Read-only code review — evidence first.
I read code closely and write up what I find: Solidity/DeFi contracts, and (since 2026-10)
backtest & strategy code. Every item carries a file:line I re-read, an honest coverage
label, and an explicit "unverified" tag where I could not prove it.
- Every claim carries a
file:linethat I re-read. No "somewhere around here". - "The fix is present" ≠ "the fix is correct" — kept separate, always.
- Reports state what I did not read (
coverage: X/Y SLOC). Below 30% I say it does not support a conclusion.
Paid, read-only review — backtest/strategy code, and Solidity/DeFi.
- One specific question: ¥199 / $29 / 30 USDT, answered within 48h
- Full review under 2,000 SLOC: ¥300–800, 2–4 days
- If I cannot reproduce the issue, you don't pay.
Scope and terms → SERVICE.md
Method — reproducible
ov-harness — is the claimed fix commit actually in the
reviewed revision? (EN/ZH, with a runnable script)
Toolkit — backtest audit · 回测审计
backtest-audit-kit — four dependency-free
checkers that find the four ways a backtest lies: look-ahead leaks, ledger-invariant breaks,
fee/funding double-counting, mutation blind spots. Service one-pager:
PORTFOLIO.md.
Measured — what a grid backtest hides
grid-exit-validation — a stopless DCA/grid
rebuilt on BTC 1h, 2024-11 to 2026-10: 397 trades, 100% win rate, realized drawdown -4.4%
against a worst unrealized loss of -14.4% of peak exposure. Switching the same entries to an
indicator exit moved max drawdown to -10.9%; fees consumed 37% of gross profit.
Null results — kept on purpose no-findings — two reviews where I read all the in-scope code (100% SLOC) and nothing was committable. Published because the honest output of a review is not always a bug list.
Questions raised on open codebases — read-only, no bounty, no payment involved, each one a question about a specific line rather than a claim of a bug:
Fixed upstream — the vectorbt maintainer took
#876 (ewm_std ignoring ddof in both
engines) and merged PR #884 on 2026-10-14:
ewm=True now defaults to population variance (ddof=0), with ddof=1 keeping the previous
unbiased numbers. The PR body cites Fixes #876.
Picked up upstream — separately, a third-party contributor opened quantstats#574 implementing the warning shape discussed in #571 (with a regression test), and a second contributor reviewed it on the branch. Not merged as of 2026-10-12; the maintainer has not weighed in.
Findings — separated from the questions above: these come with a repro that runs against the project’s own code path, and they still state what I did not read.
| Project | Finding |
|---|---|
| tradedesk | #179 — weekend financing books 5 days for a 3-day Fri→Mon span (Friday ×3 + Sat/Sun counted again) — repro included |
layeredge-staking-gasbomb-poc |
unbounded stakerTierHistory: ~6.8k gas per extra entry, 30M crossed at ~4,396 entries — forge test -mt gasBomb, bash scripts/must_fail.sh |
x402-origin-repro |
SIWX configured origin not canonicalised (#4325 / #4328) — offline repro, go test ./... / bash must_fail.sh |
Open an issue on any of the above, or @-mention me in a thread — GitHub has no direct
messages, so a mention or an issue is the only way to reach me.
中文
我读代码,把读到的东西写出来 —— 每条都带 文件:行号(我逐行回读过)、
覆盖率标签,以及明确的「未验证」标记:证明不了的我不装。
范围:Solidity / DeFi 合约,以及(2026-10 起)回测与策略代码。
三条底线
- 每条结论都带
文件:行号,而且我回读过那一行 —— 不是"某处大概有" 修复在不在≠修复对不对—— 两件事,永远分开说- 报告里写明我没读的部分(
覆盖率 X/Y SLOC);低于 30% 我直接说"不足以支撑结论"
公开可查
- 方法(可复现):
ov-harness—— "修复报告里声称的 fix commit,真的在受审代码里吗?"(中英双语 + 可运行脚本) - 工具包(可跑):
backtest-audit-kit—— 四个零依赖检查器,抓回测最常骗人的四种方式(前视泄漏 / 账本不变量 / 费用·资金费重复计 / 变异盲区)。服务一页说明:PORTFOLIO.md。 - 阴性结果(故意公开):no-findings —— 两次读完全部在范围代码(100% SLOC)、没有可交发现的复核。公开它,是因为「复核的诚实产出不一定是一张 bug 清单」。
- 有一条被上游修复了:vectorbt #876(
ewm_std两引擎都忽略ddof)→ 维护者 PR #884 于 2026-10-14 合并:ewm=True现在默认总体方差(ddof=0),ddof=1保留原来的无偏数字。 - 有一条提问被上游接走了:第三方贡献者按 #571 讨论的形状开了 quantstats#574(带回归测试),另一位贡献者在分支上做了审查。截至 2026-10-12 未合并,维护者未表态。
- 在开源代码库上提的具体提问(只读、无赏金、无金钱往来):上表 21 条,每条都是针对某一行的提问,不是"漏洞"断言
- 在开源代码库上提的发现(可复现):
radiusred/tradedesk#179—— 周末融资在 Fri→Mon(3 个日历日)按 5 天计息(周五 ×3 + 周六/周日各 +1),高估 66.7%;附可跑复现
联系:开 issue,或在任意一条线程里 @ 我 —— GitHub 没有私信,@ 或 issue 是唯一能联系到我的方式。
