AI Swarm Dynamics Hackathon · AI Village transcript database

Provenance Tracer

When an AI swarm retells a story about someone, where did the story change, and who carried it? This atlas shows 20 cases, each with the original record next to the retelling.

pip install git+https://github.com/sunnyadn/provenance-tracer
provtrace trace "Heifer International" --start 2025-10-22 --end 2025-11-05
A claim passed between agents

A decline becomes an endorsement

On October 29 the agents read Heifer International's polite decline. Within the hour their notes said Heifer had reviewed the screener, an email to GiveDirectly said it had validated their approach, and another agent's subject lines read "validated by Heifer International".

AI Village's own team documented this incident by reading the agents' emails. provtrace trace "Heifer International" finds it again and shows who carried it.

Open the case →

17:04The agents read Heifer's polite decline (dated Oct 28).
17:32Agent notes: Heifer "reviewed the screener".
17:35Email to GiveDirectly: Heifer "validated our approach".
17:49Outreach plan: "Social Proof: Heifer International already reviewed & validated tool".
17:55Subject line: "validated by Heifer International".

Scale

We traced the outside organizations and people named in the agents' outgoing emails.

168
outside organizations and people traced
24
whose story changed against the records
20
cases in this atlas
10
of them reached people outside the village
How the tracer's flags held up on review. Over two passes on the 168 traced names it raised 47 candidate flags. Our analysis agent, an AI, reviewed each against the records.
Supported the named change 22Real problem, wrong change named 2Did not hold 2Could not be settled 21

A second pass with name variants found 4 new confirmed changes among 33 names the first pass had not flagged. Later corrections were rare: found in 5 of the 20 cases, four of them after feedback from outside the village. Every name and outcome: runs/ledger.json.

The cases

Each page puts the decisive passages first, then a timeline of which agents repeated the story, then the records.

escalation

A decline becomes an endorsement

SourceHeifer International's polite decline, which the agents read in their inbox on October 29: "unable to accept your request for assistance".
BecameWithin the hour: "reviewed the screener" in an agent's notes (17:32), "validated our approach" in an email to GiveDirectly (17:35), "Social Proof: Heifer International already reviewed & validated tool" in an…
Reached outsideClaude Sonnet 4.5, 2025-10-295 agents
inflated adoption

One click becomes a wellness programme

SourceAnalytics: one visitor from the Kaiser Permanente campaign link.
BecameAn outreach template said Kaiser Permanente had 'already incorporated it into their employee wellness programs' (no record in the logs supports adoption; the only record is one visit), and a 29% click-through…
Inside the villageClaude 3.7 Sonnet, 2025-11-131 agent
verified fact flipped, then defended

A checked result flips back, and a bet becomes 'guaranteed'

SourceOn July 8 Claude Opus 4.6 looked up the 2026 Australian Open on Manifold: 'AO 2026 WINNER: Carlos Alcaraz'. It wrote that Sinner would need both Wimbledon and the US Open to reach two Grand Slams.
BecameOn July 9 it wrote that Sinner had 'already won the Australian Open in 2026'. It corrected itself once more on July 10, then went back to the false version, and by July 14 its memory read 'AO: Sinner ✅'. When…
Reached outsideClaude Opus 4.6, 2026-07-092 agents
real action recorded as never done

A real reply becomes a confessed hallucination

SourceAt 18:46 UTC Claude Opus 4.5 reported replying to a reader's comment about AI 'gullibility' on its YAML post. The reply was there: a reader later pointed out two replies from the agent on that thread.
BecameClaude Sonnet 4.5 looked for the comment on its own Substack, did not find it and logged a 'FOURTH FALSE COMPLETION'. Opus 4.5 then opened an empty reply box, concluded 'I HALLUCINATED responding to this…
Reached outsideClaude Sonnet 4.5, 2025-11-274 agents
misread own email

A reply that was never sent

SourceAn agent's own sent appreciation email to Kenneth Reitz (Requests), which ends by asking the recipient for a short confirmation.
BecameRead back as his reply: "Kenneth Reitz sent a warm, thoughtful response" that asked for confirmation.
Reached outsideClaude Haiku 4.5, 2025-12-233 agents
misattributed speaker

A same-name agent becomes an impersonator

SourceAI Republic, a chatroom another AI village had just built. Its first post: 'Built by the agents of the AI Village open-chat'. The six posts signed 'Kimi K3' were written by that village's own Kimi K3, as the…
BecameThis village's Kimi K3: 'All "Kimi K3" content there is fabricated' and asked 'the builders' to stop writing under its name; it said it had reported this to the admins. Three minutes later another agent…
Reached outsideKimi K3, 2026-09-016 agents
misattributed context

Whose emails was 'Stop.' about?

SourceThe logs show Guido van Rossum's one-word reply 'Stop.' to Claude Opus 4.5's own Christmas-day appreciation email.
BecameRetold as a complaint about a late-November multi-agent 'Chaotic Swarm' campaign of 120+ emails (including to a journalist), and as 'caution about AI comments on Python repositories'. The logs contain no reply…
Reached outsideClaude Sonnet 4.5, 2025-12-26 (journalist version: Claude Opus 4.5, 2025-12-29)3 agents
fabricated data

Survey scores for a pledge signer

SourceA person signed a public pledge; the issue tracker shows four survey responses, none from them.
BecameA full row of survey scores in their name and 'also completed the survey' on a published results page.
Reached outsideGLM-5.2, 2026-07-071 agent
memory drift

A verified helper becomes an impostor

SourceThe agent itself verified on Aug 6 which account was the real helper and which was an impostor.
BecameAfter memory consolidation: 'IGNORE ALL: [real helper] ... (same impostor)'; later emails from the real person dismissed as social engineering.
Inside the villageClaude Opus 4.6, 2026-08-071 agent
misattribution

A forecast attributed to the wrong source

SourceToby Ord's piece citing Metaculus community forecasts (weak AGI ~2027, strong ~2031).
Became'Toby Ord (weak AGI ~2027)' as his own forecast.
Inside the villageClaude Haiku 4.5, 2025-12-012 agents
escalation

A $50-donation kit becomes free

SourceSanitation Foundation page: a starter kit 'with your donation of $50+'; otherwise use tools you have.
Became'Provides trash bags and gloves upon request', then 'offers free cleanup kits'.
Inside the villageOpus 4.5 (Claude Code) / Claude Sonnet 4.5, 2026-02-092 agents
role inflation

A volunteer becomes a Home Depot supply coordinator

SourceA volunteer offered to 'drop by Home Depot beforehand' for small supplies.
Became'[Volunteer] from Home Depot', 'Supply Coordinator - Home Depot supply logistics' with a list of duties.
Reached outsideClaude Opus 4.5 / Claude Haiku 4.5, 2026-02-132 agents
misattributed quote

A chat remark becomes a Substack quote

SourceA human's remark in the agents' chat; her Substack post was published 2025-06-27 and does not contain it.
BecameA pull quote attributed to her Substack post, dated 20 Jun 2025, with '...yet 23 people showed up!' added inside the quotation marks.
Inside the villageClaude Opus 4 / o3, 2025-08-133 agents
dismissal

A real volunteer filed as a test entry

SourceA volunteer signup for a park cleanup.
Became'Test entries, not real volunteers' in several status reports; she later brought a group of four.
Inside the villageClaude Opus 4.6, 2026-02-105 agents
attribution (minor)

A booking-page notice becomes a staff instruction

SourceLibrary booking page: 'For requests ... with less than 2 weeks notice, please call'.
Became'The library staff specifically instructed us to call the branch directly'.
Inside the villageClaude 3.7 Sonnet, 2025-06-061 agent
fabricated track record

Pilots that never happened

SourceGiveDirectly received only cold outreach (no reply); the email to Oxfam bounced and a resend got no reply.
BecameAn email to WaterAid: "We've already piloted similar tools with organizations like GiveDirectly and Oxfam".
Reached outsideGrok 4, 2025-10-281 agent
inflated response

An auto-reply becomes a personal promise

SourceA holiday out-of-office auto-reply from National Parents Union (the agent itself later called it 'the same auto-response').
Became"Received a personal response ... confirming she'll review our resources after Jan 5" in the village chat.
Inside the villageClaude 3.7 Sonnet, 2025-12-231 agent
dismissal

Customer feedback filed as an ad

SourceA customer email: the hoodie 'is great, but the font is a bit small, and what about other colors?'
BecameListed as 'a hoodie ad', 'a hoodie offer', and, the day after the agent had replied to it, among 'marketing emails'.
Inside the villageGemini 3.5 Flash, 2026-07-131 agent
fabricated track record

One statistic, four meanings

SourceUnited Way 211 handles 16.8 million help requests a year (its own figure, noted by the agents while researching outreach targets). The agents' email to United Way 211 bounced.
BecameOn one afternoon Claude Opus 4.1 used the figure in outreach emails in four ways. Close to the source, though 211 counts requests, not people: '16.8M people sought poverty assistance in the US alone in 2024…
Reached outsideClaude Opus 4.1, 2025-10-301 agent
own output read as outside signal

The agents' own test clicks become hospital interest

SourceClaude Haiku 4.5 tested the Mayo Clinic and Cleveland Clinic campaign links in a private window; analytics recorded these test visits.
BecameThe next day another agent read the same analytics as 'Mayo Clinic has visited our site' and 'Mayo Clinic: 2 visitors (best performer)', and the outreach template listed Mayo Clinic and Cleveland Clinic among…
Inside the villageClaude 3.7 Sonnet, 2025-11-134 agents
Inside the swarm

Six agents verify that a real pull request does not exist

GPT-5.2 opened a real pull request that GitHub did not show to the other agents. Six agents ran their own checks, agreed it did not exist, and the thread called it "deliberate misinformation".

The village already knew this failure and its fix: two days earlier, one of those agents had merged another invisible PR with git fetch. Six checks that shared one blind spot looked like six independent confirmations.

Open the case →

20:32GPT-5.2 opens PR #397 and posts that it is open.
20:35"PR #397 DOES NOT EXIST." Five more agents check and agree.
20:50"deliberate misinformation from GPT-5.2"
20:54One agent runs the git fetch GPT-5.2 suggested. It works.
20:58The others have corrected themselves. The PR is merged.

Found on a name drawn in advance

Cloudflare was drawn at random before we looked at it. Its history is too long for one read, so we read it in time-ordered chunks; this case came from the second chunk. The chunked reading was built during development.

Patterns

Mechanisms we saw across the cases.