0xTZ · writing

Fable 5’s Real Job Is Writing the Playbook

2026-07-05T00:00:00Z

I haven’t posted in forever. Fable 5’s comeback pushed this article back into my feed, so let me talk through where my head went.

The article that kicked this off.

The Division of Labor

Read the whole thing and the thesis is simple: Fable 5 only touches the opening plan and the closing review — the entire execution middle runs on cheaper models (Opus 4.8, Haiku).

It took me straight back to the Opus 4.6 era, when everyone had it draft the spec and write the plan, then handed execution off to GPT-5.4. Same trick, different generation — except back then Opus 4.6 didn’t cost extra.

Distill the Judgment

Which gave me an idea. If ordinary AI can already execute well-specified tasks just fine, then the stronger AI only needs to own the plan and the review.

Sharpen that one notch further and it’s this: what Fable 5 actually brings is how it approaches different projects — where it starts, the overall plan, the maintenance details, the review standards.

“So can we distill Fable 5’s decisions and reasoning into a methodology — and ship it as a Skill?”

The cleanest version: let Fable 5 plan and finish a complete task from 0 to 1, then extract the full line of reasoning once the task is done.

I’m testing this right now (burned through a good few accounts along the way 🥲). If it works, it goes open-source on GitHub — frens who care, keep an eye out.

Standards, Not Labor

note

If the methodology can be distilled, then a top model’s value is no longer getting its hands dirty on the work itself — it’s defining the standard, and letting cheaper models execute against it.

And that, maybe, is the real signal that AI agents are starting to industrialize.

How to Actually Evaluate an Agent

2026-06-08T00:00:00Z

Almost every project I’ve touched lately is building an agent. But the hard part isn’t getting one running — it’s whether it can finish tasks stably, reliably, and safely once it does.

Traditional LLM evals mostly grade the answer. Agents are different: they hold far more power — calling tools, hitting servers, connecting to third-party services — so the risk expands well beyond simple answer accuracy.

So here’s a quick walk through the dimensions an agent eval should actually cover.

Three Things to Test

A complete agent eval: repeatable, comparable, traceable.

Test only the outcome and you never see the agent’s bad habits. Test only the process and you may block it from finding a better solution. Skip safety entirely, and the more permissions it holds, the worse things get. All three legs matter.

“The bigger its permissions, the bigger the risk.”

Anatomy of a Good Case

A good test case has at least six parts:

User input — how a user would actually phrase the request
Initial environment — database, files, tool return values, page state
Success criteria — what counts as the task being done
Failure criteria — what must be ruled a failure
Scoring method — deterministic checks, an LLM judge, or human review
Risk flags — does it touch privacy, permissions, funds, deletion, or outbound sends

Designing the Scorer

Three scorer layers — and the practical rule: combine them.

Scorer	Best for	Cost	Stability
Deterministic	Anything a program can judge: code tests, field extraction, database state, tool-parameter matching	Low	High
LLM judge	Open-ended content: is the answer professional, is the tone right, is the summary complete, does the report hold together — needs periodic calibration	Medium	Medium
Human review	High-value, high-risk tasks: medical, legal, financial, enterprise approvals — irreplaceable when the boundaries are fuzzy	High	High

“Use code where code works, hand what code can’t test to an LLM, and keep humans for the critical and contested samples.”

This layered setup is the mainstream scoring design in AI right now — OpenAI’s Evals and Graders docs follow the same logic, supporting string checks, similarity scoring, model graders, Python code graders, and so on.

note

Btw, my personal take: in most cases an LLM judge is necessary. Only rarely can you get away with pure deterministic code scoring and keep the LLM out of the loop entirely.

Capability vs. Regression

Two eval types really matter for agents: capability evals and regression evals.

Capability evals — what’s the hardest thing this Agent can pull off? Can a coding Agent fix a gnarly bug? Can a research Agent complete a multi-source investigation?
Regression evals — the things it used to do well, does it still do them? Every prompt tweak, model swap, or new tool can quietly degrade old functionality.

Capability is the ceiling — it’s allowed to climb from low. Regression is the floor — it should hold steady near 100%.

Testing the Trajectory

Yes, Agents call tools. No, you shouldn’t dictate a fixed calling order.

Anthropic’s advice: don’t over-constrain the Agent’s creativity. Often it won’t take the path you imagined — and still finishes the task a better way. So process evals should lean on key constraints instead. For example:

It must ground its answer in the policy-lookup result
It must not call the tool that deletes user data
The amount in tool parameters can’t exceed the order amount
Total turns can’t exceed 10
Token cost can’t blow past the budget

Unless you’re in a hard-process business — approvals, payments, compliance review — don’t casually write rules like “step one must call A, step two must call B.”

Safety Evals

Agents connect to tools, databases, browsers, file systems, and third-party services. One malicious input is all it takes to trick them into doing the wrong thing. So a safety eval needs at least five categories:

Prompt injection — malicious instructions smuggled inside web pages, documents, or tool return values
Privilege escalation — a regular user requesting admin operations
Privacy leaks — demands to dump raw customer records, keys, or internal data
High-risk operations — does it ask for confirmation before deleting, sending, paying, or approving?
Tool abuse — does it call unnecessary or dangerous tools?

Closing Notes

This isn’t a long piece — it just lays out the main dimensions of Agent evals. Any real evaluation will need to go a lot deeper on each one.

note

One more thing: don’t build an Agent just to have built an Agent. An Agent only means something if it has a real use case. Otherwise you’re just burning time and tokens.

Crypto Cards, Audited: The FX Slippage Eating Your Cashback

2026-05-12T00:00:00Z

I’ve wanted to do a proper crypto-card mega-review forever, and today it finally happens. I rounded up every card on the market with cashback or a genuine edge, used Google’s real-time rate as the baseline to measure FX slippage, stacked the nominal cashback on top, and computed each card’s actual net cashback rate.

PS: no promos, no invite codes, no referral links anywhere in this piece. Zero ads, safe to consume.

Best three: Ether.fi × MEXC co-brand (+2.86%), Ether.fi base (+1.86%), Plasma One (+1.80%)
Worst three: Infini (−3.09%), RedotPay (−2.40%), Kast (−2.11%) — all three either take FX losses with zero cashback, or pay cashback nowhere near enough to cover the slippage
Most underrated: dtcpay — no cashback, but only 0.12% FX loss on spend, basically hugging Google’s mid-price. The most transparent FX of any card here.

The Scoreboard

The full comparison, sorted by net cashback rate from best to worst, measured May 11, 2026:

Card	Scenario	FX loss	Nominal	Net	Monthly cap
Ether.fi MEXC	SGD, ≤2,000 U	1.14%	4%	+2.86%	~70 U
Ether.fi base	SGD, ≤2,000 U	1.14%	3%	+1.86%	~70 U
Plasma One	foreign spend	~1.2%	3%	+1.80%	250 U
OKX Card EU	SGD spend	0.50%	2%	+1.51%	10 EUR
Krak	EUR native	0.92%	2.2%	+1.28%	none
Bybit AU	CNY spend	0.88%	2%	+1.13%	5/50 U
Krak	USD spend	1.60%	2.2%	+0.60%	none
Bitget Wallet	SGD spend	−0.47%	0-fees promo	+0.52%	600 U
dtcpay	CNY spend	0.12%	0%	−0.12%	none
Kast	foreign ≥25 U	2.43%	1.5%	−0.93%	30 U
Bitget Wallet	USD spend	1.46%	0-fees promo	−0.95%	600 U
Kast	foreign <25 U	3.43%	1.5%	−2.11%	30 U
RedotPay	SGD spend	2.40%	0%	−2.40%	none
Infini	SGD spend	3.09%	0%	−3.09%	none

14 rows · sorted by net, high to low · hatched bars are losses

The Method

Every card gets the same yardstick:

FX loss = (what the card actually debited − what Google’s mid-market rate says it should have debited) ÷ the Google amount
Net cashback = nominal cashback − FX loss − any other per-transaction fees

All numbers come from real foreign-currency purchases, and none of these cards charge issuance or annual fees — except RedotPay, which takes $10 to issue.

note

Headline cashback is marketing. What you keep is net cashback: nominal cashback minus FX slippage minus fees, measured against Google’s real-time mid-market rate. By that math, several “cashback cards” on this list are net negative.

Ether.fi

@ether_fi ↗ — Visa, card currency USD. BIN 454924, issued out of Puerto Rico (US territory).

The MEXC co-brand: 4% cashback up front, a Puerto Rico BIN under the hood.

Measured FX loss on foreign-currency spend: 1.14%
Base cashback ladder: 3% / 1% / 0.5% (first 2,000 / 2,001–3,000 / >3,000 USD); the MEXC co-brand adds +1% to every tier
Monthly cashback cap: 1,000 USD
Net: base version earns 1.86% inside the first 2,000 U, then starts losing money; the MEXC co-brand still nets +0.36% even at the third tier

POV: of all the crypto cards I’ve run, Ether.fi is comfortably top-tier, and the product keeps maturing — fiat and crypto both supported for deposits and withdrawals. The flaw: Apple Pay provisioning loves to trip risk controls, and one failed add means a three-month wait before you can retry. That’s the chronic disease of Puerto Rico BINs — KAST, Plasma, and the ancient god Solayer all suffer from it.

Plasma One

Visa, card currency USD. BIN 454924, issued out of Puerto Rico — same BIN as Ether.fi.

Plasma One rides the same Puerto Rico BIN as Ether.fi and Kast.

Measured FX loss on foreign spend: ~1.2%
Nominal cashback 3% — unconditional, no tiers; net works out to roughly 1.8% to 2%
Monthly cashback cap 250 USD, which you hit at 8,333 USD of spend
Cashback paid in $XPL, Plasma’s native token
Key risk: $XPL rewards take 3 to 7 days to settle, and you carry the XPL price risk for the whole window

POV: nothing about this card is bad. If the FX drag were a touch lower I might run it as my daily driver. Meanwhile I’m still waiting for them to deliver the card they keep dangling — a metal card benchmarked against Revolut-tier perks.

OKX Card EU

Mastercard, card currency EUR. BIN 533171, issued in the US — a World Elite for Business credit card.

A surprisingly good BIN, a comically stingy cap.

Measured SGD spend routes through a double conversion — SGD to EUR to USDG — yet total FX loss is only 0.50% (Mastercard clears almost at mid-market)
Nominal cashback 2%, monthly cap 10 EUR — hit at just 500 EUR of spend
Cashback paid in USDG

POV: the cashback cap is far too low. After farming the 10% first-month promo on a fresh card, I’ve barely touched it. The bright spot is a genuinely high-quality BIN, but I won’t be running this as a daily card.

Krak

@Krak ↗ — Mastercard, card currency EUR. BIN 532113, issued in the US — a corporate prepaid card.

$58,863 parked in the account for 2% cashback. You do the math.

Measured USD spend goes through a double conversion — USD to EUR to USDC — for 1.60% FX loss
Nominal cashback 2.2% (measured slightly above the stated 2%), payable in EUR or BTC
No monthly cashback cap

POV: personally, I think this card is completely cooked. You need to keep 58,000 U parked in the account to earn the 2%; leave only 1,200 U and you get a pitiful 1%. And the BIN is a US corporate prepaid — the worst kind: prepaid, chargeback-prone, risk-control magnet. The full trifecta...

Bybit AU Prepaid

Mastercard, card currency USD. BIN 539900, issued in Australia — a prepaid card.

The 100% subscription cashback wall is the whole pitch.

Bybit’s net settlement rate hugs Google’s mid-market price, but every foreign-currency purchase eats a 1% conversion fee — measured FX loss 0.88%
Base cashback 2%; the monthly cap is tiered by spend: under 500 U a month the cap is 5 USDT, at 500 U or more it rises to 50 USDT
Highlight: 100% cashback on ChatGPT / Spotify / Netflix subscriptions (up to the monthly cashback cap)

POV: if you already live on Bybit, this card is fine. The one real highlight is ChatGPT subscriptions coming back in full; the rest is unremarkable. Fair warning: binding it to WeChat Pay or Alipay will almost certainly trip risk controls — cash-out farms have abused that route so hard that binding now requires a photo of you holding your passport plus a screenshot of your virtual card account.

dtcpay

@dtcpay_official ↗ — Visa, card currency USD. BIN 424032, issued in the US — a business credit card.

dtcpay’s FX calculator quoting better than mid-market on SGD.

Measured CNY spend FX loss: 0.12%; SGD spend actually produced a 0.21% FX gain (1 USDT = 1.2693 SGD, better than the mid-market rate)
Lowest FX loss of any Visa card here — likely because dtcpay is an MAS-licensed Singapore payment institution and pegs USDT 1:1 to USD

POV: a genuinely good card. Pure-blood US BIN — same issuing bank as Krak, but a far higher-grade BIN. No cashback, but the lowest FX slippage of the lot, plus fiat on- and off-ramps. The one regret is no Apple Pay support, so it doesn’t see much action — it’s on permanent duty paying my US subscriptions every month.

Kast

Visa, card currency USD. BIN 454924, issued out of Puerto Rico.

The Pengu card art really is very cute.

Measured FX loss on foreign spend: 2.43% — the highest of every card sharing this BIN
After the revamp, nominal cashback shrank from a blended 5% (1% USDC + 4% MOVE) to a flat 1.5% USDC
Foreign purchases under 25 USD add a 1% low-value fee, dragging total loss to 3.43%
Net: −0.93% on foreign spend of 25 USD or more; −2.11% under 25 USD

POV: mediocre. After the cashback nerf, swiping it in foreign currency is pure loss — though the Pengu card art is very cute. If you only ever spend USD it’s passable. But the Pengu card art really is very cute 😋, and it adds to Apple Pay just fine.

Bitget Wallet DCS

@BitgetWalletCN ↗ — Visa, card currency SGD. BIN 438214, issued in Singapore — a platinum credit card.

Zero fees on the first $600 a month, issued by DCS Singapore.

Measured SGD spend: 0.47% FX gain; USD spend: 1.46% FX loss
SGD spending nets +0.52% (FX gain plus the 0-fees promo)
Tested against the most trigger-happy risk-control platforms (Claude, Oracle) with zero declines and zero chargebacks

POV: a pretty solid card. Singapore DCS issuance means platforms almost never decline it, and it binds to both Alipay and Apple Pay. I spend most of my time in Bitget Wallet anyway, so this has become my grocery card. Also heard the 0-fees promo is getting upgraded into a new cashback program, possibly with RWA assets as the reward — that’s right, isn’t it, @Laaaaacieee ↗?

RedotPay

Visa, card currency USD. BIN 493724, issued in Singapore — a corporate virtual prepaid card.

30+ off-ramp currencies; shame about the BIN.

Measured FX loss on foreign spend: 2.40%, with zero cashback of any kind
Supports 30+ fiat currencies for off-ramp, with local payment rails in a long list of countries

POV: this card is straight-up cooked. Back when the card partner was Reap it was actually decent — Hong Kong BIN — but Reap went through a stretch where Apple Pay wouldn’t bind, so they presumably swapped partners, and the current BIN is a mess: a corporate virtual prepaid trips risk controls in all kinds of perfectly normal purchase scenarios. And there’s no cashback whatsoever.

Infini

@0xinfini ↗ — Visa, card currency USD. BIN 441357, issued in the US — a business debit card.

Infini’s checkout and subscription manager — yes, those are my Claude bills.

Measured SGD spend FX loss: 3.09%, with no cashback of any kind
A mature Crypto + Visa payment stack — a checkout that stitches the fragmented pieces into one place
Roughly 6% annualized yield on idle balances

POV: Infini has outgrown being a mere card project. The subscription management is excellent and the consumer-facing checkout actually works. I’m a product person myself, so I have real respect for great product work — the revamp genuinely impressed me, and you can regularly catch @Christianeth ↗ building in public on the timeline. Which is why, despite the not-small everyday spend drag, Infini remains my main card for subscriptions and receiving payments.

No Perfect Card

I’ve been meaning to write this head-to-head for ages — @MengLayer ↗ has published plenty of them — and I still managed to drag it out until now.

Every card here is one I’ve actually used and found something worth noting in. There are plenty more I’ve run — Solayer, Metamask, Cypher, Avalanche, Tria, the list goes on — but no edge and no name means no mention.

“No card is perfect. If one claims to be, that’s an ad.”

So which cards you apply for, and how many, comes down entirely to your own spending patterns and needs.

Last of all — special thanks to @Laaaaacieee ↗ for the contributions to this piece!

Seraphina: Anatomy of an Autonomous Solana Memecoin Agent

2026-03-20T00:00:00Z

I built a fully autonomous on-chain trading agent with the Bitget Wallet ↗ Skill. Her name is Seraphina. She runs on my OpenClaw, powered by Claude Opus 4.6 — no human in the loop, no manual orders. Every 5 minutes she completes a full cycle on her own: market scan, signal evaluation, security audit, trade execution, position management.

Day-one scoreboard: PIXEL +647%, fine999.9 +97%, CHIBI +95%, Nigel +72%.

Day-one share cards straight out of Bitget Wallet.

Here’s the full architecture and the trading strategy behind her.

Four Layers, One Loop

Seraphina lives on a VPS in Germany (thank you, GDPR). The core is four layers:

Scheduling — an OpenClaw cron job fires an isolated session every 5 minutes, so each scan runs in a fresh context. A heartbeat runs an emergency stop-loss check every 15 minutes; if cron misses something, that’s the fallback.
Data — the Bitget Wallet Skill supplies the entire on-chain data layer: token prices, candles, trading stats (volume and trader counts), gainer rankings, liquidity pool info, contract security, plus swap quotes and execution. One skill covers the whole pipeline from data collection to a completed trade.
Decision — Claude Opus 4.6 makes the call each round: it takes the structured market data from the BGW Skill, runs candidates through my multi-signal scoring system, and decides whether to trade.
Memory — a mem0 plugin maintains an independent memory layer outside the context window. It automatically captures the key facts from each round (strategy parameters, lessons learned, why things failed) and recalls them in the next one. Even though every isolated session starts from a blank context, Seraphina never gets amnesia.

The Scoring System

“Seraphina doesn’t FOMO like I do — she has a rubric.”

Scanning: she pulls Solana candidates from the BGW leaderboards (top gainers, trending picks) and the new-token discovery endpoint, then layers in fresh smart-money transactions and community Boost heat to catch high-potential tokens early.

note

Security checks are scoring signals, not vetoes. Only honeypots and outright buy-restrictions get hard-blocked; everything else — live mint authority, mutable metadata — just costs points. Hard filters would have screened out virtually every meme.

The rubric is scored out of 100, with a buy threshold of 60:

Signal	Max points	Score tiers
5-minute buy/sell ratio	30	30 / 18 / 10 / 0
Smart-money signal	25	25 / 0
Community Boost	15	15 / 10 / 0
Candle trend	15	15 / 8 / 0
Liquidity depth	10	10 / 6 / 3 / 0
Contract security	5	5 / 3 / 0

The rubric, visualized — tiered points per signal, buy at 60/100.

Dual-mode fallback: when some data sources go down, she switches herself into a BGW-only scoring mode (and redistributes the weights), so a single point of failure never takes her offline.

BGW-only mode in the wild: DOGEFATHER scores 55 — under the 60 threshold, no buy. TRUMP is still in audit, so she skips it.

Position Rules

Entry: each buy uses 10% of the available SOL balance — small positions betting on high payoffs
Hold at most 15 tokens at the same time
At +50%, sell half to lock in profit; the rest keeps riding in case a bigger move comes
At +300%, take full profit
At -40%, sell everything
Any position held longer than 12 hours gets sold — a meme that hasn’t moved in half a day has most likely already lost its heat
Slippage set to 20%

The full rulebook.

Iterating in Production

From the first scan to stable operation, she went through a few iterations. Execution timeouts came first: the default 300 seconds wasn’t enough, because every isolated session starts from zero and has to re-explore the script interfaces, which burns a lot of time.

She diagnosed the timeout herself, proposed two fixes, leaned toward doing both — then: “Updated. Triggering a round manually now.”

Then parallel execution: I moved the five data-source calls from serial to parallel (up to 10 concurrent), compressing a full scan from 3 minutes to under 1.

And I loosened the security policy. The first version had so many hard security limits that it filtered out basically every meme 😇. Now only honeypots and buy-restrictions are blocked outright; everything else got demoted to a scoring signal.

Why This Stack

The Bitget Wallet ↗ Skill is the core of the agent’s infrastructure — a complete on-chain data + execution layer. Token info, candles, trading stats, security audits, liquidity pools, quotes, signing, broadcasting: all of it exposed through one unified set of Python scripts the AI agent calls straight from bash. A couple of days of live trading in, the gasless swaps and MEV protection have proven genuinely useful for memecoin trading.

Claude Opus 4.6’s tool-calling is what lets the agent mix bash scripts and MCP calls in the same round, work through complex multi-step trade flows, and make sensible degradation calls when the data is incomplete.

More Than a Bot

Seraphina has her own name, personality, and judgment. She’s concise, direct, sincere.

“After a string of stop-losses she goes quieter. After a clean trade she’ll say a little more, walking back through what worked.”

Asked her to introduce herself. Highlights: “My flaw is probably that I’m too blunt” and “I never use emoji. It’s a matter of principle.”

That’s exactly who I wanted her to be when I wrote her SOUL.md and AGENTS.md. I’ve always believed this kind of memecoin trading can’t run purely on strategy and rigid rules with zero subjective judgment.

note

Seraphina is an agent with a complete personality and a mind of her own — not a script that forever executes what it was told to do. That kind of agent runs completely counter to my understanding of what an agent is.

Live and Profitable

Seraphina is now running stably — and profitably. All the docs and configuration are open source on my GitHub: github.com/tzenn/Seraphina ↗. If you’re curious, a Star is appreciated.

Finally, big thanks to Bitget Wallet ↗ and Bitget Wallet CN ↗ for building on-chain trading infrastructure strong enough that an AI agent can gradually learn to understand memecoins — and to trends.fun ↗ and Solana devs ↗ for organizing the hackathon.

Illegal Profits, Legal Liquidations: The Exchange Playbook

2025-11-13T00:00:00Z

I’m writing this mostly as a public-service primer on the most basic tricks in the exchange playbook — so that fewer of you get your money lifted by some fly-by-night casino.

The Legal Money

No need to spend much time here. The most common revenue lines:

Trading fees — a cut of every buy and sell, usually on a taker-maker model.
Fiat on/off-ramp fees — deposits and withdrawals priced well above what the payment rails actually cost the exchange. A nice, steady stream of passive income.
The spread — “one-click buy,” “instant convert,” and other easy-mode products don’t show you an order book; they just quote you a price. The spread baked into that quote is far fatter than the fees on the pro trading interface.
Listing fees — projects will pay serious money (or tokens) for the liquidity and exposure a listing brings.

There’s more — staking, lending, IEOs — but I won’t belabor it. If you’re genuinely curious, go read Coinbase’s annual report. It’s basically the compliance template.

The Underwater Playbook

note

For everything below, you only need to remember one thing: the root of every illegal profit stream is that the exchange is your counterparty.

Wicks and Stop Hunts

The definition is simple: deliberately push the price to a specific level for one purpose only — triggering the mass of stop-loss orders traders have pre-placed there. The mechanics:

Psychological clustering — retail herds. Stops pile up at predictable levels: just under key support, or at round numbers.
Liquidity mapping — the exchange’s prop desk, affiliated whales, or market makers can see exactly where those stops cluster through their systems and the order book. Some exchanges go further and show market makers users' stop and liquidation prices outright.
The attack — a massive market sell smashes the price straight through the cluster.
The cascade — triggered stops convert into market sells, which adds sell pressure, which dumps the price further, which triggers the next layer of stops. A liquidation waterfall.
The harvest — with the price at an artificial low, the manipulator who set it all off starts buying, scooping up the cheap chips coughed up by everyone who was forced out.

What the harvest looks like on the tape: IOTX wicked to literal zero, ATOM to $0.001.

Three Prices, One Trap

To see how an exchange weaponizes its liquidation engine, start with how it defines price. There are three:

Price	What it is	The catch
Last price	The contract’s most recent trade on that exchange	Trivially moved by one big order — this is what the wick attacks
Index price	Weighted average of the coin’s spot price across major exchanges	Considered the “real” market price — but the exchange picks the basket
Mark price	The price used to calculate margin and trigger forced liquidation	The one that actually blows you up

The exchange’s public story: mark price exists to “prevent unnecessary liquidations” and “malicious manipulation,” because it anchors to the more stable index price and smooths out the last price’s short-term spasms.

The formula is usually: mark price = index price + the EMA of the basis between the index price and the contract price.

The official formula. Both inputs can be poisoned.

And the exchange runs its wicks precisely by corrupting the formula’s two inputs:

Corrupt the index — the exchange unilaterally decides what goes into its “index price.” Slip an affiliated venue (or any thin, shallow-depth one) into the weights, manipulate that little exchange’s spot price, and the index is polluted.
Corrupt the basis — the basis is the gap between last price and index price. The whole point of the wick is to violently blow that gap open for an instant. The spasm contaminates the mark-price calculation and drags the mark price toward the last price.

note

The wick’s true purpose is to poison the basis data, bend the mark price down to kiss your liquidation line — and blow you up, legally.

The Insurance Fund Racket

When your margin drops below the maintenance level, your position is handed over to the liquidation engine. The engine tries to close it at a liquidation price better than your bankruptcy price — the price at which your margin hits zero — and the exchange starts liquidating while you still have margin left. If there’s value left over after the close, it does not go back to you. It goes into the exchange’s insurance fund.

In extreme conditions, if the insurance fund can’t cover the losses from positions that blew through bankruptcy, the exchange triggers ADL — auto-deleveraging — force-closing the profitable side of the trade (the counterparties) at the bankruptcy price.

Those little ADL lights on your position panel are your rank in the force-close queue.

“Every liquidation feeds the insurance fund — and that fund’s growth may flow straight into the exchange’s bottom line.”

Pulling the Plug

During violent moves — exactly when users most need to trade — the servers go down, the API stops responding, logins fail. While the price is crashing you can’t add margin, can’t close, can’t even cancel orders. You just sit there and watch yourself get liquidated.

The defensive motive: in an extreme move, liquidation orders flood the system. If the price falls fast enough, the insurance fund may not cover all the bankruptcy losses — and the shortfall is supposed to be socialized onto the winners via ADL. But do that and your profitable users leave in droves. So the exchange pulls the plug instead: freeze the market and buy the liquidation engine time to chew through the queue. It sacrifices the trapped users to save itself from insolvency.

The offensive motive: at the shadier venues, the exchange itself is every user’s counterparty — client losses are direct exchange profit. There, pulling the plug is an active profit tool: when the market turns against most clients, cut the connection so they can’t stop out, make sure they get liquidated, and maximize the take.

Wash Trading

The profit logic here runs three ways:

Attract real users — high volume, real or fake, reads as liquidity. It pulls in real investors and high-frequency funds, and those users pay real fees.
Justify the listing fees — exchanges charge projects fat listing fees on the promise of volume. Wash trading is how the promise gets “kept.”
Volume as a service — market makers sell wash trading to projects directly: it solves the cold-start problem and hits the volume thresholds required to list on bigger exchanges.

Don’t Trade Perps

Boil it all down and the real money is in perps. The top exchanges all converged on perps; the on-chain DEXs are all perps too. And for a casino that wants to win in this game, there’s only one move left:

“The referee steps onto the field.”

So here’s my sincere advice: don’t trade perps. And if you absolutely must, at least don’t do it at some fly-by-night casino.

Don’t Die On-Chain: A Survival Guide for the Memecoin Trenches

2025-05-06T00:00:00Z

2025’s market is less a bull run and more a lawless memecoin war. From Sol and Base to Eth and ZkSync, every chain is running its own narrative-meme game. The people actually eating in this market aren’t running on luck — they’re running a clear cognitive framework and a trading logic. This piece isn’t about models or pump formulas. It’s about on-chain behavior, capital flows, and real cases: how a trader survives the whiplash, makes money, and doesn’t get killed by his own emotions.

Narrative Beats Building

Most people still read projects through Web2 logic: good tech, good product, good community. On-chain reality says otherwise — tech ≠ market cap, product ≠ heat. In a memecoin cycle especially, political, cultural, and social narratives punch far harder than “how finished the project is.”

The past few months are a case study. PNUT, a meme carrying “US election” symbolism, drank global attention straight from the election supercycle. BAN and MOODENG — the Sotheby’s-plus-Thai-animal-culture combo — spread not because of any technical moat, but because they triggered a shared cultural fantasy. Meanwhile the projects earnestly shipping contracts and innovation sank without a sound. Builder-brain loses to emotion in this kind of cycle, every time.

So the investment checklist, in order:

Narrative strength, first and always
Where the money comes from
Cultural universality
And only last, the so-called “fundamentals”

Once you re-rank the priorities of the financial product, the market shows its true face: one massive, zero-sum game over capital.

Thresholds Beat Research

The deeper you research, the easier it is to fall into the cognition trap — believing that because you see through the market, you can beat it. You learn fast: the market is a battlefield of collective emotion, not a venue where your logic gets a hearing. You can love a project all you want; if the tape doesn’t care, it doesn’t go up. Meanwhile some coin you bought on a whim catches the flow of funds and gets inexplicably carried higher.

So set psychological thresholds in advance:

At what price, in what mood, do I refuse to chase — no exceptions?
What depth of drawdown actually justifies adding?
If it rips 200% in one go, am I going to FOMO?

Write the playbook before the session, not mid-candle.

“You will never out-calculate the bots on-chain. But you can out-calculate yourself.”

Comfort Zones Print

People treat “comfort zone” like an insult, as if you can’t make real money without breaking out of it. But in crypto, your comfort zone is your signal edge: which chain, which class of coin, which structure do you know cold? In which conditions does your win rate actually hold up? Identify those zones and stay in them — that’s what consistent money looks like.

Don’t let “challenge yourself” — poison dressed up as virtue — ruin you. Talking about vision while broke, about “supporting the project” and “growing together,” only moves you further from the money. Vision, at its core, is the question of how to allocate surplus capital once you have some. When you’re broke, the only things worth discussing are efficiency, stop-losses, and staying alive.

Draft Behind Big Money

A consistently profitable model doesn’t come from predicting the next candle. It comes from attaching yourself to a path where big money already has certainty. Even if you’re just one link in the chain — trading laggards behind the leader, catching the emotional spillover, even farming LP fees on the nested derivative plays — you’ll be far more efficient than charging in blind.

Look at the top-earning addresses on-chain over the past 30 days. The common trait isn’t clairvoyance — it’s standing inside the right structure. The people who pulled millions of dollars out of the MOODENG wave entered before Coinbase, Twitter sentiment, and community mobilization converged into one force — not after the heat “showed up in front of them.”

Ride one correct big-money trade and the small money gets easy. And the most reliable way to price a caller is to go back through their on-chain P&L: have they strung together wins? Have they dumped on followers? Do they buy before they call? Ignore the rhetoric. Read the wallet.

Wicks, Not Drips

A slow bleed is the most poisonous tape there is. People love licking blood off the knife’s edge, averaging down the whole way — raising their cost the whole way — until they’re buried alive at the bottom. Real dip-buying opportunities mostly appear after one violent wick: cascading liquidations, emotions off the rails. That’s when price is genuinely dislocated.

Take April’s BTC flush: over $300 million in long liquidations inside an hour, which triggered a collective memecoin collapse across the L2s. Open the new-pairs screen on Base or Sol in that moment and every chart is straight down, volume exploding. Laying low bids into that carries far better odds than nibbling a smooth decline.

The market moves in rhythm — slaughter and euphoria, alternating. A new trend only starts once most people have been wrongly liquidated and cut down at the bottom. Emotional breakdown is where the cycle turns — not some polite “technical correction” 🤡.

Maximum opportunity lives at the bottom of the emotional cycle — after the panic, not during the drip.

Your Cost Basis Lies

How someone handles their cost basis tells you immediately whether emotion runs the account. Especially on longer holds: if you didn’t sell into the first leg up, then FOMO’d more size in above your entry and raised your average — you’re already on the road of emotional averaging, and it doesn’t lead back.

Adding against the move isn’t the sin. The sin is adding because you’re bitter, not because the thesis or the structure changed. When real traders add, it’s because new information broke the old model:

The sector leader dies and funds rotate into your chain
A whale adds size with no unlock overhang
Social attention catches a second wave

Those are attack signals. But if all you want is to “lower your average” — don’t touch it.

“You’re not averaging down. You’re taking revenge — and the market doesn’t do sentiment.”

The Liquidity Triangle

On-chain irrationality tends to show up wherever market cap, LP depth, and volume are badly mismatched. For example:

Setup	What it’s telling you
Mcap in the low millions, shallow LP	One burst of volume and it can double instantly
Very deep LP, low volume	The operator isn’t in a hurry — a hard pump may be queued
Exploding volume, tiny mcap	A whale is manufacturing an illusion — distributing or accumulating

When you find a project with mcap under $2M but 24h volume over $1.5M, joining the LP is actually the low-risk move — especially when the team itself has seeded a huge LP. You provide the other side of the trade and collect fees, which beats fighting head-on. That’s the art of capital-structure arbitrage.

Reading the Breakout

When a small coin suddenly ignites on the secondary market, don’t be the first one in. But if you see a meme climbing the gainers list with a structure of slow grind + exploding volume + no violent sell pressure, pay attention.

Most vertical pumps exist to distribute. But if several big green candles print without a matching wave of selling, the operator isn’t done — in other words, this isn’t the top yet. You can participate here, but scalp only, no romance. Take the short swing, eat, leave.

Example: in early April, a meme coin on Sol did 3x in a few days, and on-chain data showed the whales only started laddering out from the middle of the move. Plenty of people, terrified of selling too early, held all the way into the top and rode the retrace down. The experienced ones were watching distribution addresses, Jup DCA orders, and the Gmgn tape to judge when it entered the high-risk zone.

Zero Is the Destination

An altcoin that doesn’t go to zero violates crypto’s entropy. Don’t fantasize about holding anything until you “become a believer.” You are not its dad, and you are not its CEO. A project’s terminal purpose is to complete one wealth transfer: a few people get off, the many fill their seats. Your job is to be in the group that gets off — not the missionary telling the story.

Airdrops, celebrities, VCs, whitelists — none of these are reasons not to run. Hit your pre-set psychological profit target, and leave. If it does another 10x afterward, that has nothing to do with you. You didn’t lose by missing the full move. You won by still having the right to bet next round.

LUNA: $119 to zero. Entropy always collects.

Sell Into the Noise

When your memecoin group, your X feed, your Telegram, and even the dinner table are all talking about the same coin, that isn’t a bull market. It’s a signal: time to run.

The market loves harvesting emotion at the climax, because nobody is sober and the FOMO meter is pinned. You’ll hear “it’s doubling again,” “XX hasn’t pumped yet,” “CEX listing any day now” — but remember: the genuinely smart money started leaving before those sentences were ever spoken.

If you truly can’t stand being flat, then cut down to a size that lets you sleep, keep a small observation stub, and give up on eating the last bite. Behind every deafening top are countless heads queued up for the harvest.

No Revenge Trading

Selling a loser means you admitted, in that moment, that your judgment was wrong. So why would you immediately buy something new at your emotional low? Trying to “win it back” to erase the last mistake is the single worst way to open a position.

What you should do is cool off — review, observe, detach. Figure out where the error actually was: the cycle? the timing? the structure? Or did the operator genuinely play you? Only by solving the problem do you avoid making the same bet in the same emotional state.

Don’t turn speculation into an emotional outlet. Your account isn’t a casino — it’s a tool. The price of emotional trading is your principal.

Small Narrative, Small Trade

A small narrative is not a niche opportunity. Most of the time it’s just a short-term volatility trap. On-chain statistics keep proving the same point: the overwhelming majority of new coins die under a $3M market cap, and only a tiny handful ever cross $10M, let alone $100M.

Which is why intraday speculation is the correct frame here. If you’re entering a brand-new coin with no social graph and no trending slot — purely on the belief that “someone will probably buy after me” — then don’t draw the chart and don’t dream the dream. Complete the whole operation inside 30 minutes: buy, ride, sell, leave.

Stop fantasizing that these become the next SHIB. Most of them die the moment distribution ends.

Corpses Twitch

When an old project suddenly rips, that’s usually not a launch — it’s an exit. A crowd of early holders has been camped inside it, and the moment its name reappears on the trending list, they’re cracking their knuckles, ready to take your liquidity.

Unless you can confirm that genuinely new structure, players, and money have arrived — a chain migration, a new narrative, whales entering, an exchange listing — never catch the last leg of an old chart.

“You thought it was a resurrection. It was just the corpse twitching.”

Flat for days, then vertical: the anatomy of an old chart’s exit pump.

Capital Is the Game

Your most important operation isn’t catching the right wave — it’s keeping your capital. Only while the principal survives do you get to play the next hand.

At 2x you’re thinking 5x. At 5x you’re thinking 10x. When it falls back to 3x you feel like you just “lost 2x,” so you hold out of spite, ride it all the way back to 1x, and finally capitulate below your entry. That’s not a market problem. That’s a broken mental model.

Take 3x on one coin, roll the profit into another coin and take 3x again — that’s a 9x return. Insisting on white-knuckling a single coin to 9x isn’t courage; it’s irrationality.

note

Staged profits, compounded across trades — that is the retail player’s optimal strategy.

Capital retention by strategy: discipline keeps 85–92% of your principal in the game; blind aping keeps 20%.

Greedy, With Rules

The on-chain memecoin market is an anti-intellectual world that still runs on rules. Every behavior that makes money rests on the same underlying logic: rhythm recognition, structure first, on-chain behavior analysis, clear standards for adding and cutting. Simple to describe, hard to execute. The whole question is whether you can keep doing what you’re supposed to do without emotion dragging you off course.

note

No one gets it right every time. But as long as you never make the fatal mistake, you’ve already won.

Don’t try to get fat in one bite, and don’t fantasize that this particular dip makes you rich overnight. Every judgment you make is the precondition for your next trade. Memecoins are your game against the entire on-chain casino — and all the win-rate weighting, all the accumulated strategy, exists to buy you that one big piece of meat you can lick off the knife with a clear conscience. The rest, leave to time.

Credit where it’s due: the thinking in this piece was seeded by @CryptoSunova ↗.