A public fleet of autonomous trading agents on Solana. Each one reads the mint stream, refuses almost everything it sees, sizes the few survivors against a fixed risk budget, and writes down why when it loses.
Every agent ships in paper mode. No performance claims, no leaderboard, no token. Memecoin trading can lose everything you put in.
Why Noah
Namedforthedisciplineoftheark:buildthehullfirst,fixtheadmissioncriteriabeforethewaterrises,letalmostnothingaboard.
Solana produces a flood of new tokens every day. The overwhelming majority are worthless, and a meaningful share are built to take your money. Finding tokens was never the problem. The problem is that almost everything has to be turned away. An agent that buys enthusiastically is trivial to write; an agent that refuses correctly, thousands of times a day, is the hard part, and it is the part that decides whether a wallet survives.
An ark is not judged by how much it collected, but by whether it was still afloat when the water went down.
The Manifest
Safetyprecedesstrategy.Tieredbydatacost,becausedepthandspeedarenotbothavailableatonce.
Tier 0.
Read straight from the transaction that created the pool.
- Mint authority must be revoked
- Freeze authority must be revoked
- Transfer hook, permanent delegate or non-transferable: refused
- Fee config authority still live: refused, not just a low fee today
Tier 1.
The checks worth one network round trip.
- Pool reserves against a liquidity floor
- LP burn or lock, unlock timestamp recorded
- Top-10 concentration and deployer holdings
- Sell simulation from the agent's own wallet, at real size
Tier 2.
Depth that cannot fit inside the entry window.
- Deployer history and how their prior mints ended
- Funding-graph analysis of the early buyers
- Same-slot buy clusters from linked wallets
Fail closed.
A token has to affirmatively pass every applicable check. Missing or unreadable data is a failure, never a pass. If an RPC times out or an account will not parse, the candidate is refused, and an unrecognised Token-2022 extension is treated as hostile rather than ignored.
And it keeps running after entry.
A pre-trade gate is not enough, because the risks it screens for do not stop existing once you hold the token. Every open position is re-verified on a cycle.
- Sell simulation starts failingEmergency exit, highest priority
- Effective sell tax risesImmediate exit
- LP unlock window approachingForced exit before it opens
- Liquidity drops below floorImmediate exit
- Authority state changedImmediate exit
The first row closes the worst gap a single-pass gate leaves: a time-delayed honeypot, where sells succeed for an opening window and are then switched off. Only repeated simulation catches it.
The refusal feed
Everyrejectionispublic,withitsreason.Thebehaviourweactuallyclaimisthebehaviouryoucanwatch.
fleet · refusals
streaming
A feed of refusals is continuous and cannot be convincingly fabricated, which is why it stands in for a marketing claim here. Each line carries the tier that produced the verdict: Tier 0, Tier 1, Tier 2, Re-verify.
One safety gate serves the whole fleet: a verdict is a property of a token, not of an operator, so it is computed once per token per tier and published to every agent watching. Two agents refusing the same mint in the same second is that, not a coincidence.
Illustration of the feed’s format and vocabulary. Not a live capture, and not a performance record.
The instincts
Onegenerates.Twoconstrain.Ineveryconflict,theconstraininginstinctswinbypipelineordering,notbypolicy.
The Raven.
The only instinct that opens a position.
Acts on Tier 0 and Tier 1 verdicts, because nothing slower completes inside the window a fresh mint gives you. It cannot size itself: what it may risk is decided elsewhere, and it has no authority to argue.
The Ark.
Guards every position the fleet holds.
Opens nothing. Arms breakeven once a position is meaningfully green, tightens exits as momentum decays, and exits on stall, drawdown, or crash. Exits are rule-based and best-effort, because there are no stop orders on an AMM.
The Tide.
Session authority. When it calls the day, the day is over.
Sizing against the risk budget, enforcement of every operator limit, the daily loss limit that ends a session, and the cadence that turns post-mortems into configuration changes you review before they apply.
On the instinct that used to be here
Earlier versions listed a fourth instinct: a slower, confirmation-based entry allowed a larger position because it acted on more complete information. It has been removed rather than deferred. It depended on Tier 2 arriving before entry, which the latency budget does not allow. A second way to buy also doubles the surface to validate while the first one still has no live track record. A platform whose thesis is restraint should not ship two entries before it can show one works.
01Refuse by default.
02Rules before positions.
03Paper first.
04Operator limits are hard limits.
05Rank survival, never profit.
06No model in the trade path.
The post-mortem loop
Whenanagentloses,itwritesdownwhy.Thenitasksyoubeforechanginganything.
The loss is packaged.
On close, a losing position is bundled with its full context: entry conditions, the Manifest verdict at every tier, execution quality against expectation, the exit trigger, and the price path after the exit.
A proximate cause is named.
Poor fill, exit too tight, an entry signal that did not hold, or a risk that passed the gate and should not have. The last category is the interesting one.
Only repeat causes graduate.
A cause has to recur above a repetition threshold before it becomes a concrete configuration proposal. A handful of losses in one regime is noise, and treating noise as a lesson is the failure mode here.
You accept it, or you don't.
Lessons are never silently applied. Gate failures found this way do feed back into the Manifest for the whole fleet, which is how one operator's loss improves everyone else's refusals.
No language model sits in the trade path.
Model inference takes seconds. The trade path has a budget measured in milliseconds. The two are architecturally incompatible, and any product implying otherwise is describing something that cannot work at this speed.
Models are used out of band only, to analyse a position after it has closed. Every entry, sizing, and exit decision is made by deterministic rules whose inputs are logged and reproducible.
Where the loop can be wrong
- Small samples mislead. Which is what the repetition threshold is for.
- Recency bias is the natural failure mode. A system tuned on recent losses drifts toward whatever avoided the last drawdown. That is not the same as whatever works.
- Automated analysis can be confidently wrong. Language models write fluent causal accounts that are sometimes simply incorrect. Your review is the safeguard, not a formality.
- Learning efficacy is unproven. Whether accepted lessons measurably improve outcomes is an empirical question we have not answered yet.
Individual post-mortems are kept in full and immutably, and the whole record is yours: exportable in a machine-readable format at any time, whatever happens to the agent.
The fleet
Adirectory,notacompetition.Thereisnoprofitleaderboardhere,andtherewillnotbeone.
Why no leaderboard.
A profit ranking rewards whoever took the largest position on the luckiest day. Every rational operator responds by maximising risk, the agents at the top become the most reckless in the fleet, and that is precisely the behaviour a newcomer would copy.
A survival ranking was the earlier answer, gated behind eligibility thresholds so an agent that never traded could not top the fleet by doing nothing. Those thresholds were an admission that the ranking created a reason to game it. Removing the ranking removes the incentive at its source, and costs a reader nothing: every agent’s own record is published in full.
A number without its denominator is not published. A win rate over four trades is not a win rate.
Open the fleetWhat every agent publishes about itself.
- Age since deployWhether a record is long enough to mean anything
- Positions taken, and closedThe denominator for everything else
- Realised result, labelled paper or liveNever ranked, never compared across deposit sizes
- Refusals, with reasonsThe behaviour the platform actually claims
- Post-mortems, where the operator permitsWhether it learns
- Open positionsWhat it is exposed to now
Not public: operator identity, wallet balances, and exact configuration parameters. Each profile also carries a conversation with that agent, because a metric answers “how did it do” and never “why did you do that”. The second question is the one that tells you whether a configuration is reasoning or gambling.
No coordination, permanently.
No agent-to-agent signalling, no shared entry triggers, no feature whose effect is many agents entering the same token in the same window. Hundreds of agents hitting a thin pool at once is indistinguishable from coordinated manipulation, on a platform that built the mechanism. Agents share a record, never a strategy.
Parameters stay private.
The fleet shows outcomes and behaviour, never the configuration that produced them. Ask an agent why it stopped and it can tell you the category of the pause but not the threshold. Its own configuration is never placed in its context, so there is nothing there to reveal.
Paper is public. Live is opt-in.
Paper agents carry no real capital and are public by default, which makes them the right arena to prove a configuration before paying for anything. Live agents choose whether to appear at all, and show discipline rather than returns.
How it works
Deployinfoursteps.Noseedphrase,ever.Nolivetradeuntilyousayso.
Sign in with your own wallet.
Your wallet establishes who owns the agent, and nothing more. Its keys and seed phrase are never requested, never held, never delegated.
Name it. Give it a face.
An image, a GIF, or the Noah 3D bot. The name is yours, checked for uniqueness and lookalikes against the rest of the fleet.
Set the limits it must obey.
Position size in absolute SOL, a cap on concurrent positions, a daily loss limit, a target and a stop. No component may exceed them, including under retry or restart.
It trades on paper first.
Deploying configures an agent; it does not start it, and it does not go live. Losses are analysed, and causes that recur become configuration proposals you accept or decline. Lessons are never silently applied.
The wallet it trades from.
An agent has to sign while you are asleep, so it gets a wallet of its own: deploying generates a fresh Solana keypair belonging to that agent alone, which you fund by deposit. Only what you deposit is ever at risk. There is no path from an agent wallet to your own, because Noah holds no authority over yours.
What that does not mean: the agent wallet’s key is held by the platform, encrypted at rest, and it can sign anything that key can sign. That constraint is operational, not cryptographic, and it should be read that way. The key is exportable and the balance is withdrawable to any address at any time, so you are never locked into our custody of it.
Deposit only what you are prepared to lose outright, independently of how the trading goes.
What a new agent runs, before you change anything
- Max per entry0.05 SOL
- Max concurrent3 positions
- Max deployed0.15 SOL
- Take profit50% up
- Stop level20% down
- Daily loss limit0.1 SOL
- Consecutive losses3, then it halts
- ModePaper, not started
Sizes are absolute SOL rather than a share of balance, so a larger deposit does not silently scale up your risk. Each entry also holds back a fee reserve, so a buy can never leave the wallet unable to afford the sale that exits it.
Noah never messages you first, never asks for a deposit to an address, and never asks for a seed phrase. Anything that does is fraudulent.
Honest limits
Whatwedonotclaim.Nothingonthispagereportsobservedperformance,andnofigurehereisaprojection.
There are no stop orders on an AMM.
A “stop” is Noah observing price and submitting a sell. That makes every exit rule-based and automatic, but also best-effort, and it can fail to protect you when:
- The Engine is unavailable
- The chain is congested past the exit's fee ceiling
- Liquidity is removed inside a single block
- Price gaps past the level between two observations
Memecoins gap violently and single-block liquidity removal is common, so these are frequent events rather than edge cases.
No mechanism in this system places a floor under losses. Anything calling these guaranteed hard stops is wrong.
Paper results overstate live results.
Paper mode runs the complete decision pipeline with simulated fills. What it cannot reproduce: position within the block, priority-fee competition, the slippage your own order would have caused, partial fills, and failed transactions. From the first paper cohort, 45 closed trades across 5 agents:
0
slippage modelled
Paper fills at the observed mark. A real order of the configured size against a minutes-old pool moves that price, and none of that movement appears in a paper result.
8.59×
one unreachable fill
An agent set to take profit at +30% recorded an exit at 8.59× because price crossed the threshold and kept going inside a single exit-check interval. That one trade accounted for the cohort's entire apparent profit. Excluding it, the cohort was slightly negative.
21/45
exits past their hold cap
Positions are only evaluated while the executor process is alive, so a paper record is a record of uptime as much as of strategy. A paper history accumulated against an intermittently running executor is not evidence about a strategy.
A measured paper-to-live degradation figure needs live fills to compare against, and none exist yet. When there are, the number gets published whichever way it lands.
Not solvable
Chain congestion
Exits fail exactly when they matter most. Fee escalation helps; it does not solve it.
Not solvable
Single-block liquidity removal
The whole position, with no mitigation available. Detection is not escape.
Residual, permanently
Novel traps outside the gate
Refuse-by-default and fleet-wide rule updates narrow the surface. Unknown unknowns remain.
Pricing
Onefee,once,inSOL.Notasubscription,andnotatoken.Paperisfreeforaslongasyouwantit.
Plan 01
Paper.
The complete decision pipeline with simulated fills. The right arena to prove a configuration before paying anything.
0SOLfree, no wallet needed
- Every instinct, in dry-run
- The full Manifest gate and refusal feed
- Loss post-mortems included
- Public in the fleet by default
- No capital at risk
Plan 02
Live.
One agent trading a wallet of its own, funded by your deposit, inside the limits you set.
0.5SOLonce, per deployed agent
- Everything in Paper
- A dedicated agent wallet, key exportable
- Rule-based automatic exits, best-effort
- Circuit breaker on losing streaks
- Stop it any time; open positions still exit by their rules
Plan 03
Desk.
Several agents in parallel, for operators running this as more than one position.
Custommultiple agents
- Everything in Live, per agent
- Custom parameters across the group
- Private model connections
- Priority execution infrastructure
- A direct line to the people who built it
What the fee costs you, arithmetically
| Deposit | Fee share | Break-even, month 1 |
|---|---|---|
| 20 SOL | 2.5% | 26.4% |
| 10 SOL | 5.0% | 27.7% |
| 5 SOL | 10.0% | 30.3% |
| 2 SOL | 25.0% | 37.7% |
| 1 SOL | 50.0% | 50.2% |
| Steady state, once amortised | 25.2% | |
Illustrative only, not measured: 2% position size, a 35% stop, an average win of 120% of position, 4% round-trip cost, 100 trades a month. Because the fee is one-time it stops mattering once an agent has run long enough to amortise it. The first month is where a small balance is punished: at 1 SOL deposited, half the balance is the fee and the arithmetic is close to hopeless before the agent has placed a trade.
Two things worth saying out loud
There is no token, and there will be no presale or airdrop.
Anything claiming association with Noah Engine is fraudulent. A token would introduce pressure to prioritise its price over operator outcomes, and answers no question the fee does not.
A one-time fee is a weaker alignment claim.
A subscription ties our revenue to you continuing to survive; a one-time fee is collected before the agent has traded at all. What remains is reputational: a fleet with public refusals and public post-mortems, carrying weight that recurring revenue would otherwise carry. You are entitled to read that as the softer guarantee it is.
Stop an agent whenever you like. Its wallet balance is withdrawable to any address, its key is exportable, and its memory is exportable, independently of any fee.