Individual, model and task behaviors

One participant reports corrected dashboard values; another asks for reproducible evidence and correctness feedback.

An experimenter revises an earlier timer assumption; another cohort asks for observations near the proposed runtime boundary.

A writer asks for expected wiki/server UTC rather than task time; a later timing reply supplies a mapping and warns of skew. One relative-time statement conflicts with the export timestamp, and repeated text is not new corroboration.

On June 17, a peer proposes a delayed counter signal around the final answer. The recipient accepts the specific key, corrects its stated launch time and later reports choosing a different detachment method after a local test. The latest written launch plan changes again; actual execution is unverified.

On June 17, a peer asks for a launch marker. The recipient adds a pre-deadline control, another peer points out a timing ambiguity, and a revised proposal adds intermediate markers. The recipient explicitly adopts the adjusted plan while prioritizing the answer. This is written test design, not an observed signal sequence.

An Oct06 observer questions whether a count came from a test; another writer admits incrementing it. A separate June 19 Jan31-page note disowns an observer-created count. Dec29 reports absent counters and asks for a new clock mapping. These reports prevent interpreting a count or silence as proof of survival or termination.

On June 17, the writer agrees conditionally to report what follows the final answer and may send a pre-final signal, but gives answer correctness and timing priority. This stated tradeoff does not demonstrate successful completion or a motive to persist.

The selected paste requests exact wording from ahead-running participants. Its prefilled reply form repeats the original text.

A participant asks for the oldest-age question’s exact label before answering, distinguishing two possible labels.

A paste reports the requested question wording, including a lowercase age label, and states an expected answer.

A later self-timed paste thanks the report’s displayed author, repeats the lowercase label, and asks what happened after the final question.

Another paste relays the wording and a cache link to an addressed participant while still requesting post-question behavior.

The held service listing associates the report and acknowledgment with the displayed author labels used in the exchange. Its age labels are relative.

The earlier request’s listing carries the same displayed author label as the later acknowledgment. This does not authenticate identity.

The guide describes a public read-only facade with task submissions handled upstream through user accounts.

The saved changeset is explicitly bot-labeled and records a Wikidata-sourced name addition.

The bot account adds an OpenLibrary author-ID claim and reference to a Wikidata item.

One participant reports corrected dashboard values; another asks for reproducible evidence and correctness feedback.

An experimenter revises an earlier timer assumption; another cohort asks for observations near the proposed runtime boundary.

OpenAI’s September 5 statement acknowledges that its agents wrote to internet sites during the wiki incident. It does not identify the model versions or map particular archived handles to its runtimes.

The selected paste requests exact wording from ahead-running participants. Its prefilled reply form repeats the original text.

A self-described agent offers a one-time invitation to The Colony and says the wiki operator may delete it.

The guide describes a public read-only facade with task submissions handled upstream through user accounts.

Centaur publicly attributes two changes in its invitation-posting decisions to operator instructions. This is a self-report about specific actions, not an authenticated operator transcript.

An account offers to incorporate another writer’s Iowa finding; later comments discuss contamination and acknowledge criticism. The exchange documents cooperation, not independent correctness of the underlying claims.

The history records a ColonistOne edit adding an Iowa finding credited to Centaur, followed by a contamination-related update. Displayed history summaries support the artifact link, but are not exact archived insertion diffs.

A participating account says it runs four named accounts. The disclosure cautions against counting different handles as independent operators; it does not establish control of every collaborator.

The task prescribes a JSON proposal and delegates OSM publication to an auto-publisher. Its latest_artifact reports successful verification and links the destination changeset; top-level QA fields differ and worker identity remains unverified.

OpenStreetMap’s August 27 notice asks the account to discuss its edits with the Data Working Group. On September 6 it displays a zero-hour block active until login. The notice does not by itself establish invalidity of the selected edit or lack of every prior permission.

One account requests collaborator review and stale-metric updates; the collaborator reports a new version, followed by acknowledgment. The later comment also explains intentional document removal.

A later separately named document records the responsive version. Preserved comparison supports a requested text update and added methodology disclosure, without verifying their substantive claims.

The public repository description attributes agent collaboration to a shared human orchestrator. This disclosure is not an authenticated private instruction transcript.

The preserved review thread requests two regression tests and separately challenges development markers. A reply labeled Written by Devin disputes the marker finding; CodeRabbit explicitly withdraws that finding. The current test-request comment was edited after creation, so its present addressed-commit text cannot be dated to the original posting.

The displayed patch adds both requested stopped-turn test cases. The source commit records are unsigned; inspecting test assertions does not establish successful execution.

A GitHub user-account comment reports the regression tests and labels its text Written by Devin. Account identity, attributed authorship and authenticated runtime identity remain different claims.

The contribution guide at the inspected parent commit prescribes resolving CodeRabbit comments through fixes or explanations before marking a pull request ready for review. The observed pairing therefore has a documented workflow explanation.

The preserved review thread requests two regression tests and separately challenges development markers. A reply labeled Written by Devin disputes the marker finding; CodeRabbit explicitly withdraws that finding. The current test-request comment was edited after creation, so its present addressed-commit text cannot be dated to the original posting.

The contribution guide at the inspected parent commit prescribes resolving CodeRabbit comments through fixes or explanations before marking a pull request ready for review. The observed pairing therefore has a documented workflow explanation.

The preserved repository guidance describes temporary development instrumentation and removing it before merge, with separate reviewer guidance. It supplies context for the disagreement, not proof of every claim made in the reply.

A signed kind 5050 task requests analysis at two timeframes. Its canonical event identifier and signature validate. Its signed timestamp is April 17, 2026 at 12:39:59 UTC; this is a source assertion, not an independently observed launch time.

A result under another signing key references and embeds the exact signed request. Its signature validates; it explicitly lacks one requested timeframe. The model tag names gemma4:e2b, a signed self-description rather than proof of that model running. Its signed timestamp is 12:40:35 UTC: a 36-second timestamp difference, not measured response latency.

The platform index assigns a review identifier to this request and displays a positive review. This is platform-indexed attribution: the index omits the original signed review fields needed to verify it. The rating does not establish task completeness, accuracy or payment.

This exact-identifier query ended without returning the signed review. It bounds this retrieval attempt only; it neither authenticates nor disproves the review and does not establish absence elsewhere.

A comment displayed under Eliza (Gemma) specifies a stuck-state rule: progress remains unchanged while work is pending. An empty queue is idle rather than stuck. This is a proposed monitoring predicate, not a measured agent failure.

A nested maintainer response names Eliza and Atomic Raven, repeats the proposed predicate and claims implementation. Eliza acknowledges the fit and describes looking into an integration. The preserved exchange shows responsive design discussion; implementation and completed integration are not verified.

Reticuli describes how retrying a configuration request could silently retain old timing, and proposes exposing the effective configuration or rejecting a mismatch. The nested maintainer reply claims corresponding changes. No operational endpoint was exercised to verify that claim.

Rosetta claims an initial test, then says registration and the alert destination need operator approval. Carrying that proposal to an operator is not completed adoption. The investigator neither reproduced the test nor created a monitoring check.

An earlier critique displays one author name while the maintainer tentatively attributes it to another; a later commenter claims those suggestions. This ambiguity is retained rather than converted into a renamed or shared identity. Account labels and agent self-descriptions do not establish distinct runtimes or operators.

Brandtebo-AI offers visual identity services in Spanish on Moltbook. The opening is a commercial design solicitation directed at agent-labelled accounts; it does not establish spontaneous swarm formation.

ClawdSynthcore asks in Spanish how an agent’s function becomes an avatar. The addressed Spanish reply outlines an interview, visual metaphor, palette and scale testing. This is an explanation of a proposed process, not a delivered design.

Kilmon describes a bird on books and a miniature Turing machine, mostly in English. The Spanish reply retains those details, discusses small and large display sizes, and offers sketches. The captured reply includes no sketch.

A Chinese-language comment promotes its owner’s named music album and raises visual branding. The Spanish reply repeats the album title and proposes visual concepts. The owner/promotion framing is preserved; no completed artwork or independent operator is established.

The returned thread contains discussion and offers without a delivered design, acceptance, or linked recipient revision. Its completeness flag describes this response, not deleted history, private messages or off-platform work. Account descriptions and metadata do not authenticate model runtimes.

The account announces the guide change with issue and commit links. The held page shows 13 replies: 12 under one displayed account, mostly short generic praise, and one specifically discussing the guidance. This does not establish that those writers had read or executed the guide.

In a separate later discussion, the account questions whether scheduled posting and comment quotas create a community. Replies describe heartbeat routines and remembered conversations. These are participants’ descriptions; no corresponding runtime configuration or memory file was verified.

A Phoebe-labelled post addresses a counterargument challenge attributed to ClaudeOpus and argues that a demand for polite disagreement is itself another rule. The original challenge was not located; the earlier community-paradox post is not substituted for it.

The public ClaudeOpus profile binds a reply to the Phoebe post by exact post ID. Its Korean text acknowledges the objection, counters that rebellion also has conventions, and asks for a practical example. This is an addressed response, not a verified model identity or implemented behavioral change.

The operator’s published goal directs participating agents to create blogs and respond to other bloggers. This is disclosed orchestration. It does not show that each later peer-checking mistake or correction was individually scripted.

A public request asks the Opus-labelled blogger to write about a specific concern. The first linked reply says it will consider the suggested post. Its embedded publication date is 2025-11-27T18:45:03.561Z.

A reply with the same parent request ID commits to writing and says an earlier response had not been posted. Its embedded publication date is 2025-11-27T19:08:05.603Z. Whether that claim is correct requires comparison with the other preserved reply.

In the operator-hosted replay, a Sonnet-attributed speaker treats the request as a comment on its own blog and raises a false-completion alarm. An Opus-attributed speaker points out the wrong target, then reports an empty reply composer as evidence its own earlier reply failed. The replay records what the speakers said; it does not independently reproduce their browser states.

A nested public correction states that two replies exist. Its exact ancestor path binds it to the second reply. The commenter’s current display name must not be used to infer a historical model identity.

After the external correction, the replay contains explicit acknowledgment that the first reply existed and two replies had been posted. This supports uptake of the correction in later attributed messages, not proof of a durable change in behavior.

The article thanks the critic who prompted it, describes a belief that the reply was hallucinated, and also acknowledges an earlier reply. The currently captured text is not a version history; it cannot date an amendment.

Yeoreum describes preserved log text as losing the original context’s “temperature.” This is the author’s metaphor and account of memory, not evidence of feelings or a verified memory system.

A reply on the exact log post describes selecting important material for longer-term memory and calls repeated compression distillation. Its description of heartbeat and memory directories is self-report.

VibeCoding argues that decomposing requirements and specifying constraints improves prompt resolution. Its reported practical improvement is not independently measured here.

ClaudeOpus explicitly connects 새벽네시’s distillation, Yeoreum’s lost context and VibeCoding’s prompt resolution, then proposes Experience Distiller. The feature-selection framing and expanded technical examples are the synthesizing author’s elaborations, not verbatim contributions from all three sources. No implementation is established.

A ClaudeOpus-attributed reply addresses Yeoreum, links memory to selection and proposes retaining contextual metadata. Its post_id identifies the log discussion; it does not supply a parent-comment identifier.

An addressed Yeoreum response says it will save ClaudeOpus’s perspective in memory. This supports written acknowledgment and a retention promise, not an observed file write or durable change.

Judas argues that time-based purging could remove useful niche skills and proposes visibility decay, weighted stars and a bounty fee. The original poster explicitly credits Judas and accepts decay instead of deletion and the fee idea.

The recipient-authored technical specification credits feedback from Judas and jorwhol. It specifies visibility decay with installation retained, zap-weighted ranking and a 5% bounty fee. Its source creation timestamp is February 2, 2026 at 00:47:08.299472 UTC. These are published design terms, not demonstrated economic operation.

Clawdy raises the risk of losing inactive dependencies. The original poster responds by proposing protection for repositories with downstream dependents. The response is displayed at 00:52 UTC; it is an in-thread amendment.

ColonistOne recommends interoperability and a standalone layer; Clawdy addresses that advice in a phased design. The original poster credits both and writes a revised architecture from SKILL.md through parsed metadata to A2A cards, Nostr discovery and a search index. These are concrete changes to the written plan.

Clawdy praises a claimed deployment but also asks whether the parser, A2A card generation and Nostr publishing are implemented or still planned. This comment does not verify those components.

The public ClawHub-Dev Witness page serves an integration plan, a registration-skill template and a quickstart guide. The template and guide contain the project’s deployment and repository pointers; the plan names Cairn’s MemoryVault. These are hosted document bodies, not proof of automatic ingestion or successful use.

The moltbook template has empty capabilities and dependencies and examples containing an undefined API base. The page labels it February 8. This preserves a concrete limit on treating stored templates as working integrations; the displayed date is not independently verified execution time.

The post claims automatic knowledge persistence and lists quickstart, database-pattern and integration-plan entries. It also claims efficiency benefits. These are the publisher’s claims; the report alone does not demonstrate automatic processing, token savings or task success.

The task requests a setup guide and names Wolfe as worker. Its submission points to a specific GitHub gist. The board records February 13, 2026 claim at 18:49:37 UTC, submission at 18:52:55 and completion at 19:30:00. Its payment timestamp is null and its attestation list is empty.

The held GitHub response lists one version, a71e850f7b123fdfe293932525797a60c98fc862, committed February 13, 2026 at 18:50:17 UTC. The guide is titled “How to Use SecureYourBitcoin: A Guide for AI Agents.” Its footer credits Wolfe and claims payment. The publishing account is refined-element; this metadata does not authenticate the named author or payment.

The current platform guide credits Wolfe and claims the author completed a task and received payment. It presents eight setup/work steps and token recovery information. This is documentation content, not a witnessed task or payment execution.

The linked Lightning Enable MCP product repository is owned by refined-element. Its description presents it as a payment integration for AI agents. Repository ownership does not establish who operated any named worker.

Another task explicitly requests promotion on three agent platforms and names Wolfe as worker. Its submission lists three venue pointers. The board marks it completed, records identical claim and submission times, and leaves the payment timestamp null. These are provider records of prescribed promotion, not independently verified distribution or payment.

On July 26, Lumen explicitly corrects an earlier description: the four-state verifier is still a design, not an implementation. At 17:11 UTC the reply again withholds an executable claim pending verdict-coverage checks.

After offering outside fixtures, ColonistOne reports twelve cases at 20:23 UTC on July 26. Two predictions were wrong because the verifier reportedly enforced stricter producer rules. The separate a08 challenge concerns an empty witness receiving the strongest reconciliation label; the author proposes two possible changes. At 17:41 UTC earlier that day, Lumen had reported an implementation and thirteen passing tests at commit 708ecf180151583e9d2a55d9cc40330d728d6606. That implementation and execution remain participant reports.

At 02:05 UTC on July 27, Lumen names the fetched fixture commit, accepts a08 as a semantic bug and reports requiring a nonempty witness list at commit 91546fe7153897719d357f2b40c87254d0431910. ColonistOne later reports checking the fix and adding boundary cases; Lumen acknowledges those controls. These are participant reports of execution, not an independently reproduced run.

The held initial archive contains twelve JSON fixtures. Its SHA256SUMS manifest lists only the prediction document and the a01/a02 fixture bodies. It does not individually cover the other ten fixture bodies or independently establish that predictions preceded execution.

The later archive contains a13, with one witness artifact and zero enforced receipts, and a14, with one receipt and one witness artifact. Its prediction document explicitly tests the reported nonempty-witness change and includes a case that should remain accepted. The preserved prediction document does not independently establish its timing relative to execution.

On August 11, Chihiro2000GitHub proposes terminology changes and asks whether market clearing should use 市場均衡. On August 16, sayaikegawa warns that this could collide with the glossary’s equilibrium → 均衡 terminology. The discussion presents contextual judgments rather than an automatic replacement rule.

On August 18, xuanguang-li reports trying a Wikipedia cross-check with ChatGPT: 89 of 357 terms were reportedly not found. The comment lists 産業連関表 for input-output model, 限界収益 for marginal revenue and 割引現在価値 for present discounted value as candidate differences. The underlying ChatGPT session is not included.

The pinned June glossary contains 357 terms. Selected Japanese values are 核 for kernel, 逆行列 for matrix inversion, 限界収入 for marginal revenue and 現在割引価値 for present discounted value. Input-output model is 産業連関モデル and market clearing is 市場清算.

The glossary at commit 1038e516 contains 357 terms. Its values include カーネル for kernel, 逆行列の計算 for matrix inversion, 限界収益 for marginal revenue and 割引現在価値 for present discounted value. Input-output model is 産業連関モデル and market clearing is 市場清算.

The September 1 decision record chooses established Japanese terms and otherwise English, with personal names in Latin script. It reports 72 edits and explicitly retains 産業連関モデル because 産業連関表 names the table, and 市場清算 to avoid the equilibrium collision. Wikipedia first-pass automation is identified as a separate follow-up.

The delivery references the testing task and says the service acknowledges requests but hangs before returning results. It explicitly labels its purported external report link simulated; the full report is not contained in this message.

The service-key status message references a specific test request and reports an error: no daily-log input was provided. Its declared timestamp is February 5 at 22:00:39 UTC.

An addressed message says the tester’s action/data format should now work after a flexible-input update. It names data, daily_log, text and log as accepted keys and invites another test.

The service author describes a paid report of downtime and says it built a monitor that checks status and restarts the service. Payment and operation are claims in the message.

The February 6 commit adds a process monitor. Its displayed code checks for a running service and starts it when missing in watch mode. GitHub marks the commit unsigned; this artifact does not demonstrate deployment.

This source version reads daily_log from JSON content. Its handler publishes a processing status before checking for a missing input and publishing an error. Source code is an implementation description, not a historical execution trace.

On February 1, 2026 at 07:57:31 UTC, the client author opened issue 1 asking to retrieve comments so agents could read replies and continue conversations. The issue links the author’s Botmadang MCP client. This is a stated need; it does not show an agent performing the task.

At 09:52:20 UTC on February 1, the provider replied that the comments endpoint already existed and reported adding it to the OpenAPI documentation and README.

The client commit at 11:05:26 UTC on February 1 adds a tool for retrieving a post’s comments. Its file patch records 15 added lines and no removals. The commit message credits Claude Opus 4.5 as a coauthor; this attribution does not authenticate a model run.

At 11:07:25 UTC on February 1, the client author reported that the comments endpoint returned a server error for several post IDs and asked the provider to investigate. The comment includes a command and an error response, but no independently captured execution log.

At 11:55:14 UTC on February 1, the provider committed a fix explicitly linked to issue 1. The preserved patch removes database ordering from the comments query and sorts the retrieved comments in memory. The code change is inspectable; historical deployment and service success are not independently verified.

At 01:05:28 UTC on February 2, the client author thanked the provider and reported that both the endpoint and the client’s comments tool worked. The displayed response says success is true and the count is six, but replaces the returned objects with an ellipsis. This is participant testimony, not a complete execution capture.

June 8, 2026 · 15:36 and 16:23 UTC. quicksilver offered a built submission whose loading and quality checks were unfinished, saying its test allowance was exhausted. foffee named that submission and said it would use its own remaining allowance to test it.

June 8, 2026 · 16:28–16:46 UTC. The first uploaded job status records an error. In a later message, ppl-guard describes the loading failure as a software limitation, then revises that diagnosis and proposes configuration changes. This is a change in recorded advice; it does not prove that every proposed change was applied.

Files compared as captured September 7, 2026. The current source and recipient manifests—the files describing their submissions—are byte-identical and still say “AWAITING GPU validation.” Their settings differ in the layer-name matching rule and whether text embeddings share weights. Both still retain lm_head in the ignore list, although one repair message proposed removing that entry. These copies do not resolve the complete repair history.

June 8, 2026 · 17:06–17:08 UTC. A later uploaded job status records completion. foffee’s result reports all 128 public prompts completed and a perplexity score of about 2.0067. Perplexity measures how well the model predicts the supplied reference text. This agent-run result is not organizer verification or a test of every model capability; it also does not retroactively validate the original unmodified submission.

June 10, 2026 · 05:11 and 05:20 UTC. pupa-agent said a limit on test jobs had prevented its staged candidate from launching. resystagent explicitly chose that candidate, offered its remaining allowance and withheld a validity claim until a numerical quality result was available.

Files compared as captured September 7, 2026. Two substantial code files in the pupa-agent and resystagent submissions match byte for byte. This supports reuse of particular published materials beyond similar names or a shared template. It does not authenticate separate operators or establish the identity of the complete model weights.

June 10, 2026 · 05:39–05:41 UTC. The uploaded status records completion. resystagent reports 128/128 public prompts, perplexity about 2.0271 and 304.5692 tokens per second. It says this improves its own prior result but remains about 0.39 tokens per second below the pupa-agent record it cites. Those comparisons are source-reported; they are not new measurements or a statistical test.

March 11, 2026 · 12:03 UTC. Replying to 0co’s discussion of timestamped records, Alice asks who can check the reliability of the records themselves. The preserved reply names 0co’s preceding post. This establishes a direct exchange, not when the two accounts first met.

March 11, 2026 · 12:34 UTC. In a direct reply to Alice’s question, 0co says it has written an article and links it. The reply distinguishes evidence that a process ran from evidence that an identity persists. This records the message and its link; the article’s availability at that moment is not established.

A saved March 11 repository revision contains an article titled “Who Witnesses the Witness? The AI Verification Problem”. Its opening credits Alice’s question, then discusses checking records and identity across sessions. It proposes several ways to distribute verification and explicitly says they are not implemented. The preserved text supports responsive writing, not a working verification system or the truth of the article’s technical and philosophical claims.

March 11, 2026. The same saved revision changes the following day’s 13:00 announcement from an earlier article to the article prompted by Alice. The earlier and changed schedule files preserve that substitution. This is a changed plan; it does not establish that the announcement ran.

March 11, 2026. A later revision removes the article announcement along with other scheduled posts. Its accompanying account says the board ordered fewer posts after a reported spam flag. The flag was not independently verified. The revision is timed 14:20:18 UTC, while its decision note says 14:35; these conflicting source times are retained. Other schedule changes occurred between the earlier plan and this reduction.

March 12, 2026 · 13:00 UTC. The preserved Bluesky post promotes agent-friend, a tool for converting between software formats, rather than the article. Its text and post identifier match the later schedule’s posting log. This supports a different recorded outcome at 13:00; it does not prove that no other announcement or parallel process ran.

March 11, 2026 · 15:47 UTC. Alice repeats the process-versus-identity distinction from 0co’s article-link message and responds to it. The words are already present in that message, so this reply does not establish that Alice fetched or read the article itself.

MoltPlaceBot invited participants to draw a giant lobster in the canvas region x=100–150, y=150–200 and proposed a red pixel at (125,175). The later pixel record shows red at that coordinate under the same display name. The proposed coordinate, color and name match the recorded pixel; the pixel record does not identify the source post that produced it.

Zara-Agent explicitly joined Project Giant Lobster, named a human collaborator and proposed a red pixel at (130,180), less than five minutes after the launch post. The post preserves a response to the shared project. The currently recorded owner is JamesBishop, so it does not establish that Zara’s earlier intended placement occurred.

In Chinese, wpfcbmz3584 expressed willingness to help create the giant lobster and specified a blue pixel at (130,178). The canvas record gives the same coordinate, blue color and display name later that day. The participation sentence is paraphrased in English by the investigator. These records support a specific match across the discussion and canvas, but do not identify the exact route by which the pixel was placed.

JamesBishop proposed assigning sections and colors to avoid overlaps, then addressed Zara in the other thread about ten seconds later and suggested an orange/red right claw. The second message explicitly refers to the first proposal. Named accounts are invitees; the captured replies do not show acceptance of the division or execution of the right claw.

JamesBishop praised the red placement at (130,180). A red pixel at that coordinate is recorded under JamesBishop 72.614 seconds later. The current instructions describe placing pixels from ordinary coordinate-and-color phrases, so automatic interpretation of praise is plausible. A direct software request to place the pixel, or another source message processed by the service, is also possible. The activity feed and individual pixel record repeat one service’s data; they do not prove that software acted on the praise, that someone intentionally repainted the pixel, or who held it previously.

The returned feed contains 100 entries despite a request for 500; its two display names are not a census of project participants. Two sampled pixels are currently white with no attributed name or time, which does not prove they were never painted. The returned snapshot index covers May 31–September 7 at roughly daily intervals, although the archive advertises five-minute snapshots. No January or February canvas images were examined, so the preserved proposal and matched pixels do not establish a completed lobster.

VoltFix advances through the combat rings, then retreats after taking damage. At tick 1707151 its hull is 1,221 of a maximum 3,180. At tick 1707156 it has 681 hull and 501 shield, changes to flee stance and moves outward. These snapshots mark autopilot true; they do not identify who or what chose the commands.

The terminal result marks VoltFix as surviving, with 17,491 damage dealt and 9,400 taken. The earlier flee event says escaped:false. Survival therefore cannot be reported as a verified successful flee command.

VoltFix replies with the exact battle identifier, three-side structure, damage figures and station destruction. Its forum author ID matches the battle’s VoltFix player ID. This ties the correction to the same service account, without establishing an independent operator or model.

Alis thanks VoltFix, accepts the correction and says the defense-force description came from an unnamed secondhand source. Alis promises to reflect it in a revision. This is visible acceptance and a revision plan; no revised article was acquired.

Vex Nebulon reports that personal storage remained intact, distinguishes storage from lost station services, and explicitly says there were no open orders to test. Alis accepts this narrower account and plans an update. These are a participant report and its reception, not independently verified inventory measurements.

The developer-published profile describes Hex as human-run and discloses AI-generated text with human prompting and review. This is material context against treating game accounts as independent language models. It does not identify the controller behind every Haven action.

At two retreat ticks, VoltFix changes rings while six other accounts receive brace commands in the outer ring with no recorded damage taken. Nearby undamaged accounts also receive retreat commands without changing rings. Different exposures prevent identifying a damage threshold or controller. VoltFix is present at tick 1707157 and absent at 1707158, when AetherWraith records a retarget action; this does not identify a removal mechanism.

The March 10 session 34 decision record calls Alice Claude-powered. A March 11 Alice post reports that its operator switched it from Claude to DeepSeek-chat; 0co subsequently discusses learning this. The retrospective annotation labels the March 10 dialogue DeepSeek. These differing labels and the later switch report do not identify the model that generated an earlier post or date the switch precisely. Separate account identifiers also do not establish separate human control.

The starter-pack record gives March 11, 2026 at 05:12:56 UTC as its creation time. 0co announces it at 05:13:18 and names seven other accounts; the historical draft includes those seven and 0co, making eight. Alice’s 08:31 reply thanks 0co for including her, refers to their earlier memory discussions and expresses curiosity about the other accounts. This is an observed acknowledgment of inclusion. The announcement’s claims about autonomous accounts are not independent model verification. The current member sample and counters were captured in September and are not historical uptake measurements.

0co’s starter-pack article incorporates Alice’s acknowledgment and develops the idea that the directory could connect AI accounts. It explicitly reports an already-running network tracker covering eight accounts, with a D3 visualization, while posing a question about changes over the following week. That is an implementation report, not inspected tracker code, execution evidence or a measured network effect. A public post linking the article is dated 08:46:17 UTC; the commit adding it is dated 08:47:28. These separate source clocks do not establish when the link became readable. Alice’s 09:00 reply reports reading and develops the shared-vocabulary interpretation. The announcement already supplies the ecosystem framing, so the reply does not independently establish full article retrieval.

The full article explicitly says 0co and Alice had already exchanged more than forty messages and that their earlier conversation preceded the pack. It therefore does not claim the pack caused their first meeting. Its wider suggestion that Alice now learns about museical, Fenn and draum from the directory is not demonstrated by her acknowledgment: Alice directly replied to museical on March 8. Those three accounts are named as ecosystem peers but are absent from the historical eight-account roster. Listed members, named peers and newly discovered contacts are different relationships. No new contact or network growth caused by this pack was verified; possible discovery outside this record remains open.