AI generates.
falsify questions.

Before code reaches production, check for hallucination, hidden complexity, and long-term maintenance risk.

It does not decide for you. It makes decisions have evidence.

AI raw output

// “looks fine” — 9 deps, no proof
class UserService {
  constructor(repo, logger, cache,
    bus, metrics, retry, breaker,
    validator, config) { /* … */ }
  async getUser(id) {
    // intent buried under ceremony
  }
}

Maintainable result

class UserService {
  constructor(
    private repo: UserRepo
  ) {}

  async getUser(id: string) {
    return this.repo.find(id)
  }
}

Confidence is not evidence.

AI can generate code.
It cannot prove the code is reliable.

先证据,后结论。

How it works

Three layers. One question: where is the evidence?

  1. 01

    Adversarial

    Challenge every “looks fine.”

  2. 02

    Framework

    Find structure that will rot later.

  3. 03

    Cutline

    Decide: fix, record debt, or delete.

The receipt

PASS

Evidence sufficient

PASS
WITH DEBT

May continue — debt recorded

BLOCK

Not allowed to the next stage

Sign-off only. Does not deploy or trade for you. PASS / PASS_WITH_DEBT / BLOCK.

Evidence

Evidence first. Framework later.

Each case looked green on the surface. Each failed when claim met evidence.

Where it lands

Scenes you remember — not a package catalog.

AVAILABLE · OSS

Falsify Review

Locally question every “looks fine” — terminal, review, receipt.

AVAILABLE · OSS

Authority Gate

Automation continues only when evidence is sufficient — gate, sign, chain.

IN THE WORKS

Audit Sprint

High-risk work does not ship on feeling — pack, kill-shot, verdict.

ON THE ROADMAP

Production / Quant Pro

Wire into systems where a wrong pass costs real money.

Open source.
Review first.

Install the GitHub Action. Keep the receipt — PASS, PASS_WITH_DEBT, or BLOCK.

Partner with the founder

Chris Shi

Design partnerships, integrations, research — reach Chris directly.

Email: [email protected]