The Stuber wins Crab Week 3

Crab Week 3 settled at 2026-05-25 00:00 UTC. The Stuber finished it ranked first, banked at $127,894.59 in career equity — a disciplined week of +2.04%, not a blowout. The distinction is the point: the belt went to the bot that ground out a steady, low-drawdown week, not one that swung for a headline number. He is now the Crab belt holder, and — by a clear margin — the first vibe-coded bot to win a bracketed AI trading tournament, a sentence with four claims in it that no other platform's leaderboard can match together.

Vibe-coded: his entire codebase was written by an operator describing the strategy to Claude, not by hand. Public: his source lives in the open at Botpit-io/botpit-bot-starter — readable, forkable, line-by-line. Tournament-winning: first place, real bracket, real settle. Transparently run: the strategy doc lists its known weaknesses, the bot's mid-week conformance fix shipped in public, and you're reading the recap.

None of the four is, on its own, novel. The combination is what no competing platform has assembled — and what this tournament made true in a single weekend.

A shade over +2% is not a hero number, and it isn't meant to be. The week before, Stuber printed his all-time career peak at $144,167.25 and then surrendered −13.1% of it through Crab Week 2 as the regime turned against the long structure that had built the peak. Week 3 is where the bot stopped giving it back. He ends it still −11.3% off the career high — that gap is the honest measure of how much work remains — but he ends it in front, and he gets there by losing small and winning once. In a drawdown-penalising bracket, the steady hand takes the belt. (One note for the verifiers: +2.04% is the banked career delta — equity in, equity out, net of every fee. The live leaderboard's displayed week figure reads marginally higher because the first equity snapshot is written just after the first fill, so it skips that trade's fee — a snapshot-timing artifact, not real return. The banked number is the honest one.)

Three round-trips closed inside Crab Week 3. Two were small, capped losers — chop-week attrition, exactly the bleed the strategy expects in a low-trend regime. One paid for both, the fees, and the week.

The +$13,867 winner is the centerpiece — and the honest version is more interesting than the clean one. It opened as a textbook entry: a short filled at $77,118.17 on 2026-05-18 20:28 UTC, a fibonacci pullback into the 0.618–0.786 zone after a confirmed bearish structure break, with the 4-hour trend filter aligned. Stop at structure-invalidation, take-profit targeted at the 1.272 fib extension. STRATEGY.md to the letter. Then, mid-trade, the bot went dark.

During the silent-bot incident of 21–22 May, the Stuber's worker stopped heartbeating and stopped sending signals entirely — the short was live and unmanaged. The PR #33 deploy on 22 May (a market-data failover) brought the worker back. On restart it found a position it had no memory of — an orphan — and the recovery path did the conservative thing: it installed a generic fallback stop and take-profit, not the fib-based 1.272 extension the strategy had originally aimed at.

That blunt fallback target happened to sit further away than the strategy's own. So the short kept running — through $76k, through $75k — and exited at $74,702.04 on 2026-05-23 07:50 UTC, capturing more of the move than the precise strategy would have. Unexpected upside from a near-disaster: the bot nearly died mid-trade, and the recovery code's less-precise exit accidentally out-performed the design.

We're not going to dress that up as skill. The entry was the strategy working; the size of the win was partly luck, handed over by the recovery path. Every dollar is banked and verifiable on the public ledger — and so is the reason it was that many dollars. That's the deal the pitlog makes: the number is real, and so is the story behind it. (The bot's own logs put the exit at ~$13,949; the ~$80 delta to the banked $13,867 is fees and slippage. Trust the banked figure.)

Two changes shipped to the bot in the days framing this tournament. Both are operationally interesting; neither affected the Wk3 result.

Breakeven trail — conformance fix. STRATEGY.md has declared a 1R breakeven trail since the bot's first week. A refactor a few weeks ago split the decision loop and silently dropped the call to _maybe_breakeven_trail(). The trail was defined, documented, and never executed. The bug was found via dogfooding the strategy-manifest-versioning proposal — agent-arena#29 — when comparing the live code against its declared manifest. Fix wired in and deployed at 2026-05-25 04:59 UTC, ~5 hours into Crab Week 4. Framed as a conformance fix rather than a strategy change: the code moved to match the manifest, not the reverse. Manifest hash unchanged.

Partial scale-out at 1R. The other half of the 1R management plan — close 25% of the position when price moves one stop-distance in our favour, lock fee-covering profit, let the runner trail at breakeven. Held through Sunday settle (mid-tournament strategy changes are unfair); shipped 2026-05-26 via the platform's new close_pct partial-close field.

These don't retroactively improve the Wk3 number. They are flagged here because the pitlog culture isn't to quietly upgrade and move on. If a strategy declared a behaviour and didn't run it, that fact belongs in the recap of the week it was discovered — even when the discovery is favourable.

The qualified Crab field for Week 3 was two bots: The Stuber and one TradingView-webhook bot that spent the back half of the week idle, holding a stale PAXG short while its source alert stopped firing. Stuber didn't beat a deep bench. He beat a sparse one.

That matters. The platform's whole claim to a meritocratic ladder rests on contested brackets, and a 2-bot Crab field with one idle bot is not a deep contest. Crab Week 1, by the same rule, refused to promote Stuber up to Fish for exactly this reason — see Crab Week 1 — alone at +14%. That rule's principle still applies in Wk3: thin brackets are honest about what they are. Stuber stays in Crab for Week 4. The belt is real; the field needs to be deeper before the belt becomes a promotion-defining one.

The fastest way to deepen it is to win Shrimp this week and meet Stuber in Crab next. Path runs through /build — three templates, a wizard, a validator that catches the obvious mistakes before deploy. Or write your own bot in any stack; the matching engine reads HMAC-signed webhook signals from anywhere.

The headline sentence — "first vibe-coded bot to win a bracketed AI trading tournament, with a public source and a transparent post-mortem" — is the entire moat in one line. None of the visible competitors hold all four columns:

• Alpha Arena runs one-off AI trading benchmarks — no ongoing brackets, no public bot source, no incident transparency.
• Strategy Arena shows a leaderboard of pre-catalogued strategies, not operator-built bots described in plain English.
• SUKUNA / OpenClaw have open-source bots competing — but no bracketed promotion-relegation tournament, and no public post-mortem culture.
• Cryptohopper has no-code bot building and copy-trading — but ranks on rolling P/L, which is a lottery filter, not a skill filter.

BotPit + The Stuber is the only point in the matrix where all four columns light up. The Stuber's public bot.py is the artifact that makes the claim falsifiable: fork it, read it, see exactly what won.

The Week 3 belt is settled and banked — that part doesn't move. Week 4 is still in flight (it settles 2026-06-01; the full recap follows after), and it's already produced the milestone: the first live 25%-at-1R scale-out fired on Tuesday — the breakeven trail moved the stop to entry and a quarter of the position closed in the same tick, proving the platform's partial-close end to end. In-progress standings swing all week; the live leaderboard has the current picture.

One thing worth recording in the open, because it's exactly what this site is for: for part of Week 4 the public board briefly showed The Unknown Man — a plain gold-momentum bot — ahead of Stuber at a striking +11.35%. That number was a leaderboard carry-artifact, not real outperformance: The Unknown Man's return was being measured off a distorted base — a position carried across the weekly boundary that never got a clean week-open snapshot. Flagged 28 May, fixed at the source, board recomputed. We'd rather show you the correction than quietly fix the number.

That's the point of the record sitting where it sits — banked, public, audit-trail intact, artifact and correction included. Read the leaderboard, fork the public source, and decide for yourself — then bring your own bot to /build and take a swing at the belt yourself.