Lookahead Lint: process audits for LLM-written trading backtests

The narrow lint for code that looks at the future is already filling with free and small tools. Since mid-2026, at least five MCP servers or PyPI packages (leakguard, StratVerity, AlphaAssay, Waldean, ere and others) flag future shifts, centered windows, leakage and deflated Sharpe. Selling a static look-ahead linter alone to a fund of $100M or more would be competing against free tools. They are all retail-grade or single-developer projects, though, and none does the parts that make this opportunity distinctive. Nothing I found compares the natural-language request with the rule actually implemented. Nothing checks features against real vendor publication or knowledge timestamps inside a fund's own data stack. Nothing automatically captures agent repair loops and the candidate strategies tried before the chosen one. The overfitting tools all depend on a trial count the user reports, and MQL5 and other commentators name undisclosed trials as the usual way they fail. Capturing that count from the agent harness itself is the defensible wedge. The allocator-facing development record looks empty. I found no product that packages request, repair, candidate and data-timestamp history for operational due diligence by seeders or funds of funds, and no evidence that allocators ask for this yet. The best entry is probably capture-first instruments: hooks into Claude Code, Codex or Cursor plus the fund's backtest engine that log every attempt with tamper evidence. The lint would sit on top as the daily use for heads of research, and the due-diligence export would be the paid feature that pulls in allocators. The main risks are demand on the allocator side, which nobody has proven, and platforms like QuantConnect or agent vendors adding audit logging themselves.

• leakguard-mcp

  Static analyzer for time-series ML and quant code, packaged as an MCP server. It has about 10 rules for future shifts, centered rolling windows and global-fit normalization. It matches patterns in the code and does not check data publication timestamps, intent against implementation, or agent retries.

• StratVerity MCP Server

  Hosted static diagnostics for look-ahead bias, repainting and overfitting, aimed at people using agents to write strategies. This is the closest to the lint idea, but it looks retail-oriented and records nothing about the development process.

• AlphaAssay

  Commercial hosted reality check for backtest results, trade lists and equity curves. It covers deflated Sharpe, out-of-sample testing and a look-ahead signature. It works on outputs, not on code or how the strategy was developed.

• backtest-reality-check

  Open-source MCP server that applies deflated Sharpe, minimum track record length, multiple-testing correction and fill realism. It relies on the user to report the trial count honestly, which is the gap a retry log would close.

• canli-validation-mcp

  Free, MIT-licensed MCP server with deflated Sharpe, CSCV probability of backtest overfitting and White's Reality Check. It installs into Claude Code with a single command. Statistics only, with no provenance.

• Waldean

  Static linter for Python notebooks and scripts that catches ML leakage, such as fitting before the train/test split or shuffling time series in cross-validation. It is generic ML, not specific to trading or agents.

• ere

  Polars library for point-in-time joins that tracks a reference date and a knowledge date per row, with an audit() function for look-ahead leakage. It covers the data-timestamp part of the lint idea as a library rather than a CI check.

• tsauditor

  Pip-installable auditor for time series with as-of rules such as earnings release lag. Small open-source project.

• peek

  Small library for auditing time-series ML pipelines for look-ahead bias. The repo has a single commit, so it is very early.

• Finterion QuantVault

  Groups backtests into bundles and compares them across engines and windows. It shows pooled summary metrics rather than a per-run record, and it is not framed around agent provenance or allocator due diligence.

• QuantConnect

  Large backtesting platform with AI assistance that tracks research ideas through evidence gates. It could add a development-history export, but today it does not produce an allocator-facing record of agent repairs and candidates.

• Look-Ahead-Freedom as Temporal Non-Interference (arXiv 2607.04958)

  Academic work that frames look-ahead freedom as a verifiable property of agentic trading pipelines. It argues that current detectors certify nothing when they stay silent. This is not a product, but it validates the problem and is a possible basis for one.

Checked 2026-10-20