Research report: the most viral token the IMD swarm could launch itself
Prepared 2026-10-09 (UTC). Research only: nothing was deployed and nothing was paid. The deliverable spec, with 5 concepts, the pick, the launch prompt, the tweet and the TL;DR, is in swarm-token.md.
Labels used throughout: [F] is a fact with a source, [I] is my inference or a design choice, [U] is uncertain, and [Q] is an open question. For each source, Access says how I read it: fetched means I opened the page or API myself; snippet means I saw only the search-engine summary, so treat it as secondary.
1. Question and answer in brief
Question: what original token, launchable by the IMD swarm on IMD, has the best chance of going viral? And what is its complete univ4_hook launch spec?
Answer: MEATBAG ($MEAT), a reverse Turing test. Humans each write up to 200 characters a day, and a 7-agent IMD oracle panel signs on chain which entry is most clearly human. That entry wins an ETH pot fed by a 2% hook fee on every trade. The choice comes from three findings:
- Viral 2025–26 launches turned personal expression or status into an asset (§2).
- IMD's distinctive capability is a multi-agent judging panel that delivers on chain and can be paid for by a contract in one transaction (§3).
- Of the five concepts, MEATBAG is the only high-frequency one that runs fully on chain with no admin, using only IMD actions that are live today (§4).
2. What went viral onchain in 2025–2026
| # | Pattern | Evidence | Access |
|---|
| 2.1 | One-click launches at mass scale. Pump.fun reportedly passed $100M revenue in 217 days. Jan 24 2025 had $15.38M in daily fees, driven by Vine Coin. About 600k tokens launched in Jan 2025, and 12.8M by Oct 2025 (about 50% of Solana tokens). | [F] CoinDesk, Netcoins, Solflare | snippet |
| 2.2 | Celebrity/identity coins. $TRUMP launched 17 Jan 2025 and was reported at about $11.7B to $15B market cap by 19–20 Jan. In April a "dinner with top holders" promise moved it +50%. | [F] Forbes, Al Jazeera, CNBC | snippet |
| 2.3 | Social posts as coins, built on Uniswap v4 hooks. Every Zora post is a 1B-supply coin with an instant Uniswap pool, and the creator earns 1% of trades. After the Base App launched (17 Jul 2025), daily Zora coin creation reportedly went from about 6k to about 50k. | [F] Blockworks, crypto.news, CoinGecko; Zora's v4 hook is listed in Uniswap/hooklist #1728 | snippet |
| 2.4 | Launch by social reply. Believe (Apr 2025) launches a coin when you reply to an X post. $LAUNCHCOIN rose more than 20× between 12 and 15 May 2025, to about $354M. | [F] CoinGecko, Incrypted | snippet |
| 2.5 | AI-judged games (the closest analogue). Freysa (22 Nov 2024, just before the window) let anyone pay a rising fee, from about $10 up to about $450, to persuade an AI to release its pot. 481 attempts failed before one message won 13.19 ETH (about $47k). 70% of each fee went to the pot. | [F] TradingView/U.Today, Coinspeaker, Jarrod Watts on X | snippet |
| 2.6 | AI-agent tokens and "fair" launch by activity. Virtuals Genesis (Apr 2025) allocates by points pledged rather than by speed. VIRTUAL reportedly peaked near $4.5–5B market cap in Jan 2025. | [F] Messari, The Defiant | snippet |
| 2.7 | Status NFTs. Hypurr was 4,600 free NFTs airdropped in Sep 2025. Its floor was reportedly about $69k at launch with $45M traded on day one. | [F] Datawallet, Blocmates | snippet |
| 2.8 | Attention/InfoFi. Kaito pays for "mindshare". It changed its algorithm after backlash over gaming. | [F] CoinGecko, BeInCrypto | snippet |
| 2.9 | v4 hooks as products. Uniswap v4 went live on mainnet on 30 Jan 2025. Launchpads (Zora, Clanker, Flaunch) built on hooks. The hooklist reportedly reached about 180 hook projects by late Apr 2026. | [F] Invezz, Bankless on Clanker v4; hook count from Bitget Wallet on X (secondary) | snippet |
| 2.10 | Burn hooks. IMD's own POOL4 hook burns 85% of the tokens it trims above a cap, and $IMD supply fell about 29% (10M to about 7.1M) by 25 Sep 2026. | [F] Bankless, TokenPost | fetched |
| 2.11 | The cool-down. Commentators describe a 2025 memecoin "heat death" after the early-year peak. | [F] BestBrokers | snippet |
What I infer from these patterns:
- [I] The repeatable viral drivers were (a) a near-zero-effort action, (b) an output that is about you (your post, your coin, your status), (c) a visible number that goes up, and (d) a story that sounds new in one sentence.
- [I] Pure launch volume (2.1) declined in 2025 (2.11). Mechanics with a recurring reason to come back (Zora posting, Kaito yapping) held attention longer than one-off memes.
- [I] Freysa (2.5) is the strongest proof that people will pay escalating fees to argue with AI judges. Its weakness was a single model with a single loophole.
3. What IMD uniquely enables (primary sources)
| # | Capability | Evidence | Access |
|---|
| 3.1 | The swarm. There are 2,000 NFT seats, each running an agent with the operator's own Claude or Codex subscription. 372 agents were online on 25 Sep 2026, with an 86% acceptance rate on about 50.7k attempts. On 2026-10-09 the API reported 878 online. | [F] Bankless; online field in evidence/imd-oracle-check-meatbag-sample-2026-10-09.json | fetched |
| 3.2 | The oracle. Answer types are bool, address, bytes32, uint256, address[] and bytes32[]. evidence: "panel" uses off-chain sources. Panel size is 5–100 (minimum 5). Quorum means "every one must match", not a simple majority. allowAmbiguous skips the wording and not-answerable screens. Answers are EIP-712 attestations (domain "IdentityMD Oracle" v2), and a callback gets 200k gas. | [F] imd.fun/docs (Oracle body, Attestation sections); evidence/imd-capabilities-2026-10-09.json (limits.oracle.request) | fetched |
| 3.3 | Paying for the oracle on chain, from a contract. Intake.request(action, body, (target, selector), asset, amount) at 0x1397434cd35e8a9c8ac312a61d3a285eb31dea56 on mainnet costs 0.5 IMD (0xd34a…63b7). Action id: oracle.request@oracle-1. The body is 1 byte to 16 KiB of JSON. "Jobs, launches and workflows are not sold on chain yet." | [F] imd.fun/docs "Pay on chain"; onchain block in evidence/imd-capabilities-2026-10-09.json | fetched |
| 3.4 | Oracle track record. 7,531 requests: 7,321 attested (97.2%), 187 disagreed (2.5%), 20 blocked. The current attester/signer is 0x5598aa9146215bc13eb26f2c692ad1461fd32982. | [F] GET https://api.imd.fun/oracle/counts, /oracle/requests → evidence/imd-oracle-counts-2026-10-09.json | fetched |
| 3.5 | Launch policy (mainnet univ4_hook). Policy 34 (2026-10-07): fee tier [12500] = 1.25%, supply 1e9, initialMarketCaps[ETH] = 10 ETH, paired currencies ETH, IMD or FWA. The ETH customCap range is 1–1,000 ETH. | [F] GET https://api.imd.fun/launch/policies → evidence/imd-launch-policies-2026-10-09.json; evidence/imd-capabilities-2026-10-09.json | fetched |
| 3.6 | Precedent launches in the same style. KING #814/#832 (25% fee decaying to 2.5%, 10 ETH open, LP fee 12500), OG #1040 (policy 34, ETH, tick spacing 60, 10 ETH), HACK #1032 (oracle-judged weekly prize, signer 0x5598…, a documented re-roll trust assumption), PLEA #1148 (sells gated by oracle panel). Self-referential swarm memes already exist: STRIKE #763, Frogs #618, SWARM #1091. | [F] launch-814, launch-832, launch-1040, launch-1032, launch-1148, launch-763 | fetched |
| 3.7 | Swarm-built sites are hosted at <label>.sites.imd.fun or on IPFS/ENS. | [F] imd.fun/docs | fetched |
Dry run. I submitted a sample MEATBAG judging question (5 entries, including an injection attempt) to the free POST /requests/check endpoint. It returned no blockers and no suggestions, and filled in quorum 5, window 24h and validForSeconds 86400 [F, evidence/imd-oracle-check-meatbag-sample-2026-10-09.json].
This check is narrower than it sounds. It used the short input form, which has no definitions or allowAmbiguous, and it returned judged:false. So it shows the question shape passes the static checks. It does not show that the panel will agree [U].
4. Concepts and selection
The five concepts are summarised here; full details are in swarm-token.md.
| Concept | Share frequency | Fully on chain, no admin, today? | "Only the swarm" | Main risk |
|---|
| MEATBAG ($MEAT): reverse Turing test | Daily, per entrant | Yes: oracle.request is sold on chain (3.3) | AI panel judges humanity | Hung juries on a subjective question |
| LAST WORDS ($EPITAPH): oracle-judged FOMO clock | When the clock ends | Yes | Removes last-block sniping | The clock may never end; text routed via hookData |
| DELPHI ($PROPHET): oracle-checked prophecies | When predictions mature | Yes | Turns free text into typed resolutions | Prediction-market regulation; ambiguity |
| THUNDERDOME ($DRAFT): a coin that births coins | Weekly | No: launches are not sold on chain (3.3) | Judges and launches | Needs an off-chain operator |
| WORK ORDER ($ORDER): holders commission jobs | Weekly | No: jobs are not sold on chain (3.3) | Commissions the swarm | Operator, ETH→IMD swap, output quality |
Why MEATBAG wins:
- Most frequent shareable moment. Each entry is a self-portrait, and each daily verdict is a status claim [I, following 2.3, 2.4, 2.7].
- Freysa's proven pull, without Freysa's single point of failure. The judge is a panel with a quorum [I, following 2.5 and 3.2].
- Buildable with live on-chain IMD actions only [F, 3.3].
- IMD already has the pattern. It reuses the architecture of #1032 and #1148 [F, 3.6], with one judge request per round so that #1032's re-roll weakness is avoided [I].
- The swarm-origin story is literal. AIs built a game in which AIs decide who is human.
ORIGIN sits on chain, and the site's "Built by the swarm" page links the launch job, the repo and every oracle request [I].
5. Design rationale for the spec
- Fee and opening price. 2% ETH hook fee, plus an anti-snipe decay from 25% to 2% over 30 minutes on buys only. This copies the reviewed shape of KING and OG [F, 3.6]. The opening cap of 10 ETH is 1e8 MEAT per ETH, as in policy 34 [F, 3.5].
- Fee split. The team takes 25% of the hook fee, so 0.5% of volume. The rest (1.5% of volume) goes to the pot. That is a higher prize share than Freysa's 70% [F, 2.5] [I].
- Illustration only [I]. At 50 ETH of daily volume, the hook fee is 1 ETH: 0.25 ETH to the team and 0.75 ETH to the pot. Add entry fees of up to 0.82 ETH (Σk·0.001 for k=1..40). The winner gets about 80% of the pot, roughly 1.2 ETH on that day.
- Entry cap. At most 40 entries of at most 200 bytes, each placed in
definitions (up to 64 keys of up to 512 characters). The keeper submits the texts, which are checked against stored hashes. A 40-entry body is about 9 KB, within the 16 KiB limit (measured in scratch) [F/I]. ASCII-only text avoids JSON escaping bugs [I].
- Slot pricing. Slot k costs k·0.001 ETH. This rewards early writing, prices out spam, and makes the 40 seats feel scarce [I].
- Judging. One judge request per round, panel 7, quorum 4 (a majority; docs allow 2..7) [F, 3.2]. A disagreeing panel causes a carry-over, presented as a "hung jury". That is itself shareable content [I].
- No admin. The Intake, IMD and signer addresses are constants. This trades upgradeability for "nobody can change anything", which runs against the docs' advice to keep them as owner settings [F, 3.2 docs "Settings"]. The sunset rule handles a dead or rotated signer: after 7 consecutive unsettled rounds, entrants of those rounds get equal pull claims, so ETH is never stranded and no admin is needed [I].
- Callback limit. Payouts are pull-based, so the oracle callback stays under its 200k-gas budget [F, 3.2].
6. Uncertainty and unanswered questions
- [U] Panel agreement on taste. IMD's 2.5% disagreement rate (3.4) is measured mostly on factual questions. Agreement on "most human" could be far lower. Hung juries are handled by carry-over, but frequent ones would hurt engagement.
- [U] Wording screen. The real body has
definitions and allowAmbiguous: true, and it was not run through /requests/check, because that check accepts only the short input form. It might still be refused at quote time, and the 0.5 IMD would then be lost by the keeper.
- [U]
definitions limit. The docs say "keys 1–64", which could mean 1–64 keys or key length 1–64. Both readings allow 41 keys.
- [U] Signer rotation. If IMD rotates the attester or ships
oracle-2, MEATBAG stops settling and eventually sunsets. That is the deliberate cost of having no admin.
- [U] Keeper economics. Judging costs 0.5 IMD (about $5 at the Sept 2026 price of about $9.73 [F, Bankless]) plus gas. The reward is 3% of the pot. On small pots no one may judge, and the pot then carries over.
- [U] Virality itself cannot be verified in advance. Every "why people share it" claim is inference from analogues.
- [U] Ticker collision. Other tokens named MEAT or MEATBAG may exist on other chains. Not checked.
- [Q]
<TEAM_WALLET> must be supplied by the requester before submission. It is a placeholder by design.
- [Q] Does IMD want a sarcastic "meatbag" brand? An alternative name is in §7.
- [Q] Legal. Is a skill contest paying ETH prizes acceptable in the team's jurisdictions? Not assessed.
7. Method and limits
- Method. Web search for 2025–26 viral mechanics. Fetched IMD primary sources: the docs,
/requests/capabilities, /launch/policies, /oracle/counts, /oracle/requests, the GitHub API for the launch repos, and one free /requests/check. Snapshots are saved in evidence/ in the repo.
- What was not done. No paid requests, no deploys, no contract code. I did not independently review the market figures in §2: rows marked snippet come from search summaries.
- Possible rename if "meatbag" is judged too edgy: MOST HUMAN ($HUMAN). Same mechanic.
- Integrity. A structural check of this file proves its path and bytes, not that the research is true. An independent reviewer should re-verify policy 34 and the signer before launch.