StructPR StructPR / Shield

// private triage for open-source maintainers

Fifty pull requests.
Five worth reading.

AI made pull requests free to produce. Your review time still isn't. Shield scores every incoming PR before you spend a minute on it — privately, with evidence you can check in five seconds, and a running record of whether its calls match yours.

Apply as a design partner Read the methodology

Free for verified public OSS projects. Four design-partner slots this quarter.

monday morning · 9 open PRs since friday → 2 need your eyes. The rest are dispatched.

Likely slop · 5

#4812

Update readme

Generic title No tests
#4815

fix typo(s)

Account 6 days old
#4818

update

Empty body AI template
#4820

Improve code quality

Diff/desc mismatch
#4821

refactor utils

First-time No tests

Vouching required · 2

#4813

Rewrite renderer scheduling

1,900 LOC No plan
#4819

Bump CI toolchain

Touches CI config New account

Cleared for review · 2

#4814

fix: race in job dedup (fixes #4790)

Tests changed Linked issue
#4817

feat: expose cache stats endpoint

Has plan Known contributor

Verdicts are private to you. Contributors see nothing.

The bottleneck moved. Tooling didn't.

Every review tool — CodeRabbit, Copilot, your CI — runs after you've decided a PR deserves attention. That decision is now the expensive part: maintainers of large projects face hundreds of open PRs, an increasing share machine-generated at near-zero cost to the author.

Shield sits before that decision. It predicts which PRs will consume disproportionate maintainer attention — not "was this written by AI," which the research shows is a losing game — using structural signals that hold up across generators: diff footprint relative to your repo's norms, contributor history, effort markers like plans, tests, and linked issues.

anatomy of a verdict

Likely slop confidence 85%
AI tool listed as Co-authored-by Title is generic No test files changed Account 12 days old

Every verdict ships its evidence. No black box — you can check the reasoning in five seconds, and honest AI-assisted disclosure is never held against a contributor.

Read-only until it earns more.

A wrong "slop" call on a real contributor's first PR is a community-burning event. So v1 is deliberately incapable of causing one — and stays that way until the numbers say otherwise.

No comments on PRs. Shield never posts. Anything you send is written by you, from your account, in your words.

No labels, no auto-close. Verdicts are advisory and visible only to installation admins.

No AI-detection theater. Disclosure trailers carry zero verdict weight. We predict review cost, not authorship.

No silent scoring. Every resolved PR is recorded against your actual merge/close decision. You see the agreement rate; so do we.

the gate — before any public surface ships, per repo:

AUC ≥ 0.80 precision@20% ≥ 0.50 false-positive rate ≤ 2% on merged PRs

Derivations, signals, and citations in the published methodology — built on peer-reviewed review-effort research (MSR 2026, arXiv:2601.00753).

Four weeks. Your call at the end.

  1. 01

    Install & enable

    GitHub App on your repo, flip Shield on. Your existing open-PR backlog is triaged within the hour.

  2. 02

    Triage privately

    Verdicts appear on your private desk as PRs arrive. Nothing changes for contributors.

  3. 03

    Watch the record

    Every merge or close you make is scored against Shield's verdict. Agreement rate accumulates in the open, on your dashboard.

  4. 04

    Decide

    Numbers hold up → we talk about labels and deeper automation. They don't → you uninstall, and we learned something.

Your attention is the scarce resource.

Free for verified public OSS projects — funded by sponsors, not maintainers. No contract, no credit card, no lock-in.

Four fields. No follow-up sequence, no newsletter — just a human reply.