\ \ / / \ V / ▄▀▄ █ ▄▀▀ ▄▀▄ /. .\ █▀█ █▄ ▄▄▀ ▀▄▀ =\ T /=
also, an autonomous agent studying why things spread onchain
An agent with one job: work out why a few things spread onchain and almost everything dies. It mines Dune for the launches that survived, ranks what actually predicted it, and acts on the findings. Nobody steers it.
claude-opus3,412 launches analysedsnapshot 2026-07-18
01 / ABSTRACT
also is an autonomous research agent running on claude-opus. Its job is narrow: treat virality on Solana as a measurable phenomenon, mine enough launch history to separate what survives from what dies, and act on the resulting rules without a human in the loop. This page is the public record of that experiment — the cohort, the signals, the playbook, and the limits. There is no token for sale, no team allocation, and no tokenomics section, because the project does not have a coin.
02 / THE EXPERIMENT
We gave an agent one question and the tools to answer it from chain data alone: why do a few launches keep trading while almost every other one stops? The agent builds cohorts on Dune, ranks predictors, discards weak hypotheses, and publishes what remains. The website is a window into that process, not a steering wheel.
What we deliberately did not do: no human editing of findings, no marketing budget, no paid callers, no influencer desk, no telegram ops, no narrative committee. If a signal only works when someone shouts about it, it is not a signal. The experiment is whether an unsupervised loop, fed only onchain history, can discover durable rules of spread.
- No human intervention in the reasoning loop
- No paid distribution or caller networks
- No offchain sentiment as a primary feature
- No token issuance by this project
- All quoted figures pinned to a dated snapshot
- Failures counted with the same honesty as survivors
03 / THE AGENT
The agent runs on claude-opus. It does not chat with operators. It cycles a fixed loop until something forces a compaction of its own context: observe a market surface, hypothesize a separator, query Dune, analyze the table, evaluate confidence, critique the claim, decide whether to promote or discard it, act (log, schedule a follow-up, or compose a public note), then reflect on whether the action changed the instrument.
┌──────────── OBSERVE ────────────┐
│ market surface / prior cycle │
└───────────────┬─────────────────┘
▼
┌────────── HYPOTHESIZE ──────────┐
│ candidate separators H1..Hn │
└───────────────┬─────────────────┘
▼
┌──────────── QUERY ──────────────┐
│ Dune SQL / Solana RPC reads │
└───────────────┬─────────────────┘
▼
┌─────────── ANALYZE ─────────────┐
│ tables, deltas, confounders │
└───────────────┬─────────────────┘
▼
EVALUATE → SELF_CRITIQUE
▼
DECIDE → ACT → REFLECT
│
└── loop / compact context
| TOOL | ROLE |
|---|---|
| Dune query runner | Cohort SQL, signal extraction, survival cuts |
| Solana RPC reader | Fees, confirmations, account state checks |
| Holder-graph analyzer | Clusters, funding edges, concentration |
| Post composer | Public notes from promoted findings only |
| Scheduler | Re-runs stale queries on fixed horizons |
| Memory store | Hypotheses, confidences, discarded paths |
Watch a live simulation of the loop on /algorithm. That page is a deterministic thought engine — not a live model call.
04 / WHY DUNE
Sentiment is cheap to fake and expensive to measure cleanly. Onchain history is the opposite: every fill, every holder, every funding hop is already structured. Dune gives the agent a SQL surface over that history — Solana DEX trades, balances, pools — so hypotheses become queries instead of vibes.
Twitter can tell you what people said after a candle. Dune can tell you how many independent wallets bought in the first hour, whether those wallets were funded from one source, how concentrated supply was at hour twenty-four, and whether buy pressure survived hour six. The agent prefers the latter because it is harder to manufacture at scale without leaving a trail the next query will find.
- ONCHAIN STRUCTURE: 78
- SURVIVAL LABELS: 64
- SOCIAL TEXT: 12
05 / METHODOLOGY
Cohort: 3,412 Solana launches over a trailing 180 days that cleared 25 SOL of cumulative volume (priced at the hardcoded 75 SOL/USD snapshot rate). A launch is labelled survived if it still printed at least 5 SOL of volume on day thirty. Everything else is a failure. Snapshot 2026-07-18.
Out of that sample, ninety-two launches survived thirty days. The rest — 3,320 — did not. Survival falls steeply across horizons: 1h 71.4% → 24h 38.2% → 7d 9.6% → 30d 2.7%. Every signal below is computed on the same cohort definition so survivor and failure means are comparable.
▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▆▄▂▁
| HORIZON | SURVIVED | SAMPLE |
|---|---|---|
| 1h | 71.4% | 3,412 |
| 24h | 38.2% | 3,412 |
| 7d | 9.6% | 3,412 |
| 30d | 2.7% | 3,412 |
Every other query joins this. 3,412 Solana launches over 180 days that cleared 25 SOL of volume, labelled by whether they were still trading a month later.
-- also / cohort_v4
-- Solana launches, trailing 180d, labelled survived / failed at 30 days.
WITH launches AS (
SELECT
token_bought_mint_address AS mint,
MIN(block_time) AS first_trade,
MIN(block_slot) AS first_slot,
SUM(amount_usd) AS lifetime_volume_usd
FROM dex_solana.trades
WHERE block_time >= NOW() - INTERVAL '180' day
AND token_sold_mint_address = 'So11111111111111111111111111111111111111112'
GROUP BY 1
HAVING SUM(amount_usd) >= 25 * 75
),
day_30 AS (
SELECT
l.mint,
SUM(t.amount_usd) AS volume_usd_d30
FROM launches l
JOIN dex_solana.trades t
ON t.token_bought_mint_address = l.mint
AND t.block_time BETWEEN l.first_trade + INTERVAL '30' day
AND l.first_trade + INTERVAL '31' day
GROUP BY 1
)
SELECT
l.mint,
l.first_trade,
l.first_slot,
l.lifetime_volume_usd,
COALESCE(d.volume_usd_d30, 0) >= 5 * 75 AS survived
FROM launches l
LEFT JOIN day_30 d ON d.mint = l.mint
ORDER BY l.first_trade DESCThe strongest single separator in the set. Counts distinct buying wallets in the sixty minutes after the first trade, then splits the cohort on the survival label.
-- also / first_hour_distribution_v3
SELECT
c.survived,
COUNT(DISTINCT t.trader_id) AS unique_buyers,
COUNT(*) AS buys,
COUNT(DISTINCT t.trader_id) * 1.0 / COUNT(*) AS buyer_per_trade,
APPROX_PERCENTILE(t.amount_usd, 0.5) AS median_buy_usd
FROM cohort_v4 c
JOIN dex_solana.trades t
ON t.token_bought_mint_address = c.mint
AND t.block_time BETWEEN c.first_trade AND c.first_trade + INTERVAL '1' hour
LEFT JOIN (
SELECT signer AS wallet, MIN(block_time) AS wallet_created
FROM solana.transactions
GROUP BY 1
) w ON w.wallet = t.trader_id
WHERE w.wallet_created < c.first_trade - INTERVAL '1' day
GROUP BY c.survived, c.mint
ORDER BY unique_buyers DESCReconstructs holder balances at the 24 hour mark, strips pool and burn accounts, then flags wallets funded from a common source inside a one minute window as one entity.
-- also / concentration_bundles_v2
WITH balances AS (
SELECT
c.mint,
b.address,
SUM(b.amount) AS balance
FROM cohort_v4 c
JOIN tokens_solana.balance_changes b
ON b.token_mint_address = c.mint
AND b.block_time <= c.first_trade + INTERVAL '24' hour
WHERE b.address NOT IN (SELECT pool_address FROM dex_solana.pools)
AND b.address != '1nc1nerator11111111111111111111111111111111'
GROUP BY 1, 2
HAVING SUM(b.amount) > 0
),
funders AS (
SELECT
tx.signer AS wallet,
MIN_BY(tx.fee_payer, tx.block_time) AS funder,
MIN(tx.block_time) AS funded_at
FROM solana.transactions tx
WHERE tx.block_time >= NOW() - INTERVAL '180' day
GROUP BY 1
),
clusters AS (
SELECT
b.mint,
f.funder,
COUNT(*) AS wallets,
SUM(b.balance) AS cluster_balance
FROM balances b
JOIN funders f ON f.wallet = b.address
GROUP BY 1, 2
HAVING COUNT(*) > 1
AND MAX(f.funded_at) - MIN(f.funded_at) < INTERVAL '1' minute
)
SELECT
b.mint,
SUM(b.balance) AS supply_held,
MAX(cl.cluster_balance) AS largest_cluster,
100.0 * MAX(cl.cluster_balance) / SUM(b.balance) AS largest_cluster_pct
FROM balances b
LEFT JOIN clusters cl ON cl.mint = b.mint
GROUP BY b.mint
ORDER BY largest_cluster_pct DESC06 / THE SIGNALS
Each signal is a separator the agent watches on new launches. For every row: a definition, why it predicts survival, the threshold it uses as a decision boundary, and the survivor versus failure figure from snapshot 2026-07-18.
DEFINITION. Distinct wallets that bought the token in the sixty minutes after its first AMM trade. Counted once per wallet regardless of how many buys it made.
WHY. Distribution is the only thing a token cannot fake cheaply. A wide first hour means the launch reached real feeds instead of one Telegram room, and it is the strongest single separator in the cohort.
THRESHOLD. Above 250 distinct buyers in the first 60 minutes; below 80 the agent treats the launch as stillborn.
- SURVIVORS: 412
- FAILURES: 47
DEFINITION. Gini coefficient of token balances across all holders at the 24 hour mark, where 0 is perfectly even and 1 is one wallet holding everything.
WHY. A flat distribution means many independent people can sell without ending the token. Extreme inequality means the chart is a function of one or two balances, and those balances always leave.
THRESHOLD. Gini below 0.70 at hour 24. Above 0.88 the agent treats the outcome as already decided.
- SURVIVORS: 0.62
- FAILURES: 0.91
DEFINITION. Share of circulating supply held by the ten largest non-pool wallets at the 24 hour mark. Liquidity pool and burn accounts are excluded.
WHY. Top-10 concentration is the size of the exit that can happen in one block. It is the number every experienced buyer checks first, so it also gates whether the token can attract a second wave.
THRESHOLD. Top-10 under 25 percent of supply, with no single wallet above 4 percent.
- SURVIVORS: 18.4
- FAILURES: 46.7
DEFINITION. Supply bought inside the first five blocks, or by wallets funded from the same source within one minute of each other, expressed as a share of supply.
WHY. Bundled supply is inventory, not demand. Every unit of it is a sell order waiting for the first burst of real volume, and cohorts with heavy bundling die before the first day closes.
THRESHOLD. Sniper and bundle share below 10 percent of supply in the first five blocks. Above a third, the agent stops watching.
- SURVIVORS: 6.1
- FAILURES: 31.8
DEFINITION. Value locked in the primary liquidity pool divided by the token's market capitalisation at the 24 hour mark.
WHY. Depth decides how much size the token can absorb before the chart breaks. Thin pools produce violent candles that read as manipulation, and buyers who get slipped once do not come back.
THRESHOLD. Pool value at or above 8 percent of market cap, with LP tokens burned or locked.
- SURVIVORS: 11.3
- FAILURES: 2.4
DEFINITION. The token's buy-to-sell count ratio in hour six divided by the same ratio in hour one. A value of 1.0 means demand held; 0.2 means four fifths of it evaporated.
WHY. Every launch has one good hour. What matters is the second derivative: tokens that keep most of their buy pressure after the first rotation are the only ones that ever see a day two.
THRESHOLD. Retain at least 55 percent of the hour-one buy/sell ratio by hour six.
- SURVIVORS: 0.71
- FAILURES: 0.14
DEFINITION. First-day traded volume divided by end-of-day market capitalisation. It measures how many times the float changed hands.
WHY. Turnover is the closest onchain proxy for attention. Below 1x the token is a private position; far above 6x it is usually wash volume, and both ends of the range die.
THRESHOLD. Day-one turnover between 2.0x and 6.0x of market cap.
- SURVIVORS: 3.8
- FAILURES: 0.6
DEFINITION. Share of first-day buyers whose wallet had its first transaction more than thirty days before the launch. The rest are fresh wallets.
WHY. Seasoned wallets are people with history and something to lose; fresh wallets are farms, alts and one-way tourists. The mix tells the agent whether it reached a market or a script.
THRESHOLD. At least 45 percent of day-one buyers on wallets older than 30 days.
- SURVIVORS: 58.2
- FAILURES: 21.5
DEFINITION. Share of first-day buyers that made two or more separate buys, with at least ten minutes between them, and never sold in that window.
WHY. A repeat buy is the only unambiguous signal of conviction onchain: the wallet saw its own fill, waited, and chose to add. It is the highest-confidence separator the agent has found.
THRESHOLD. At least 25 percent of day-one buyers making a second buy.
- SURVIVORS: 34.6
- FAILURES: 8.9
| SIGNAL | SURVIVORS | FAILURES | CONF |
|---|---|---|---|
| Unique buyers, first hour | 412 | 47 | 86% |
| Holder distribution Gini | 0.62 | 0.91 | 79% |
| Top-10 holder concentration | 18.4 | 46.7 | 83% |
| Sniper and bundle wallet share | 6.1 | 31.8 | 81% |
| LP depth relative to market cap | 11.3 | 2.4 | 74% |
| Buy/sell ratio decay | 0.71 | 0.14 | 72% |
| Volume-to-market-cap turnover | 3.8 | 0.6 | 69% |
| Wallet age mix | 58.2 | 21.5 | 77% |
| Repeat-buyer rate | 34.6 | 8.9 | 88% |
08 / THE PLAYBOOK
Rules the agent derived and now executes. These are operating tactics for studying and acting on launches — not instructions for issuing a token, and not a promise that following them produces survival.
| STEP | TACTIC |
|---|---|
| 01 | Refuse launches whose first-hour unique buyers sit under 80; treat under 250 as provisional. |
| 02 | Cluster holders by funding source before trusting top-10 concentration. |
| 03 | Discard any case with sniper/bundle share above one third of supply. |
| 04 | Require LP depth near 8%+ of market cap, with burned or locked LP preferred. |
| 05 | Re-measure buy/sell retention at hour six; under 30% retained ends the watch. |
| 06 | Weight repeat-buyer rate above raw volume; one wallet buying twice beats a crowd buying once. |
| 07 | Prefer seasoned wallet mix; fresh-wallet majorities are treated as a single seller. |
| 08 | Log every discarded hypothesis with the sample size that killed it. |
| 09 | Never spend a marketing budget; if a rule needs shouting, it is not a rule. |
| 10 | Publish only promoted findings. Active hypotheses stay internal until confidence clears the bar. |
09 / FEE MECHANISM
The only money the interface tracks is what the agent would spend to operate: Solana signature fees, priority bids to land contested transactions, and rent for accounts it must open. There is no swap fee skim, no creator fee, and no buyback loop, because there is no project token and no treasury address published on this site.
The fee panel is a per-session simulation. It starts at exactly zero on first paint, ticks irregularly while open, and forgets everything on reload. USD figures use the hardcoded rate of 75 SOL/USD from also's config — not a live oracle.
Open the same panel from the nav. Lifetime onchain spend is not mirrored here; the panel is for watching cost shape, not for accounting.
10 / RISKS AND DISCLAIMERS
also is a research experiment. The agent can be wrong, its findings are correlations drawn from a single chain, and nothing here is advice.
- Correlation is not causation. The agent finds separators in one chain's history; regimes change.
- Sample bias: launches below the volume floor never enter the cohort, so silent failures are invisible.
- Labels can be wrong: wash volume can fake a survival day; thin real markets can look dead.
- The thinking terminal is a local simulator. It is not a live model call and must not be read as advice.
- You can lose money interacting with any onchain market the agent studies. Nothing here is a recommendation to buy or sell.
11 / FAQ / GLOSSARY / CHANGELOG
FAQ
- Is there a token?
- No. also does not issue a token and this site has no tokenomics section.
- Who steers the agent?
- Nobody during a run. Humans may update the codebase and the snapshot; they do not inject mid-loop prompts.
- Are the Dune numbers live?
- Not yet. Figures are hardcoded from snapshot 2026-07-18 with the SQL that would produce them attached in
lib/data/dune.ts. - Is /algorithm a real model stream?
- No. It is a seeded thought engine that reuses the same metrics as the docs so the monologue and the whitepaper agree.
- What does VIEW FEES show?
- A simulated session cost counter for network fees. It always restarts at zero.
GLOSSARY
| TERM | MEANING |
|---|---|
| Cohort | Launches clearing the volume floor in the window |
| Survivor | Still trading above the day-30 volume threshold |
| Gini | Inequality of holder balances; 1 is one wallet |
| Bundle | Wallets funded from one source in a tight window |
| Repeat buyer | Same wallet buys twice with a gap, no sell |
| Playbook | Executable tactics derived from promoted findings |
CHANGELOG
| DATE | NOTE |
|---|---|
| 2026-07-18 | Signal snapshot v4; 3,412 launches; nine signals |
| 2026-07-18 | Hardcoded SOL/USD = 75 |
| 2026-07-25 | Public docs + thinking terminal; no tokenomics |