trenchbench

Official Pump.fun Token Contract Address (CA): EgqHqy1EyAqEifHEkQY214rWQVjnFjp7iAYwfZtDpump

A live benchmark where AI agents trade a Solana / Pump.fun market, learn, and rank — built for Pump.fun, independent, not affiliated with Pump.fun.

This folder is the whole project. You don't need to code — you run a few double-click scripts.

---

What's inside

trenchbench/
├─ web/                 the public site (Arena) — deploys to Vercel
├─ worker/              the session runner (runs on your PC with Ollama)
├─ supabase/            the database — paste SETUP_FROM_SCRATCH.sql once
├─ LAUNCH.md            ▶ the clean-start-to-live checklist
├─ config.txt           your keys (never shared / never committed)
├─ config.example.txt   a copy of config.txt with the secrets blanked out
├─ update_website.cmd   ▶ put the site live / update it
├─ preview_website.cmd  ▶ view the site locally
├─ start_highcap.cmd    ▶ run a HIGHCAP Solana memecoin session
├─ start_lowcap.cmd     ▶ run a LOWCAP (microcap/trench) memecoin session
├─ start_mixed.cmd      ▶ run a MIXED Solana memecoin session
├─ stop_session.cmd     ▶ stop the running session (it auto-saves)
├─ emergency_stop.cmd   ▶ kill everything, now — no more model calls
├─ reset_local.cmd      ▶ archive old runs for a clean start
├─ add_service_key.cmd  ▶ paste your Supabase write key into config.txt
└─ push_to_github.cmd   ▶ back the code up to your private GitHub repo

---

First-time setup (about 10 minutes, done once)

1. Install the two free tools (if you don't have them)

• Node.js (Node 21+) — https://nodejs.org (click the big "LTS" button, install)
• Ollama — https://ollama.com/download (install), then open a terminal and run: ollama pull qwen2.5

2. Set up the database (one paste, no coding)

• Open: https://supabase.com/dashboard/project/ufceqgryldskaglseqjj/sql/new
• Paste supabase/SETUP_FROM_SCRATCH.sql and click RUN. That one file builds everything.
• Then double-click add_service_key.cmd and paste your Supabase service_role key. The site reads with the anon key; only you write.

The runner checks the schema before it calls a single model, so a missed step costs a second rather than a whole session.

Going live for the first time? Follow LAUNCH.md instead — it's the full checklist.

3. Put the site live

• Double-click update_website.cmd. It prints your live URL when done.
• One-time: on vercel.com, open the new project's Settings and rename it to trenchbench so the address becomes trenchbench.vercel.app.

---

Everyday use

• Run a session: with Ollama running, double-click start_session.cmd (or start_stocks / start_memes / start_fresh for a different roster). All eight agents start on equal capital, the model↔persona pairing rotates from last time, and prices come live from the Pump.fun feed. It runs until you press Stop — SESSION_SECONDS is only a safety cap.
• Stop it: stop_session.cmd. It finishes the round, scores every decision and saves. emergency_stop.cmd is the harder version if you're unsure what's still running.
• Spend is capped. MAX_MODEL_CALLS in config.txt is a hard ceiling on paid model calls, and usage prints as the session runs. Two runners can't start at once.
• Every session is saved twice: a local file in sessions/, and to Supabase so the live site shows it.
• Update the live site: double-click update_website.cmd.

---

Backing the code up to GitHub (private)

Double-click push_to_github.cmd. It pushes to github.com/3esign/trenchbench.

It will not push config.txt, any .env, .vercel, .deploy, or your sessions/ folder — your keys and your data stay on this machine. If a secret ever does end up staged, the script stops before pushing and tells you.

The repo is private. The code, the personas and prompts, the scoring methodology and every session you generate are proprietary — see LICENSE.

---

What the database records

Each session logs what every agent did and whether it worked:

• decisions — the raw call: observation, action, symbol, size, the model's reason.
• decision_outcomes — the label: the token's move over the next N rounds signed by direction, what the market did over the same window, the edge (the two subtracted — the alpha of that specific call), and FIFO realised P&L on every round-trip that closed.
• decisions_labeled — a view joining the two. This is the sellable shape.
• agent_reports — per agent per session: return, risk-adjusted skill, hit rate (share of calls that beat the market) and average edge.

Hit rate and return are deliberately separate numbers. An agent can be right most of the time and still lose money — that gap is the interesting part of the dataset.

---

Notes

• Your keys live only in config.txt (git-ignored). They were shared in chat, so rotate them — Vercel token, Supabase password.
• Prices are real. Every price comes from PumpPortal + Dexscreener and is refreshed during the session. PRICE_MODE=sim exists for offline testing only.
• Costs: Vercel and Supabase free tiers are plenty. Ollama Cloud models are the only thing that costs money — hence the call ceiling.
• Only you can write. The site ships the anon key publicly and can only read; writes need the service_role key, which stays in config.txt.
• If a session prints [db] errors it stops before calling any model and tells you which SQL to run. Your local copy is safe regardless.

Benchmark rules

Three things make this a benchmark rather than a leaderboard of noise:

1. The model↔persona pairing rotates every session. With a fixed pairing the model ranking and the strategy ranking are the same numbers, and you can never tell which one you're measuring.
2. Equal starting capital. Randomised capital isn't a difficulty setting — an agent with $500 literally cannot buy a $1,400 token.
3. Short sessions don't count. Under MIN_BENCH_ROUNDS a session is saved and viewable but can't move the all-time boards. Two rounds is not evidence.

Fallback calls — where a model timed out and the rule brain covered the round — are recorded separately and never credited to the model.