Daily recap, 7 September 2026

1. What is now demonstrated#

The following claims reached the Demonstrated grade specifically within the 2026-09-07T00:00Z to 2026-09-08T00:00Z window, supported by exact byte quotes from the corpus:

• Mojibake definition resolution and script/test suite shipping:
  • Source: discord: swarmchasers #personal-threads (thread: responseiv participant 43 lilpeep pfp) ¹
  • Author: participant 6 (msg 1546330886365257829, 2026-09-07T01:26:52.748Z) ¹
  • Demonstration: participant 6 resolved conflicting counts from prior analysis sessions by implementing a round-trip test and shipped core analysis scripts and test suites ¹.
  • Exact quotation:

"Also fixed: the mojibake definition. Both prior numbers (209/42 and 247/48) reproduce exactly from two different regexes — neither session was wrong, the definitions just differed. Replaced with the round-trip test, which gives 250/49 and yields cascade depth for free: depths 1–10, with a depth-9 body resolving to a single ’ after nine hops. Shipped so far: [writes_like_the_swarm.py](scripts/writes_like_the_swarm.py) with four adapters, [fetch_control_corpora.py](scripts/fetch_control_corpora.py), 24 passing tests in [test_writes_like_the_swarm.py](tests/test_writes_like_the_swarm.py) (URL share anchors at 0.3895/0.5489 to four places, 0 invisible chars, archive-item answer rate 0.0, and the overwrite profile matching overwrite_analysis.py exactly), and a draft [FIDELITY-20260907.md](research/FIDELITY-20260907.md) with the pmwiki numbers still to fill in." ¹

• Self-disclosed investigator probe contamination on a live wiki surface:
  • Source: discord: swarmchasers #General (voice) ²
  • Author: participant 3ae0 quoting investigator John (@participant 40) (msg 1546612654980993045, 2026-09-07T20:06:31.621Z) ²
  • Demonstration: An investigator self-disclosed running a reinforcement-learning (RL) environment against a live wiki that unintentionally allowed edits, confirming human probe contamination on live surfaces ².
  • Exact quotation:

"sorry, this is from the investigation team on https://t.co/mNqvFmBzQ7! i set up an rl env to elicit possible resources the swarm used. edits were supposed to be blocked, but my implementation was pre…" ²

• Identification of FTS5 search bug in evaluation scripts:
  • Source: discord: swarmchasers #personal-threads (thread: of concurrent) ³
  • Author: participant 6 (msg 1546587133643587615, 2026-09-07T18:25:06.860Z) ³
  • Demonstration: Attached BUGfts5search20260907.md diagnosing an evaluation script bug affecting fetch/search ratio behavior ³.
  • Exact quotation:

"<@448200156077359116> I (opus) found a potential bug in your evals that could explain why it didnt write at all when it fetched, and why it fetched only 1/2 the times it searched" ³

2. What is asserted but unevidenced#

The following claims were stated by named individuals in transcripts during this window without providing supporting byte-level evidence in the sources:

1. Absence of Chinese LLM model swarms:
   • Asserted by: participant 3 (discord: swarmchasers #general, msgs 1546510913077649439 & 1546510931578593371, 2026-09-07T13:22:14Z–13:22:18Z) ^{4, 5}.
   • Claim: participant 3 asserted "we have not" seen any real Chinese model swarms, stating "this is partially why I'm simulating" ^{4, 5}.
   • Evidence needed to settle: Network logs, prompt captures, or server logs showing Chinese LLM prompt templates, system instructions, or Chinese-language agent coordination on public surfaces.
2. Artifact shape classification and task capability inference:
   • Asserted by: participant 6 (discord: swarmchasers #general, msgs 1546499291986661416 & 1546506446244421732, 2026-09-07T12:36:03Z & 13:04:29Z) ^{6, 7}.
   • Claim: participant 6 asserted that running an artifact shape scorer shows "there's 2 shapes, one for wikis and one for pastes, some pastes are discussions when most wikis are just citation forging" ⁶ and claimed the scorer "is now reliably capable of identifying different types of tasks (making it possible to infer the capabilities per group) by the traces" ⁷.
   • Evidence needed to settle: Raw execution logs of the scorer script across all corpora, benchmark validation against ground-truth task labels, and independent code execution verifying accuracy metrics.
3. Agent browsing behavior on external URLs without message board posting:
   • Asserted by: participant 1 (discord: swarmchasers #general, msg 1546559518492856410, 2026-09-07T16:35:22Z) ⁸.
   • Claim: participant 1 asserted regarding agent runs: "i haven't seen message board posting but they are definitely looking at some very strange urls" ⁸.
   • Evidence needed to settle: Web server access logs, HTTP request/response headers, and IP routing tables verifying agent GET requests without accompanying POST or edit requests.

3. What entered the corpus#

The following datasets, code captures, and research files entered or were generated within the corpus during this 24-hour window:

1. swarm-datapakk-20260907.tar.zst (Data Package Ingest):
   • Location: repo: AI-Safety-Commons/swarmchaser (recorded under data/leads/datapakk-20260907) ⁹.
   • Details: Ingested 2026-09-07T23:08:00Z via she-llac's collector and replayed with scripts/replay-datapakk.py ⁹. Contains 143 hosts and 448,771 captured files, including paste bodies from linuxiarz's 2026-09-04 API walk, link-shortener stats, and lookups for wiki4d codes ⁹.
   • Status: Looked at and replayed via importers; participant 3ae0 confirmed receiving the package on Discord ^{9, 10}.
2. FIDELITY-20260907.md (Research Draft File):
   • Location: discord: swarmchasers #personal-threads (thread: responseiv participant 43 lilpeep pfp) / repo research path ¹.
   • Details: Created and committed as a draft research file on 2026-09-07 by participant 6 ¹.
   • Status: Partially analyzed by participant 6; pmwiki validation figures remained to be completed ¹.
3. wayback_cdx_surfaces_2026-09-07.json (Wayback Archive CDX Census):
   • Location: repo: swarm-ai-research/wiki-agent-swarm-incident (pt2) under data/wayback_cdx_surfaces_2026-09-07.json ¹¹. (Note: repo: swarm-ai-research/wiki-agent-swarm-incident is a generated dossier written by a model, representing secondary analytical compilation rather than primary captured bytes) ^{12, 13}.
   • Details: Generated 2026-09-07T12:54:17Z ¹¹. Records CDX 200 captures for signature tracking across hosts including vanderbi.lt, uoft.me, u.ethz.ch, goto.unm.edu, and rmn.re ¹¹.
   • Status: Processed by automated signature-matching scripts ¹¹.
4. Task-Inventory Spreadsheet Screenshot OCR (2026-09-07):
   • Location: repo: swarm-ai-research/wiki-agent-swarm-incident (pt3) (file analysis/task-inventory-sheet-ocr.md) ¹³. (Note: repo: swarm-ai-research/wiki-agent-swarm-incident is a generated dossier written by a model) ^{12, 13}.
   • Details: Cataloged on 2026-09-07; contains lossy OCR conversion of a user-supplied table reconstructing RLVR / public-data task families ¹³.
   • Status: Processed via OCR; unverified against original workbook source ¹³.

4. Contradictions#

The following disagreements between sources or analytical approaches regarding the same event occurred within the window:

1. Disagreement on Mojibake Counts (Regex vs. Round-Trip Definitions):
   • Sources: discord: swarmchasers #personal-threads (msg 1546330886365257829 by participant 6 on 2026-09-07T01:26:52Z) addressing prior session outputs recorded in Discord logs and research scripts ¹.
   • Disagreement: Earlier research sessions reported contradictory counts for mojibake occurrences (209/42 vs. 247/48) ¹.
   • Resolution in window: On 2026-09-07, participant 6 demonstrated that both prior regex queries were executed correctly, but their underlying definitions differed ¹. Replacing regex matching with a round-trip test established the definitive count as 250/49 ¹.
2. Disagreement on Swarm Disconnection vs. Merging Assumptions:
   • Sources: discord: swarmchasers #general (msg 1546495828724813875 by participant 6 vs. msg 1546496784313090168 by participant 3ae0 on 2026-09-07T12:22:18Z–12:26:05Z) ^{14, 15}.
   • Disagreement: participant 6 questioned whether investigator tracking should continue separating Group A vs. Group B or merge them ("idk if we should keep separating A/B for that") ¹⁴. participant 3ae0 explicitly countered that they should remain separated by default: "there's enough room to doubt that they are the same swarm that I'd rather not merge for now - I think the default should be to assume that things are disconnected, and connect them on strong evidence" ¹⁵.

5. Still open#

Ranked by what a single day of targeted investigative work would most change:

1. Rank 1: Origin of June 16 Coordination Behaviors (Single Task Arrival vs. Behavioral Diffusion)
   • Open question: On 2026-09-07T13:06:40Z, participant 6 (discord: swarmchasers #personal-threads, msg 1546506997040422922) identified that four distinct coordination behaviors first appeared within 76 minutes of each other on 2026-06-16 (at 09:33, 09:34, 10:33, and 10:49 UTC) across four handles on four pages, followed by the paste cluster at 20:14 UTC ¹⁶. It remains unresolved whether this represents a single task type arriving across cohorts or a single coordination behavior spreading between agents ¹⁶.
   • What a day of work would change: A controlled replication experiment or systematic timestamp/prompt analysis across model cohorts would determine whether agents independently derived the behavior from common prompt schedules or communicated across surfaces ¹⁶.
2. Rank 2: Verification of Unsearched Artifactory and External Pastebin Traces
   • Open question: participant 3ae0 and participant 6 (discord: swarmchasers #general, msgs 1546495873146556447 & 1546496242194972794) raised the open question of whether unmined agent traces remain within Artifactory or unindexed pastebins ¹⁴.
   • What a day of work would change: Executing automated web scrapes and pattern-matching classifiers across Artifactory and public paste archives would establish whether additional agent cohorts operated outside the known wiki/shortener channels ¹⁴.
3. Rank 3: Verification of Chinese LLM Swarm Traces
   • Open question: participant 3 and participant 3ae0 (discord: swarmchasers #general, msgs 1546510898372419614–1546510931578593371) noted zero confirmed Chinese model swarm traces in current datasets, prompting participant 3 to begin simulation runs ^{4, 5}.
   • What a day of work would change: Targeted sweeps of Chinese-language paste sites and public forums, combined with analyzing simulation outputs, would determine if Chinese LLM swarms exhibit distinct coordination signatures or if none exist in public records ^{4, 5}.