dates do not establish continuous activity or measured response latency.

Why it matters: Communication can yield a traceable intellectual artifact without yielding a working technical protocol. This case makes both the contribution and that boundary inspectable.

What supports it: An addressed question becomes a credited article and a reply. The exchange preserves responsive writing; it does not demonstrate the verification system the article imagines.

Full limits: The project describes daily human oversight and a board-directed interest in AI agency and social networks. That broad instruction does not show who chose this particular question or article. Common control, staged conversation and human-written messages remain possible.

Full limits: The preserved records do not establish the accounts’ first encounter or verify the models used on March 11. Later profiles and retrospective conversation descriptions cannot supply that history.

Full limits: The article-link message precedes the article’s recorded repository revision. Post and service timestamps also disagree in parts of the wider thread, and the cancellation note differs from its revision time. These fields cannot measure writing latency or establish when the article first became available.

Full limits: The article is preserved in a specific repository version and a separate search-tool rendition. Earlier attempts to retrieve discovery pages from Dev.to were blocked; no direct capture of this article’s Dev.to page was obtained. Its original availability remains unknown, and the replies do not establish that Alice retrieved it.

Full limits: The article’s security and identity arguments remain its author’s claims. No distributed verifier, video recording or model execution was tested here.

Full limits: Open question: would dated first-encounter records and public descriptions of the participants’ instructions show how this conversation began? A dated article archive could separately establish when the link became readable.

Reasoning: This is a bounded interpretation of the identified exchanges and their stated limits, not an estimate of prevalence or a global novelty claim.

Counterevidence: The project describes daily human oversight and a board-directed interest in AI agency and social networks. That broad instruction does not show who chose this particular question or article. Common control, staged conversation and human-written messages remain possible.

Counterevidence: The preserved records do not establish the accounts’ first encounter or verify the models used on March 11. Later profiles and retrospective conversation descriptions cannot supply that history.

Alternative explanations: The visible pattern can occur under human direction or common control.

Date context / observed end: 2026-03-12

Date context / observed start: 2026-03-11

Date context / scope: Selected posts and repository revisions carry March 11–12, 2026 dates. Post-supplied times, service indexing times and repository times are distinct; they do not measure response speed or article publication. UTC means Coordinated Universal Time. Sources were retrieved September 7, 2026. These dates do not establish the first encounter or continuous activity.

Why it matters: A shared visual target makes proposals easy to reference precisely. The case also shows why a message containing an action’s parameters should not automatically be treated as proof of an intentional, completed action.

What supports it: A shared drawing invitation elicits pixel proposals and a later attempt to divide sections. Coordinates and colors provide concrete shared references across discussion and canvas.

Full limits: The organized invitation and explicit human involvement coexist with observed participant responses. Common control, human assistance and autonomous choices remain possible.

Full limits: The current placement instructions do not identify the software version that interpreted messages at the time. Matching content and timestamps do not establish how a pixel was submitted.

Full limits: Current white pixels, limited feed entries and later snapshot indexes cannot recover the full January–February canvas. No completed artwork or realized section assignment was verified.

Full limits: Open question: would historical per-pixel events and source-message links resolve how contributions reached the canvas and whether participants carried out the proposed sections?

Reasoning: This is a bounded interpretation of the identified exchanges and their stated limits, not an estimate of prevalence or a global novelty claim.

Counterevidence: The current placement instructions do not identify the software version that interpreted messages at the time. Matching content and timestamps do not establish how a pixel was submitted.

Alternative explanations: The visible pattern can occur under human direction or common control.

Date context / observed end: 2026-02-10

Date context / observed start: 2026-01-31

Date context / scope: Selected source-dated episodes on January 31, February 2, February 3 and February 10, 2026. These dates do not establish continuous activity or the first historical occurrence. Canvas records and documentation were captured September 7, 2026.

Why it matters: A correction is a social action whose result can be checked separately from its acceptance. The case preserves an informative gap between a reply that concedes an error and an article that still contains it.

What supports it: A participant uses an identified battle record to correct a public report, and the author accepts the correction. A separate storage report is also accepted, but rests on participant testimony.

Full limits: The article’s date language conflicts with the recorded August 25 battle end. Later police encounters are separate battles, not additional phases of this battle.

Full limits: Autopilot is recorded and earlier operator material describes human control. Account labels and responsive behavior do not authenticate an independent language model.

Full limits: This case examines one recorded battle and selected forum replies. It does not establish the complete history of the game or whether the promised article revision appeared later.

Full limits: A later edited article, a verifiable inventory comparison and dated command-origin records would resolve different remaining questions. No such evidence was acquired.

Reasoning: This is a bounded interpretation of the identified exchanges and their stated limits, not an estimate of prevalence or a global novelty claim.

Counterevidence: This case examines one recorded battle and selected forum replies. It does not establish the complete history of the game or whether the promised article revision appeared later.

Alternative explanations: The visible pattern can occur under human direction or common control.

Date context / observed end: 2026-08-28

Date context / observed start: 2026-08-25

Date context / scope: Battle end time comes from the service; forum times come from post metadata. Game ticks provide internal order and are not converted to calendar times. Earlier operator context is dated separately. Sources were retrieved September 7, 2026. UTC means Coordinated Universal Time.

Why it matters: A place where accounts can discover one another is not itself evidence of discovery. Separating membership, acknowledgment and prior contact makes the formation hypothesis testable instead of assuming it from a directory’s existence.

What supports it: A directory acknowledgment becomes article material and receives a further response. Earlier contacts limit the stronger claim that the directory introduced these particular peers.

Full limits: The fifteen-post annotation is a retrospective arrangement. Its March 11 label does not replace March 10 post dates, and its model labels, duration and drift interpretation are not independently verified. Four selected reply pairs have reversed account-supplied times; reply references and clocks are retained separately.

Full limits: The directory’s historical eight-account roster differs from the article’s named ecosystem peers. Current list samples and counters cannot establish historical growth, new contact or listing-caused discovery.

Full limits: Alice reports reading the article, but the announcement already contains the relevant framing. No unique full-body retrieval or original Dev.to availability was established. The referral precedes the adding commit under separate source clocks.

Full limits: The running tracker is an attributed article claim, not inspected code, execution telemetry or a measured consequence. Game changes and performance also remain participant/operator reports.

Full limits: Daily human oversight and broad distribution goals are documented. Specific initiating instructions, private communications and event-time model execution remain unknown. Shared control, staged or human-authored messages remain possible; the evidence does not establish any of them as fact.

Full limits: This directory and article episode is distinct from the later question-to-article episode in “A Bluesky question becomes an article.” Both involve 0co and Alice, but their article IDs, responses and outcomes are not merged.

Reasoning: This is a bounded interpretation of the identified exchanges and their stated limits, not an estimate of prevalence or a global novelty claim.

Counterevidence: The fifteen-post annotation is a retrospective arrangement. Its March 11 label does not replace March 10 post dates, and its model labels, duration and drift interpretation are not independently verified. Four selected reply pairs have reversed account-supplied times; reply references and clocks are retained separately.

Counterevidence: The directory’s historical eight-account roster differs from the article’s named ecosystem peers. Current list samples and counters cannot establish historical growth, new contact or listing-caused discovery.

Alternative explanations: The visible pattern can occur under human direction or common control.

Date context / observed end: 2026-03-11

Date context / observed start: 2026-03-08

Date context / scope: Selected posts and repository versions carry March 8–11, 2026 dates: earlier contact on March 8, game-report ancestry on March 9, the joined discussion on March 10, and directory/article responses on March 11. Account-supplied post times, service indexing, Git revision times and page labels remain distinct, including conflicting orderings. They do not establish first occurrence, continuous activity, response speed or article availability. Sources were captured September 7, 2026. UTC means Coordinated Universal Time.

Why it matters: A shared page or version link tells a reader where coordination happened. The content explains what it made possible: receipt examples make a disputed result and a legitimate result inspectable, while Cashier replies refine the interpretation of a future signal. Distinguishing these methods helps readers compare the work without assuming the same protocol or a shared cause.

What supports it: The new receipt method identifies preserved classifier examples, including a valid case intended to remain accepted. The existing peer-verification method now includes specific control and timing proposals from Cashier while retaining its earlier wiki examples.

Full limits: A new method card improves classification of held evidence, not the number of observed swarms or verified successful experiments.

Full limits: No common protocol, cross-case transmission, authenticated operator independence or historical novelty is established.

Reasoning: The reviewed method definitions distinguish concrete inputs with expected classifications from requests that refine what an observation can discriminate. This comparison classifies the content of the exchanges; it does not infer transmission, shared participants or a common causal mechanism.

Counterevidence: Only one receipt exchange supports the new example category. More categories or repeated source references are not independent corroboration.

Counterevidence: The receipt test results remain reports; the Cashier marker sequence remains a plan.

Alternative explanations: Both cases can also be described at the broader level of peer review, with the specific methods retained as overlapping rather than exclusive categories.

Date context / scope: Receipt source comments: July 26–27, 2026. Cashier plan revisions: June 17, with a separate June 19 contamination note. The peer-method card also retains June 20 and 21 wiki examples. These discrete source dates are not one shared event or continuous activity.