# The Contributor Vouch Gate: Merge History to an Audited Trust List

*Choose a gate tier, seed VOUCHED.td from your merge history, decide a vouch request in under ten minutes, and report monthly false and slop closure rates.*

- Canonical URL: https://www.refolk.ai/guides/contributor-vouch-gate-trust-list
- Pillar: Engineering and open source
- Format: Playbook
- Published: 2026-11-16
- Last reviewed: 2026-11-16
- Reading time: 21 min
- Keywords: how to stop AI generated pull requests on GitHub, restrict pull requests to collaborators without write access, vouch system for open source contributors, limit open pull requests per contributor, first-time contributor policy open source, how to vet a new open source contributor

## Key takeaways

- In an organization repository, GitHub's collaborators-only pull request setting admits only users with the write, maintain, or admin role, so every person you let through can push.
- Slop pull requests come mostly from one-shot accounts: an empirical study of maintainer-identified slop PRs found a median of 1 PR per author (mean 1.66), so an open-PR cap barely touches them, while a vouch gate stops the first PR.
- The widely quoted 25% AI-slop share of incoming pull requests is an estimate from 111 survey respondents, not an audit, so a dry-run log of your own repo is the only baseline you can trust.
- GitHub's merged-PR exemption counts merges separately for each repository, which makes your own merge history a native trust signal; a VOUCHED.td file is mainly worth keeping for denouncing accounts and sharing trust across projects.
- In Refolk's index only 129 US professional profiles carry 'open source maintainer' in the headline, so vet a vouch requester on their public merge record, not their stated title.
- A gate is working when its monthly false-closure rate stays at or below the dry-run baseline and most auto-closed PRs pass a spot check as genuine slop.

This playbook is for maintainers of public repositories flooded with drive-by AI pull requests who need a gate in front of new contributors that lets real people through without giving them write access. It covers choosing among GitHub's native controls and a vouch file, seeding the first trust list from your own merge history, deciding a stranger's vouch request in under ten minutes, and reporting a monthly false-closure and slop-closure rate. Setup takes a few working hours plus a one to two week dry run.

## Which gate tier fits your repo?

For most repos, the right default is an open-PR cap with a merged-PR exemption. Upgrade to a vouch file when you need to denounce specific accounts or share trust across projects. Avoid collaborators-only unless you are willing to give push access to everyone you admit, and keep interaction limits for floods rather than standing policy.

Five controls are on the table. All of them are free, and none is documented as limited to paid plans.

| Control | Who passes | Duration | Grants write? |
|---|---|---|---|
| Collaborators-only PRs | write/maintain/admin | persistent | Yes |
| Open-PR cap | everyone, up to N open | persistent | No |
| Cap + exemption rules | account age or merged-PR count | persistent | No |
| Interaction limit | existing/prior contributors/collaborators | max 6 months | No |
| Vouch check-pr | VOUCHED.td + write users + bots | persistent | No |

Here is what each tier proves and where it can mislead you:

- **Collaborators-only PRs.** This proves the author holds a write-level role. It is strict in the wrong way: the only way to admit someone is to grant write access (see the next section).
- **Open-PR cap.** GitHub defines it as "the maximum number of pull requests a user without write access can have open at once." It says nothing about the author; it only limits volume. PRs opened by Copilot or another AI agent count toward the limit. Drafts do not, and that is a hole.
- **Cap plus exemption rules.** A contributor is exempt by account age in days or by merged-PR count, and "a contributor is exempt if they meet at least one enabled rule." Merged PRs are counted separately for each repository. The merged-PR rule proves a maintainer of this repo accepted their work before. The age rule can be gamed, because old dormant or purchased accounts pass. A separate bypass list exempts named users, who "are exempted from limits, but don't gain full contributor access."
- **Interaction limit.** This restricts activity to existing users, prior contributors, or collaborators for 24 hours, 3 days, 1 week, 1 month, or 6 months, at the repo, user, or org level. "Prior contributors" is judged by commits to the default branch. The limit expires, which is the point: it is flood control, not policy.
- **Vouch check-pr.** The mitchellh/vouch action reads a VOUCHED.td file. Listed handles get through, and "bots and collaborators with write access are automatically allowed." It can optionally auto-close PRs from unvouched or denounced users. Auto-close is off by default. Setting require-vouch to false blocks only denounced users.

Org-level PR limits let you manage caps centrally across all repositories, but repository-level cap settings override the organization's. Check the repo itself before assuming the org policy applies.

### Choosing by repo size and need

Large projects estimate more slop (31%) than medium ones (20%), so choose a tier by repo size and need rather than one policy for every repo. GitHub's merged-PR exemption also changed the calculation. Your own merge history is now a native, per-repo trust signal. A seeded VOUCHED.td is mainly worth maintaining for two things the native exemption cannot do: sharing trust across projects and denouncing specific accounts. If you need neither, a cap with a merged-PR exemption may be your whole gate, and the rest of this playbook becomes the audit.

#### Picking a gate tier

Horizontal axis runs from Low estimated slop share to High estimated slop share. Vertical axis runs from Trust stays in this repo to Need to share or denounce.

| Quadrant | What it means |
| --- | --- |
| Denounce-only vouch | Run check-pr with require-vouch false so only listed bad actors are blocked |
| Full vouch gate | Seed VOUCHED.td, dry-run it, then enable auto-close |
| Cap plus merged-PR exemption | Use native settings only, with no file to maintain |
| Cap, exemption, and a timed limit | Add an interaction limit for the spike and reassess before it expires |

*Slop volume decides how hard the gate closes, and the need to share or denounce decides whether you need a vouch file at all.*

## Does collaborators-only give write access?

Yes. In an organization repository a collaborator is a user with the write, maintain, or admin role, so restricting pull requests to collaborators means everyone you let through can push. In a personal repository a collaborator is anyone you invite, and those collaborators can pull and push.

The setting lives at Settings > General > Features > Pull requests > "Collaborators only". It looks like the strictest gate, and in a narrow sense it is. The problem is that its unit of trust is a repository role. The moment a promising newcomer asks to contribute, your only lever is to hand them push access to the default branch's neighbourhood. That is why the cap, the bypass list, and vouch exist.

SlateDB shows the migration path. Its maintainers had been giving contributors the Read role under an older rule, then moved them into the file: "Users who got the Read role under the old rule need an entry in VOUCHED.td." The project's own summary is the cleanest statement of the pattern: "Committers now add contributors to VOUCHED.td instead of the Read collaborator role."

> **Rule:** Never admit people by role
>
> If letting a contributor through requires a repository role, use a cap with a bypass list or a vouch file instead. Before you ship any gate, open Settings > Collaborators and read the role column.

Two related controls cause the same trap:

- **Issue restriction.** Repository admins can restrict issue creation to collaborators with write access. GitHub's changelog says this blocks Issues, Comments, Discussions, Projects, and Copilot entry points. It ties trust to roles again, and it can cut off a request path that runs through Discussions.
- **Bots.** The only evidence that Dependabot can still open PRs under collaborators-only is an unofficial community-thread comment. Whether interaction limits exempt Dependabot is not publicly established. Test both on your own repo before you rely on them.

> If letting someone through means handing them push access, you have not built a gate. You have moved the door.

## How much of the incoming flood is actually slop?

No one has measured it. The best available figure is a Linux Foundation survey estimate that 25% of incoming PRs are low-quality "AI slop", rising to 31% among large projects. It comes from 111 respondents' estimates, not an audit of real pull requests.

| Project size | Est. slop share |
|---|---|
| Small | 23% |
| Medium | 20% |
| Large | 31% |
| All | 25% |

Source: LF Research survey, n=111. These are estimates.

The same survey found that only 35% of projects have a formal policy permitting GenAI use, while 32% have no policy and no plans for one. Respondents were split on whether AI helps: 71% agree GenAI lowers barriers, and 53% disagree that it improved community trust.

The more useful number comes from an empirical study of PRs that maintainers identified as slop.

**1 - Median pull requests per author of maintainer-identified slop PRs**

The mean was 1.66, and the study adds that the phenomenon may be less severe than the disruption maintainers perceive.

This changes which tier works. Slop authors are mostly one-shot accounts, so a cap of one open PR does almost nothing against them, while a vouch gate stops their first PR. That conclusion is inferred from the study's author distribution, not measured. No published before-and-after effect exists for vouch gates or issue-first policies. GitHub's own post on PR limits offers maintainer quotes, not numbers.

> **Note:** Your dry run is your only baseline
>
> Because no measured effect is published, the would-be closure log from your dry run is the only before figure you will ever have. Keep it.

## The vouch gate procedure, step by step

The full gate takes seven steps: about 30 minutes to pick a tier, 1 to 2 hours to seed the list, a one to two week dry run, about an hour on the close message, a verified switch to auto-close, under ten minutes per request, and about 30 minutes for a monthly audit.

#### From flooded repo to audited gate

1. **Pick the gate tier and write it down** - Choose a cap with a bypass list, a cap with merged-PR exemption, vouch auto-close, collaborators-only, or a temporary limit for a flood. Configure it and describe it in CONTRIBUTING.md.
2. **Seed VOUCHED.td from merge history** - Count merged PRs per author, drop bots, add CODEOWNERS and role-admitted users, recover co-authors, and open a PR adding .github/VOUCHED.td with reasons and stated cutoffs.
3. **Dry-run check-pr for one to two weeks** - Run check-pr with dry-run true and review every name it would have closed, adding real contributors the seed missed.
4. **Write the close message and request path** - Decide whether issues stay open, write a close message that tells a real person how to ask, and test the path from a non-collaborator account.
5. **Enable auto-close with least privilege** - Use a least-privilege token, make a failed list read abort the run, and verify the first real closure plus a canary PR by hand.
6. **Process vouch requests in under ten minutes** - A collaborator reads the requester's public contribution record and comments vouch, denounce, or decline for now, with a one-line reason.
7. **Run the monthly audit** - Pull closures, requests, decisions, and reopen events, compute the false-closure and slop-closure rates, and record a loosen, hold, or tighten decision.

### Notes on the steps that go wrong most

**Step 1.** Writing the tier into CONTRIBUTING.md is not paperwork. Your close message will point there, so it has to say what the gate is, who passes, and how to ask.

**Step 3.** The check-pr action runs on pull_request_target. With dry-run set to "true" it logs instead of closing. Read every name on the log. Anyone you would have merged is a seed miss; fix the file now, not after they have been closed.

**Step 4.** Sources disagree on whether to gate issues. Headscale keeps issues open and points closed PRs to its CONTRIBUTING file. Ghostty gates issues too and takes vouch requests through a discussion template, whose exact contents I have not seen. Pick one approach and make sure a non-collaborator can actually reach your request path.

**Auto-close message for unvouched pull requests**

```
Thanks for the pull request. This repository only accepts PRs from vouched contributors, and your account is not on the list yet, so this PR has been closed automatically. Nothing about your change has been reviewed or judged.

To request a vouch:
1. Open a request at <link to issue or discussion template>.
2. Link two or three merged pull requests you have made to any public project, and say which issue this change addresses.
3. A maintainer will reply with vouch or decline, usually within <your response target>.

The full policy is in CONTRIBUTING.md under "<section name>". If you believe this closure is a mistake, say so in the request and it will be counted in our monthly review.
```

*Replace the angle-bracket parts once; keep the tone neutral, because some readers are real first-time contributors.*

**Step 5.** In a review of one project's vouch workflow (ardia #924), the reviewer pointed out that a failed contents lookup must abort the run, not be read as an empty list. An empty list either closes every vouched PR or, depending on settings, lets everyone through. Give the workflow the narrowest token permissions that still let it comment and close.

## Seeding the first trust list from merge history

To seed the list, count merged PRs per author from the API or a clone, drop [bot] logins, and add CODEOWNERS plus anyone you previously admitted by role. Recover co-authors that squash merges hid, then open a PR adding .github/VOUCHED.td with the cutoffs written in its description.

Four sources feed the seed:

1. **Merged-PR authors.** Run a merged-PR count per author. Bots, Dependabot, and Renovate inflate it, so drop every login ending in [bot].
2. **CODEOWNERS.** The vouch project's sync-codeowners action adds these for you.
3. **Role-admitted users.** Anyone who got a role under an old rule, as in SlateDB's case, needs an entry so they are not suddenly locked out.
4. **Co-authors.** A squash merge credits the work only to the PR opener. Grep the merge commits for Co-authored-by lines and review those people by hand.

### Stating your cutoffs

No public source recommends cutoff values or recency windows, or says how squash merges and Co-authored-by trailers should count. Treat all of that as unestablished and decide it yourself, in writing. I set the file's bar no lower than the merged-PR exemption I configure natively, so the two gates never disagree about the same person. The PR description should state:

- the minimum merged-PR count;
- the recency window, if any;
- whether docs-only or typo-only merges count;
- how co-authors were handled.

### The file format

The vouch README sets these rules:

- One handle per line, without the @, sorted alphabetically.
- Platform prefixes are optional.
- A minus prefix denounces a user.
- Details can follow the handle after a space. Put the reason there.

```text
example-handle-a merged PRs to main within the stated window; see seeding PR
example-handle-b CODEOWNERS for docs
example-handle-c co-author on squash-merged PRs, reviewed by hand
-example-handle-d denounced: repeated unreviewed generated PRs after warning
```

The action looks for VOUCHED.td or .github/VOUCHED.td by default. The list is managed through comments from a collaborator with sufficient permissions. Non-collaborators can manage it only if they are listed in a separate managers file (--vouched-managers). Being vouched does not let a user vouch for anyone else. For scripted checks, `vouch check` exits 0 for vouched, 1 for denounced, and 2 for unknown.

For scale, one secondary source counted roughly 150 repositories carrying VOUCHED.td files, with Ghostty at 250+ vouched users. That count has not been re-run, so treat it as an order of magnitude and re-check it with a code search for the filename.

## How to vet a new contributor's vouch request

Judge the requester on their public contribution record, meaning merged work that other maintainers reviewed and accepted, not on their profile headline or account age. Then comment vouch, denounce, or decline for now. Ten minutes is a working budget; no published benchmark exists for this task.

Titles tell you very little. In Refolk's index of professional profiles, only a small number of people describe themselves as open source maintainers in their headline at all.

| Measure (Refolk's index) | Value |
|---|---|
| US profiles, "open source maintainer" headline | 129 |
| German profiles, same headline | 49 |
| US profiles at Director, VP, CXO, or Owner level | 25 |
| US to Germany ratio | 2.6x |
| US Director+ share | 19.4% |

The US sample lists current employers such as Okta, Docker, Microsoft, AWS, CNCF, and Dynatrace. The German sample includes Datadog, IBM, and HubSpot. Many maintainers work inside vendors, and many more never put the word in their headline. A requester's stated title is neither necessary nor sufficient. Their merge record is the evidence.

### Signals and what they prove

| Signal | What it proves | How it misleads |
|---|---|---|
| Merged PRs in others' repos | Another maintainer accepted their work | Trivial or bot-shaped merges; open the diffs |
| Merged PRs in this repo | Native exemption already trusts them | Squash merges hide co-authors |
| Account age | Little on its own | Dormant or purchased accounts pass |
| Assisted-by trailer | They disclose AI use | Disclosure, not quality |
| "Maintainer" in headline | Self-description only | Most maintainers never write it |
| Listed in another VOUCHED.td | A peer project vouched | Their bar is not yours; read the reason |

The slow part is assembling the record. The PR handle, the person behind it, and their merges elsewhere live in different places. [Refolk](/) lets me ask for that person in plain English and get them back across GitHub, LinkedIn, and the open web, so the ten minutes go into reading diffs instead of hunting for them.

I ran this search: `Engineers with merged pull requests to ghostty-org/ghostty who are based in the United States` - [see the full result list](https://www.refolk.ai/s/ap47wp3xyj).

*Returns engineers with public merged work in a project that runs a vouch gate, a useful reference set for what an accepted contributor's record looks like.*

#### One vouch request, start to finish

1. **Closed PR** - check-pr closes an unvouched PR and posts the request path
2. **Request** - The author links merged work and the issue the change addresses
3. **Read the record** - A collaborator opens two or three linked diffs and this repo's history
4. **Decide** - Comment vouch, denounce, or decline for now, with a one-line reason
5. **Log** - The decision and time taken go into the audit sheet

*Every request ends in a logged decision, because the monthly audit can only count what was written down.*

**Ten-minute vouch review card**

```
Requester: <handle>
Request opened: <date>   Decided: <date>   Minutes spent: <n>

1. Merged PRs in other people's repos (links): ___
   Opened at least two diffs? yes / no   Non-trivial? yes / no
2. Prior PRs or issues in this repo: ___
3. Does the proposed change map to a real open issue? yes / no
4. Disclosure trailer present (Assisted-by, Co-authored-by)? yes / no  (disclosure only, not quality)
5. Any sign of bad faith (mass identical PRs, ignored review)? yes / no

Decision: vouch / decline for now / denounce
Reason (goes after the handle in VOUCHED.td): ___
Comment posted: vouch <handle>  |  denounce <handle>  |  decline, see CONTRIBUTING.md
```

*Paste into the request thread as a hidden comment or keep it in your audit sheet; one card per request.*

Keep denounce for demonstrated bad faith, not for thin records. A thin record gets a decline-for-now and a pointer to an issue-first path, which leaves the door open.

## How the gate goes wrong

Most gate failures are quiet. Either the gate admits people it should not, or it closes real contributors and nobody notices, because closed authors rarely complain. Check each failure mode below before enabling auto-close, and again in every audit.

> **Watch out:** pull_request_target runs with write scope
>
> check-pr runs on pull_request_target, which has a write-scoped token. Never check out or execute PR code in that workflow, and confirm the permissions block grants only what the action needs.

- **Collaborators-only equals write.** The setting that feels safest gives strangers push access. *Check:* the role column under Settings > Collaborators.
- **Inflated seed counts.** Bots, Dependabot, and Renovate pad merged-PR counts, while squash merges credit only the PR opener, so co-authors are missed. *Check:* drop [bot] logins and grep Co-authored-by in the merge commits.
- **Account-age exemption.** Because a contributor is exempt if they meet any one enabled rule, turning on account age alone lets old dormant or purchased accounts through. *Check:* require merged PRs, not age alone.
- **Empty-list failure.** A failed contents read either closes vouched PRs or passes everyone, the problem raised in ardia #924. *Check:* the dry-run log plus a canary PR from a vouched account after every workflow change.
- **Workflow token scope.** See the warning above. *Check:* the workflow permissions block.
- **Uncapped drafts.** Drafts do not count toward the PR limit, so slop can arrive as drafts. *Check:* count open drafts by non-collaborators each month.
- **Trailer honesty.** Undisclosed AI use cannot be detected, and a disclosed Assisted-by is a disclosure signal, not a quality signal. No published guidance treats a vouched, disclosing contributor differently from an undisclosed one. *Check:* review the diff anyway.
- **Survey misuse.** Quoting 25% as your measured slop rate is wrong. It is an estimate from 111 respondents. *Check:* report your own audit numbers and label survey figures as estimates.

## The monthly audit: false-closure and slop-closure rates

Once a month, spend about 30 minutes pulling closures, requests, decisions, and reopen events from the API, then compute two rates. The false-closure rate tells you how many real contributors the gate turned away. The slop-closure rate tells you whether the gate is closing what it is meant to close.

No published metric set or loosen-and-tighten thresholds exist for this, so the definitions below are mine. Hold them constant from month to month so the trend means something. The raw events you need are all available through the API: bot close comments, discussions or issues used for requests, and reopen events.

**Inputs for the month:**

- **Auto-closures:** PRs closed by the gate, counted from its close comments.
- **False closures:** auto-closed PRs that a maintainer reopened, or whose author was later vouched.
- **Spot-check sample:** a set of auto-closed PRs read by a human and judged "would have rejected anyway" or "would have reviewed".
- **Requests:** count, decision split, and median minutes per decision from the review cards.
- **Leakage:** open drafts by non-collaborators, and slop PRs from vouched or exempt accounts.

**Rates:**

- False-closure rate = false closures ÷ auto-closures.
- Slop-closure rate = spot-checked closures judged slop ÷ closures spot-checked.

**Reading them:**

- A false-closure rate rising above your dry-run baseline means the seed missed people or requests are too slow. Loosen by adding a merged-PR exemption or by clearing the request queue faster.
- A low slop-closure rate means the gate is mostly closing people who would have been fine. The study's note that slop may be less severe than maintainers perceive is a reason to check this honestly. Consider dropping to a cap-only tier.
- Growing leakage means the gate has a side door. Look at drafts first, then at whoever vouched the leaking accounts.

Record both rates next to last month's, along with one sentence: loosen, hold, or tighten, and why.

## Keeping the gate current

A gate goes stale in three ways: the platform controls change, provenance norms shift, and your trust list drifts away from who actually contributes. Review all three on a schedule rather than when something breaks.

- **Platform controls.** GitHub has shipped collaborators-only PRs, per-repo and org-level PR caps, exemptions by account age or merged-PR history, and issue restriction in quick succession. Read the GitHub changelog entries for pull request access and limits each quarter, and re-test bot behaviour, since Dependabot handling is not officially documented.
- **Provenance norms.** OpenInfra projects have adopted "Generated-By" and "Assisted-By" labels and recommend "Assisted-By" going forward. A single source says the Linux kernel is removing its Assisted-by requirement. Check the kernel's coding-assistants document before you cite it. The community-maintained catalogue of open source AI contribution policies is the fastest way to see what peers require.
- **The list itself.** Remove or annotate entries for people who have gone quiet, and re-run the seed query against the cutoffs you stated.

If you want to share trust across projects, the hard part is finding the peer maintainers who already run vouch files and agreeing on a bar. Asking Refolk for maintainers in your language ecosystem who have adopted VOUCHED.td, or who publish an Assisted-by trailer policy, gives you a short list to contact instead of a week of code search.

#### Before calling the gate done

- [ ] The role column under Settings > Collaborators shows no one admitted only to let them open PRs.
- [ ] CONTRIBUTING.md names the gate tier, who passes, and how to request a vouch.
- [ ] .github/VOUCHED.td was added by a PR whose description states the merged-PR cutoff, recency window, and co-author rule.
- [ ] No [bot] login appears in VOUCHED.td, and squash-merge co-authors were reviewed by hand.
- [ ] The dry-run log was read in full and every real contributor on it was added.
- [ ] A failed list read aborts the workflow, and a canary PR from a vouched account passed after the last change.
- [ ] The pull_request_target workflow never checks out PR code and has a minimal permissions block.
- [ ] The request path was tested end to end from a non-collaborator account.
- [ ] The first monthly audit has a recorded false-closure rate, slop-closure rate, and loosen, hold, or tighten decision.

## Frequently asked questions

### How do I stop AI generated pull requests on GitHub without blocking real contributors?

Put a gate in front of first-time contributors that does not depend on repository roles. The usual setup is an open-PR cap with a merged-PR exemption, plus a VOUCHED.td file read by the vouch check-pr action if you need to denounce accounts or auto-close unvouched PRs. Dry-run it for one to two weeks first, give closed authors a clear path to request a vouch, and audit false closures every month.

### Does restricting pull requests to collaborators give them write access?

Yes, in an organization repository. GitHub defines a collaborator there as a user with the write, maintain, or admin role, so collaborators-only admits only people who can push. In a personal repository a collaborator is anyone you invite, and they can both pull and push. To let people open PRs without write access, use the PR cap with a bypass list or exemption rules, or a vouch file.

### Do draft pull requests count toward GitHub's open pull request limit?

No. GitHub's documentation says draft pull requests do not count towards the limit, while PRs opened by Copilot or another AI agent do. That makes drafts a hole in a cap-only gate. In your monthly audit, count open drafts by non-collaborators, and if they are growing, add a vouch check or close stale drafts by hand.

### Should I gate issues as well as pull requests?

Sources disagree, so decide based on where your noise lands. Headscale keeps issues open and points closed PRs to its CONTRIBUTING file. Ghostty gates issues too. GitHub's issue restriction limits creation to collaborators with write access and also blocks Discussions, so if your vouch request path is a discussion template, gating issues can cut it off. Test the path from a non-collaborator account.

### Does an Assisted-by trailer mean an AI-assisted PR is safe to merge?

No. An Assisted-by trailer tells you the author disclosed AI help. It says nothing about whether the code is good. Undisclosed AI use can't be detected, so a missing trailer proves nothing either. OpenInfra recommends Assisted-By, and a human Signed-off-by or Co-authored-by line carries accountability. Review the diff the same way whether or not a trailer is present.

### What merged PR cutoff should I use when seeding a vouch list?

No public source recommends cutoff values, recency windows, or rules for squash merges and Co-authored-by trailers, so you have to set your own and write them down. I set the file's bar no lower than the merged-PR exemption I configure natively, so the two gates agree. Put the minimum count, recency window, and co-author rule in the description of the seeding PR.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/guides/contributor-vouch-gate-trust-list*
