A forensic record of building, breaking, and measuring a solo crypto trading bot project, April–June 2026. Documents how three classes of measurement bugs inflated paper PnL by ~135×, why paper EV was anti-predictive of live EV (ρ = -0.43), and what survives the strategy failure: methodology, infrastructure decisions, and engineering discipline.
The project ended in a small loss. The methodology is the deliverable.
| Format | Length | Time | What it is |
|---|---|---|---|
| Reflective summary | 2 pages | ~5 min | TLDR with the headline findings, lessons, and project-at-a-glance metrics. Start here. |
| Technical autopsy | 15 pages | ~25 min | Full forensic record: methodology, infrastructure, three bug classes, anti-prediction finding, limitations. |
| Forensic source documents | 6 docs | varies | The original in-project working notes the autopsy distills from. Working artifacts, not polished. |
250+ hours of focused work.
2,409 wallets classified.
180 paper bots deployed.
417,008 paper trades simulated.
$157,115 in apparent paper PnL.
Three classes of measurement bugs invalidated 99% of it.
$300 of real capital deployed.
105 live trades.
-$96 net.
One bot, +$2.93 across 37 fills.
The fleet's apparent paper PnL of +$157,115 was inflated ~135× by three classes of measurement bugs; clean paper PnL was +$1,162. Live deployment produced -$96 across 105 fills. Across the 6 deployed architectures, the Spearman rank correlation between paper EV and live EV was -0.43 — paper EV was inversely predictive of live performance. The bots my paper trading said were best performed the worst live.
polymarket-autopsy/
├── autopsy.pdf ← Full 15-page technical autopsy
├── autopsy.md ← Markdown source for the autopsy
├── reflective_2page.pdf ← 2-page TLDR summary
├── reflective_2page.md ← Markdown source for the TLDR
├── figures/ ← 5 figures (timeline, workflow, infrastructure, paper-vs-live scatter, PnL inflation chart)
├── scripts/ ← Figure generation scripts (reproducible from data)
└── docs/forensic/ ← Curated forensic working documents from the private research repo
├── proxy_pnl_bug_day16.md
├── phantom_mechanism_forensic.md
├── day17_session_summary.md
├── execution_pipeline_setup.md
├── edge_surface_findings_day24.md
└── heatmap_3d_findings_corrected.md
A 45-day project that ended in loss still produced durable engineering artifacts that I'm carrying forward:
- A 3-layer LLM classification pipeline for behavioral wallet analysis (Haiku → Sonnet → Opus by confidence, designed as cost architecture)
- A dual-Claude workflow that turned solo engineering into something closer to a small team's output, with explicit role separation (Planner / Critic / Executor)
- A bug-class forensic methodology for diagnosing silent failures in distributed systems
- Honest measurement discipline — including the discovery that paper EV can be anti-predictive of live EV when paper data is contaminated
The full reasoning lives in the technical autopsy.
The two data figures (fig1_paper_vs_live.png, fig2_pnl_inflation_over_time.png) are generated by matplotlib scripts that query the project's SQLite databases (fleet.db, live_state.db). The three diagrams (timeline, workflow, infrastructure) are also programmatic. All figure-generation scripts are in scripts/. Source data is not included in this public repo (it lives in the private polymarket-vault working environment) — available on request for serious technical reviews. See the autopsy's reproducibility appendix for details on data sources.
- polymarket-live — Live execution environment with the bot configs, trader variants, kill switches, and operational scripts that ran the May 2026 deployment.
- polymarket-vault — Research and infrastructure working repo. Permanently private. Access available on request for serious technical reviews.
The files in docs/forensic/ are working documents extracted from the private research repo. They were written as in-project notes, not for publication. VPS IPs, personal wallet addresses, and internal paths have been redacted; otherwise the content is unchanged. They preserve the analytical state of the project at the moments they were written, which is sometimes before the bug that contaminated those very findings was discovered. The edge_surface_findings_day24.md document in particular carries a contamination caveat — it pre-dates the Phantom Vector 3 discovery by three days.
These documents are supplementary evidence for the autopsy, not standalone claims. They show the work; the autopsy distills it.
Built solo over 45 days (April 21 – June 4, 2026) on $1,500 of starting capital, as the two months before college graduation. The project produced no detectable trading edge, but the engineering work it forced — measurement bug debugging, infrastructure under geo-blocking constraints, LLM-augmented solo workflow — is what I'm carrying into whatever role comes next.
For questions or context, the easiest entry point is the 2-page reflective summary.