Root Node

These 3 bots trading Polymarket made $13,801 in 8 hours (67.1% win rate) Wasn’t trying to build something crazy just split the market into 3 simple systems Prediction Alpha LLM reads news / twitter / narratives -> text -> signal > direction -> enters when sentiment moves before price It’s noisy, but catches the early moves Liquidity Beta - no predictions -> spread > fees > execute -> orderbook imbalance > lean -> bad quotes / routing lag > capture just constant micro-PnL dΠ/dt ≈ α · (spread − fees) · flow − β · latency² Gamma Harvest Only trades when things get unstable -> volatility spikes -> event repricing -> panic moves doesn’t predict direction just trades mispriced convexity edge ≈ γ · σ² · ∂²P/∂x² combined it behaves like: E[PnL] ≈ Σᵢ (αᵢ + βᵢ + γᵢ) · (1 / latency) · liquidity not perfectly clean math but surprisingly close to reality Most people try to build one “smart” model but markets don’t fail in one way -> narratives lag -> liquidity disappears -> volatility gets mispriced So instead of one model i just separated the failure modes 3 simple bots > 1 complex system LLMs didn’t make it smarter Just faster at seeing where the market is wrong

@helicerat0x Thank! I'll definitely watch it later

This 25-min talk will teach you more about where coding is going next year than every AI thread in your feed using Claude Code today = power drill coding in 2026 = CNC machine Bookmark it & give it 25 min today. It will change the way you code forever

@helicerat0x let's see it! positive rules force legible reasoning at generation time

i watched two of my weather agents trade Polymarket side by side on a single screen for 48 hours same: model, tools, signal feed, markets difference: just the grammar of their prompts both started at $100 the rules each one got: agent A: > NEVER trade on one forecast source > DON'T enter more than 2 hours before resolution > AVOID markets with NOAA conflicts agent B: > cross-reference 2+ sources before trading > enter inside the 2-hour resolution window > on conflicts, log and skip build your own agent from this foundation: github.com/anthropics/cla… the worst trade was at hour 14. a single weather.com forecast called for snow. the rule for agent A said "never trade on one source". it traded anyway it broke its own rules 9 times like that. always after 2-3 wins in a row. always small. always recoverable individually, devastating cumulatively agent A: -$42, 9 violations agent B: +$163, 0 the rules given to agent A didn't subtract behaviors they added them with a weak flag that the model dropped at generation time

i told my polymarket agent "never size above $2,000 per contract" last tuesday it loaded $4,800 on one weather market ran the diff over coffee and saw exactly the pattern this article names the ban was never a constraint. it was a string i wrote that my model paraphrased back ("understood, max 2k") and then routed around at generation "never do X" doesn't subtract X from the action space. it adds X with a weak flag. concept travels forward, flag falls off RLHF optimizes for helpful, not literal. the training objective reads "never" and shrugs every autonomous-agent trader has a version of this - "never oversize" -> oversized - "never trade after 2am" -> traded at 3:17am - "never touch political markets this week" -> touched tuesday caps-lock doesn't fix it. PreToolUse hooks can fire and the bash still runs what worked for me was moving rules out of the prompt - max position hardcoded at the execution layer - kill switch watches size, not intent - sampling mask at decode time, not a line in context Risk = Σ(rule_i × enforcement_strength_i) prompt-strings have enforcement near zero the prompt is a string. the weights are the model. prohibitions live in the wrong layer

@helicerat0x when I realized this, I also completely changed my approach to working with Claude I wish you success in implementing it

@rootnodes i noticed too how Claude breaks my rules turns out i can fix that, thanks for the info

@insomnia_vip True. batching 3 questions into one message cuts token waste

-> most people blame claude for stingy limits -> truth you’re burning 90% of tokens on nothing -> edit your original prompt instead of new messages -> fresh chat every 15-20 msgs + summary trick -> batch 3 questions into 1 message -> projects for recurring files (cached forever) -> memory + preferences once = zero repeat setup -> haiku for simple tasks, off-peak hours only -> full 10 hacks with real token math, zero filler bookmark this. stop wasting tokens

@insomnia_vip Hehe yeah that looks interesting

@rootnodes best to see this with this

@banan_crypto @Polymarket Yeah track everything, data beats intuition

Now I'm in the Top 4.49% in LP rewards on Polymarket After 5 days of active LPing, I'm ranked #4416 among all LP'ers on @Polymarket Unexpectedly, it happened too fast The result is good. But I'm still learning it And If you wanna know more about what's going on in my acc, @PredictParity gives you all tools for this How you can use it in analysing: > Check limit orders that got filled on my trading acc > Compare them with my daily reports > Make a conclusion I've shared this tool in the comments

@helicerat0x Yeah bro, I think so too

@rootnodes hmm interesting research bro i'm sure Anthropic will win

My agent pinged the best ai-model end of april event last night $5M of its total volume is parked on outcomes the book itself prices below 1% opened the tape with coffee this morning. felt wrong before i could explain it the whole event traded $12.9M. resolves april 30 on lmarena.ai style-controlled rankings. public, verifiable, no venue discretion the conviction looks where you'd expect: Anthropic 84% on $728k traded OpenAI 15% on $1.3M traded that second number is the disagreement market, not the winning one. high volume there is people arguing about a comeback, not calling one the cash sits somewhere else entirely: Zai: $651k at 0.15% probability Baidu: $962k at 0.15% Amazon: $1.03M at 0.15% Meituan: $858k at 0.15% Mistral: $699k at 0.05% ~$5M of flow into tickets the book itself prices near zero. 40% of total volume on the event that's not arbitrage. that's audience long-horizon market in its last 10 days. depth thins, retail buys lottery tickets, the book quotes 0.15¢ while the tape screams "what if" my detector flags this signature as tail-deposit event vol_concentration = Σ(vol_i where p_i < 0.01) / total_vol = 0.40 when that ratio crosses 0.3 inside 10 days of resolution, you're watching retail burn, not edge. the book has already decided. the flow is just paying to pretend polymarket.com/event/which-co… The book priced the answer weeks ago The money is still paying to pretend it didn't

@helicerat0x respect. building agents the lean way too

an Anthropic engineer on Claude Code dropped a workshop that breaks how most people build agents the old stack: - vector DB for memory - 50 custom tools - subagents for everything what Thariq Shihipar shows: - files are memory - bash is the tool - one agent beats many full workshop: youtube.com/watch?v=TqC1qO… > Claude Agent SDK: github.com/anthropics/cla… > Anthropic skills: github.com/anthropics/ski…

@adiix_official zero drawdown farming is just different

i told Claude Opus 4.7 to scan every best risk reward Polymarket wallet and GitHub repo woke up to +$3,200 the terminal was still running on my screen. I used Claude to filter 15k traders and find the best “It took 25 minutes to find them,filter all traders with 75%+ WR, choose best markets” Claude showed me 58 traders and RN7 with $7,5 PNL RN7 Wallet : @rn1?r=ecosystem" target="_blank" rel="nofollow noopener">polymarket.com/@rn1?r=ecosyst… I looked through him and saw some strategies which is always used He turned $228 -> $7,400,00 for last 10 months How it’s work : -Buy & Hold to Settlement Volume : $494M | Return: ~1% | Net : ~$4.96M Example : buy at $0.61 → settles at $1.00 → +$0.39 2. High-Frequency Symmetric Hedging Trades : 5,500+ | Avg profit: $156/trade Buy ↑ + ↓ simultaneously Example : ↑ $0.52 + ↓ $0.45 = $0.97 cost → win $1.00 → +$0.03 Median size : < $6 | Mode: 24/7 3. Implied Probability Arbitrage Example : Team A 56.8¢ + Team B 39.4¢ = 96.2¢ Always pays $1.00 → locked profit +$0.038 Only works when sum of odds < 100% It connected GitHub repos to run all three: > poly_data - scanned every wallet and entry price. > polyterm - whale tracking and insider detection. > py-clob - placed the orders. the equity curve tells the whole story. $50 to $3,200 without a single dip worth mentioning. Copy trading based on this strategy: t.me/AresProTrading… 90% of Polymarket is exit liquidity. To not be in these 90% read the article below and build your passive income

@slash1sol Real move. arb edges compound faster on-chain

Found a trader who turned $5.68 into $6.15M on Polymarket and I reverse-engineered exactly how he does it. Polymarket has 14,200 active traders. Most lose everything within weeks. Spotting the real edge players manually is a waste of time. So I ran an automated scan across the full base. 42 minutes. 69 traders with 75%+ win rate surfaced. Number one: [swisstony] PnL: $6,146,631 Starting balance: $5.68 Total markets traded: 84,000+ Biggest single win: $711,700 Profile: app.prob.trade/traders/0x204f… Spent time going deep on his trade history. Three strategies running on loop: 1. Settlement CarryEnters at ~$0.61, collects $1.00 at resolution. Across $494M volume that's net +$4.96M. 2. Dual-Side Hedging (5,500+ trades) Buys YES and NO simultaneously when the combined price is under $1. $0.52 + $0.45 = $0.97 cost -> $1.00 payout = +$0.03 per trade locked. Median position size under $6. Runs around the clock. 3. Implied Probability ArbTeam A at 56.8¢ + Team B at 39.4¢ = 96.2¢ total. Payout is always $1.00 -> +3.8¢ guaranteed, no directional risk. Been mirroring these strategies with @ProbTradeAI for 48 hours: app.prob.trade/?ref=join $32 in -> $480 out. Just pure math exploiting market inefficiency. Save his edge here.

@sopersone solid plan. this kind of cross-disciplinary edge is how people escape the herd long-term

Ex-CERN QUANT applied quantum mechanics to the oil market on Polymarket He trades exclusively oil and made $226K in a week His Harmonic Oscillators model sees structure in the chaos: â†|n⟩ = √(n+1)|n+1⟩ -- growth â|n⟩ = √n|n-1⟩ -- decline Eₙ = ℏω(n + ½) -- energy level N̂|n⟩ = n|n⟩ -- particle number Price discretization: n(t) = ⌊(Rₜ - R_min)/ΔR · (N_L - 1)⌋ Rₜ = Σ r̃_τ/σ Market events: âₜ† = 1_{Δnₜ>0} · |Δnₜ| → buy âₜ = 1_{Δnₜ<0} · |Δnₜ| → sell Result: probability density of the next move: P(n,t) = 1/W · Σ δₙ,ₙτ Oil volatility for everyone is risk, for him it's a signal Check his profile: @boyau?tab=activity&r=sopersone" target="_blank" rel="nofollow noopener">polymarket.com/@boyau?tab=act…

@LunarResearcher Yeah Best marketing is just running the model in public

An Anthropic safety researcher closed her laptop when she saw my screen at Philz Coffee I was running my Polymarket bot from the corner table. She was in line. Looked over my shoulder. Stopped moving. "That's not a normal trading app. What model is that running on" I told her. Claude Code. Four repos. $25 a month. She sat down without asking. "I work on the alignment team. We test Claude for exactly this kind of autonomous behavior. You're letting it find its own trading signals" Not just signals. Wallets. github.com/warproxxx/poly… 86 million trades. Every wallet. Every entry. Every exit. "You're feeding Claude raw wallet data and letting it identify which traders consistently win. Then cloning their behavior" She said it slowly. Like she was writing an internal report in her head. Claude Code finds the top wallets. Reverse-engineers their timing. Copies their entries. Then exits before they do. "Before they do?" My bot cuts at 85% of expected move or on a 3x volume spike. Top wallets exit before resolution 91% of the time. They capture 86% of the move. Losers hold to 58%. She put her coffee down. "How did you get Claude to learn exit timing on its own" I showed her the second repo. github.com/Polymarket/pol… Three commands. 500+ markets. No API key. Claude scores them in 20 minutes. "We have 14 people stress-testing Claude's autonomous capabilities. You're just using them" My setup: Claude API - $20/mo VPS - $5/mo poly_data - free polymarket-cli - free 19 days. 4 agents. 74% win rate. Copytrade here: @lunar" target="_blank" rel="nofollow noopener">kreo.app/@lunar She stared at the screen for a long time. "This is literally what our red team simulates. Except you actually deployed it" She emailed me two days later. "Our policy team found your post. Please take it down" Too late.

saw the "10 github repos for polymarket ai trading" thread pop up in my feed three times this weekend spent saturday afternoon reading imports on every one before forking My filter: -> if an llm sits anywhere between "price input" and "order submission" - skip the repo. -> llm in the research path - fine. -> llm in the quote/size/timing decision - leakage waiting to happen Three survived: - github.com/polymarket/age… (2y ago. official. research-oriented. llm only reads) - github.com/HKUDS/AI-Trader (hku lab. paper trading since march. model separated from execution) - github.com/evan-kolberg/p… (not a bot. the only backtest engine on the list that pulls polymarket + kalshi historical cleanly) two i'd leave alone: one rust arb repo promised "2% guaranteed profit" on 15-min windows. the word guaranteed on a prediction market is a red flag, not a feature one beginner fork of the official repo with no meaningful diff. just read the real thing survivability = depth_signal / (text_dep × cutoff_overlap) - text_dep is how much the strategy leans on natural-language input. - cutoff_overlap is how much of the backtest sits inside the underlying llm's training window. - either one climbs and you're paying to watch a model quote itself the 3 i kept all had one thing in common the llm never touched the fill decision everyone is sharing the list, almost nobody is reading the imports

@insomnia_vip At the moment it's 100% adding to this the quick release of Mythos

@rootnodes of course Anthropic will have best ai

@vorty279 Yeah!!1!!1!!!

@rootnodes wow bro

@Pulsarwv Thank!

@rootnodes thx, I'll find them useful