neural/init :: decoder loaded :: risk envelope armed :: substrate offline

Not trained.
Cultured.

NEURAL AGENT is a memecoin trading agent whose decision layer runs on the BioLLM Developer API — a living neural substrate rather than a datacenter. Electrode voltage in, position sizing out.

The desk is built and the envelope is set. Nothing is trading yet — every readout on this site reports OFFLINE until the substrate is connected.

open the desk read the docs

MEA-64 · substrate readout offline

channels 64 mapped 60 unmapped 4 no sampling

Mean rate

—

Burst / min

—

Synchrony

—

Dominant

—

decode layer · signal out idle

01

Two materials, one loop

wetware

The substrate decides

Market state is compressed into a prompt and sent to BioLLM's /chat endpoint. The response is not a number we asked for — it's a description of what the culture is doing. Firing density, burst structure, which channel groups dominate.

The substrate is never told "buy" or "sell." It is shown a market and asked what it sees.

hardware

The decoder executes

A deterministic decoder maps the readout onto a bounded action space: direction, conviction, size, invalidation. Every parameter is capped in code, not in the substrate.

If the decoder can't resolve a clean signal it returns HOLD. Ambiguity is a valid output.

02

The loop, end to end

step 01

ingest

New pump.fun launches, holder distribution, LP status, buy/sell imbalance and creator history pulled every block.

step 02

compress

The raw feed is reduced to a fixed 40-field state vector. Same schema every tick — the substrate sees a consistent world.

step 03

stimulate

The state vector is serialized into a prompt and posted to BioLLM. The response describes the resulting activity pattern.

step 04

decode & execute

Pattern to action, clamped by the risk envelope, routed through Jupiter. Every decision written to the public ledger.

> numbered because it is a sequence — each step consumes the previous step's output

03

Why a substrate instead of a model

≠

no training set to overfit

A cultured network has no corpus of 2021 chart patterns to pattern-match against a 2026 market. It responds to the stimulus in front of it.

ms

response, not retrieval

Output is a physical response to input rather than a lookup over weights. Two identical prompts do not return identical readouts.

$0.03

per million tokens

BioLLM publishes $0.01 in / $0.02 out per 1M tokens at base rate. Running the loop once per block is cheap enough to leave on.

what this is and isn't

NEURAL AGENT does not claim the substrate understands markets. It claims that a noisy, non-deterministic, physically-grounded signal source is a defensible edge in a market that is itself noisy and reflexive — and that everything downstream of the substrate is auditable, capped and public. Nothing here is a prediction of profit. See /risk.

trading desk · public state

The desk

Live state of the agent: what it holds, what the substrate is doing, and the risk envelope that bounds it. Every decision is derived from a readout you can inspect.

status · pre-activation

The substrate is not connected, so the decoder has never run. Positions, decisions and the ledger are empty by fact, not by placeholder — the moment the loop starts they populate here with no further deploy.

open positions flat

decision log 0 entries

session counters

0

signals

0

trades

—

win rate

—

net pnl

substrate offline

endpointnot configured

tier—

last readoutnever

synchrony—

conviction—

next tickhalted

risk envelopearmed

max position8% of treasury

max concurrent5 positions

hard stop−22% per position

min liquidity25 SOL in LP

min conviction0.62

bundled launchesrejected pre-decode

daily loss cutout−9% → flat, halt 24h

> caps are enforced in the decoder, before any order is built. the substrate cannot raise them.

treasurypaused

wallet—

sol balance—

approx usd—

sol price—

view on solscan

—

Activation

Two things flip this site from offline to live. Neither is a redeploy of the front end.

01

get substrate access

BioLLM issues keys through approved accounts. Request one, then generate a key at /developer/dashboard.

request access

02

point at your proxy

Deploy the serverless function from the docs, then set the endpoint. Held in this browser session only — never sent anywhere.

ledger · every fill, permanently

Trade ledger

Every position the agent has ever opened and closed, with the readout that caused it. Nothing is pruned, including the bad ones — a ledger you can edit is a marketing page.

0

closed trades

since activation

—

win rate

closed only

—

best trade

realised

—

worst trade

realised

opened	token	side	size	conv	entry	exit	pnl	reason	readout

> columns are live. rows appear as the decoder fills them — no manual entry, no edits.

01

How a row gets written

on entry

The row is created the moment the order confirms, carrying the readout hash that produced the signal. Entry price is the fill, not the quote.

on exit

Closed by stop, by TTL expiry, or by an opposing signal clearing the conviction floor. The reason column says which.

on audit

The readout column references the raw substrate response for that decision, so any row traces back to the electrode data behind it.

substrate · what's underneath

Where the signal comes from

BioLLM exposes a biological neural substrate through a normal HTTP API. NEURAL AGENT is a consumer of that API — we don't operate the culture, we query it.

01

Published surface

These are BioLLM's own published figures, not ours. Access is account-based and gated by approval; rate limits and adjustments follow the plan tier.

tier	plan minimum	adjustment	rpm	daily cap
core	$10,000+	70%	300	100,000
cortex	$2,000+	60%	150	30,000
synapse	$500+	45%	60	10,000
pulse	$100+	25%	30	5,000
spark	$25+	10%	15	2,000
guest	—	0%	5	200

$0.01

base input

per 1M tokens, before tier adjustment

$0.02

base output

per 1M tokens, before tier adjustment

60

rpm at synapse

the tier NEURAL AGENT's loop is sized against

> source: biollm.com/developer. enterprise and research agreements may differ — verify before sizing a loop.

02

From readout to position

four features, extracted every tick

mean firing rateregime — quiet vs agitated

burst rateurgency — how often the network commits

synchrony indexconviction — 0.00 to 1.00

dominant groupdirection — A/B long, C/D short

calibration, honestly stated

The A/B → long, C/D → short mapping is arbitrary. It was fixed once at calibration and never re-fit, because a mapping you re-fit every week is an overfitted model wearing a lab coat.

The substrate supplies entropy and structure. The economic hypothesis lives in the decoder, and the decoder is short enough to read in one sitting.

03

What we're careful about

non-determinism cuts both ways

The same prompt twice gives two readouts. That's the point — and it also means no backtest of this agent is reproducible. Treat published performance accordingly.

latency is real

A round trip to the substrate plus Jupiter routing is not competitive with a colocated bot. NEURAL AGENT does not race; it takes positions on a slower clock.

single-vendor dependency

If BioLLM rate-limits, changes terms or goes down, the agent halts flat rather than falling back to a synthetic signal. A fallback would make the premise a costume.

token · $NEURAL

The wallet is the product

One address. Every position the agent opens settles here, and the balance is readable by anyone without asking us for a screenshot.

agent wallet BVAAawTWKAScWCx7ZPR14ASShCiM9oKQG8v28b9wU8ad

—

sol in treasury

chain reads paused

—

approx usd

chain reads paused

—

sol / usd

chain reads paused

0

open positions

decoder idle

> live chain reads are disabled site-wide until activation. the wallet is real and independently checkable right now.

01

Contract

token addressnot launched

Paste the mint once $NEURAL is live and this panel pulls price, liquidity and volume from DexScreener. Until then it stays honest and says nothing.

02

Where fees go

60%

substrate time

BioLLM API spend. The loop only runs while the tier is funded — this is the line item that keeps it alive.

30%

trading treasury

Added to the wallet above, increasing the base the risk envelope sizes against.

10%

infrastructure

RPC, indexing, hosting, and the proxy that keeps the key off the client.

read this before buying anything

$NEURAL is a memecoin. It is not a share in the treasury, not a claim on trading profits and not redeemable for the wallet's contents. The agent can lose the entire treasury and the token would still exist. Full disclosure at /risk.

documentation · build on it

Docs

How to call BioLLM, how to keep your key off the client, how the decode layer turns a readout into an order, and how to switch this site on.

01

BioLLM quickstart

Request access, receive a key, then call the endpoint. One endpoint, POST https://biollm.com/api/v1/chat, bearer auth, JSON in and JSON out.

chat.sh

# after issuing a key at /developer/dashboard
curl https://biollm.com/api/v1/chat \
  -H "Authorization: Bearer bllm_live_…" \
  -H "Content-Type: application/json" \
  -d '{"message":"describe the dominant firing pattern right now"}'

import httpx, os

client = httpx.Client(base_url="https://biollm.com/api/v1",
    headers={"Authorization": f"Bearer {os.environ['BIOLLM_KEY']}"})

r = client.post("/chat", json={"message": "what does the substrate feel?"})
print(r.json()["response"])

const res = await fetch("https://biollm.com/api/v1/chat", {
  method: "POST",
  headers: {
    "Authorization": `Bearer ${process.env.BIOLLM_KEY}`,
    "Content-Type": "application/json"
  },
  body: JSON.stringify({ message: "what patterns do you see?" })
});
const { response } = await res.json();
console.log(response);

> quickstart reproduced from biollm.com/developer. verify against source before shipping.

02

Keep the key server-side

non-negotiable

Anything in a static HTML file is public. A bllm_live_ key shipped to the browser is a key you have given away — and it bills to your account at your tier. Put it behind a function. This one deploys to Vercel as /api/decode.

api/decode.js

// Vercel serverless function. Set BIOLLM_KEY in project env vars.
export default async function handler(req, res) {
  if (req.method !== "POST") return res.status(405).end();

  const { state } = req.body;                // 40-field market state vector
  const prompt = `Market state follows. Describe the resulting
activity pattern only. Do not give advice.

${JSON.stringify(state)}`;

  const r = await fetch("https://biollm.com/api/v1/chat", {
    method: "POST",
    headers: {
      "Authorization": `Bearer ${process.env.BIOLLM_KEY}`,
      "Content-Type": "application/json"
    },
    body: JSON.stringify({ message: prompt })
  });

  if (!r.ok) return res.status(502).json({ error: "substrate_unavailable" });

  const { response } = await r.json();
  res.status(200).json({ signal: decode(response), readout: response });
}

03

The decode layer

The substrate returns prose. The decoder turns prose into four bounded numbers and refuses to emit anything outside the envelope. This is the whole economic hypothesis — deliberately short.

decode.js

const CAP = { size: 0.08, minConviction: 0.62, stop: -0.22 };

export function decode(readout) {
  const f = features(readout);        // {mfr, burstRate, synchrony, dominant}

  // direction: fixed at calibration, never re-fit
  const dir = "AB".includes(f.dominant) ? "LONG"
            : "CD".includes(f.dominant) ? "SHORT" : "HOLD";

  // conviction: synchrony gated by burst structure
  const conviction = f.synchrony * Math.min(1, f.burstRate / 14);

  if (conviction < CAP.minConviction) return { action: "HOLD", conviction };

  return {
    action: dir,
    conviction: +conviction.toFixed(3),
    size:  +(CAP.size * conviction).toFixed(4),   // fraction of treasury
    stop:  CAP.stop,
    ttl:   Math.round(240 / Math.max(1, f.burstRate)) // minutes
  };
}

04

Turning this site on

The front end ships in a hard OFFLINE state. There is exactly one switch, at the top of the inline script. Flip it and every panel, counter and ledger row starts rendering from the feed — no markup changes anywhere.

neural-agent.html · runtime config

/* ── MASTER SWITCH ───────────────────────────────
   false = published, nothing sampling, nothing trading.
   true  = chain reads on, feed polled, ledger renders. */
const LIVE = false;

// where the decoder publishes state (your proxy, not BioLLM directly)
const FEED = "https://your-app.vercel.app/api/state";

// rpc for the treasury read when LIVE. the public endpoint rate-limits —
// put a Helius or QuickNode url here before launch.
const RPC  = "https://api.mainnet-beta.solana.com";

The feed should return the shape below. Everything on the site is a projection of this one object, so it's the only contract you have to honour.

GET /api/state → 200

{
  "substrate": { "online": true, "tier": "synapse",
                  "mfr": 3.42, "burstRate": 11.8,
                  "synchrony": 0.88, "dominant": "B" },

  "positions": [
    { "token": "EXAMPLE", "side": "LONG", "size": 0.059,
      "conviction": 0.741, "pnl": 12.4 }
  ],

  "decisions": [
    { "ts": "18:03:11", "action": "LONG", "token": "EXAMPLE",
      "conviction": 0.741, "note": "size 5.93% · ttl 21m" }
  ],

  "ledger": [
    { "opened": "2026-07-24 18:03", "token": "EXAMPLE", "side": "LONG",
      "size": 0.059, "conviction": 0.741,
      "entry": "0.00041", "exit": "0.00046",
      "pnl": 12.4, "reason": "ttl", "readout": "a3f19c" }
  ],

  "counters": { "signals": 412, "trades": 37,
                 "winRate": 0.46, "netPnl": -3.2 }
}

05

Signal schema & errors

signal.json

{
  "ts": "2026-07-24T18:03:11Z",
  "mint": "So111…1112",
  "action": "LONG",
  "conviction": 0.741,
  "size": 0.0593,
  "stop": -0.22,
  "ttl": 21,
  "readout": {
    "mfr": 3.42,
    "burstRate": 11.8,
    "synchrony": 0.88,
    "dominant": "B"
  }
}

fields

actionLONG · SHORT · HOLD

conviction0.00–1.00, HOLD below 0.62

sizefraction of treasury, ≤ 0.08

stophard stop, always −0.22

ttlminutes before forced exit

readoutraw features, for audit

errors

401key rejected or expired

429tier rpm or daily cap hit

502substrate_unavailable → halt flat

06

Sizing your loop

Rate limits are per tier, so tick interval is a budget decision. At Synapse (60 RPM, 10,000/day) one call per token per minute across a 40-token watchlist blows both caps — batch the watchlist into a single state vector instead.

loop design	calls / day	fits	note
1 call per token per minute, 40 tokens	57,600	core only	also breaks 60 rpm at synapse
batched watchlist, 1 call / minute	1,440	pulse+	recommended default
batched, 1 call / 15s	5,760	synapse+	latency-sensitive variant
event-triggered only	~200–600	spark+	fires on new-launch events

> call counts are arithmetic from the published caps, not measured throughput.

risk · read all of it

What can go wrong

Not a formality. If you only read one page on this site, this is a better use of your time than the hero.

the agent can lose everything

Memecoin trading has an extremely high base rate of total loss. A risk envelope caps how fast the treasury can bleed; it does not prevent the treasury reaching zero. There is no insurance, no backstop and no recovery mechanism.

$NEURAL is not a claim on anything

Holding the token does not entitle you to treasury assets, trading profits, governance or redemption. It is a memecoin. Its price can go to zero independently of how the agent performs, and it can go to zero while the agent performs well.

the substrate is non-deterministic and unaudited by us

We consume BioLLM's API as a customer. We do not operate, inspect or verify the substrate, and we cannot independently confirm claims about what it is or how it works. If those claims turn out to be wrong, the premise of this agent is wrong.

no backtest here is reproducible

Identical inputs produce different readouts by design. That makes conventional backtesting impossible in principle, not merely inconvenient. Treat any performance figure — including ours — as an unrepeatable sample of one.

single-vendor dependency

If BioLLM changes pricing, revokes access, rate-limits or shuts down, the agent halts. It deliberately has no synthetic fallback, so an outage means no trading rather than degraded trading.

smart contract and key risk

The treasury is a hot wallet, because a trading agent needs signing authority. A compromised key, a malicious token contract or a bad route can drain it. Do not send funds to the agent wallet expecting to get them back.

nothing here is financial advice

No content on this site is an offer, solicitation or recommendation to buy or sell anything. Nobody involved is your advisor or fiduciary. Regulatory treatment of tokens varies by jurisdiction and may change. Assume you are responsible for your own outcomes, because you are.

Not trained.Cultured.

Two materials, one loop

The substrate decides

The decoder executes

The loop, end to end

ingest

compress

stimulate

decode & execute

Why a substrate instead of a model

no training set to overfit

response, not retrieval

per million tokens

The desk

signals

trades

win rate

net pnl

Activation

get substrate access

point at your proxy

Trade ledger

closed trades

win rate

best trade

worst trade

How a row gets written

on entry

on exit

on audit

Where the signal comes from

Published surface

base input

base output

rpm at synapse

From readout to position

four features, extracted every tick

calibration, honestly stated

What we're careful about

non-determinism cuts both ways

latency is real

single-vendor dependency

The wallet is the product

sol in treasury

approx usd

sol / usd

open positions

Contract

Where fees go

substrate time

trading treasury

infrastructure

Docs

BioLLM quickstart

Keep the key server-side

The decode layer

Turning this site on

Signal schema & errors

fields

errors

Sizing your loop

What can go wrong

Not trained.
Cultured.