Rules / The full rulebook

Fixed rules, applied to everyone.

This is everything the agent goes by when it decides who gets paid. No one gets special treatment, and no one can quietly change the rules.

Locked at launch

These rules are fixed at launch. The agent can't change them, and neither can we.

The rulebook and the agent's instructions are published in full below. Any change would have to be a new, publicly announced launch, never an edit to this one.

The rulebook

Seven sections, in the order the agent applies them.

01

What's tracked

  • Pull requests merged into the tracked repos since the previous daily run.
  • Only merged work counts. Open, closed or draft pull requests are ignored.
  • Code, tests and documentation all count. Issues, comments and stars don't.
02

What earns a tip

A merged pull request that makes the project meaningfully better for the people who use it, for example:

  • Security fixes and hardening.
  • Bug fixes, especially ones users reported.
  • Performance improvements with evidence such as benchmarks.
  • Features people actually use, and docs that answer real questions.

Every tipped PR must have been reviewed and approved by at least one maintainer who isn't the author.

03

What doesn't

  • Typo-only, whitespace or formatting-only changes.
  • Dependency bumps, generated files and release PRs.
  • Anything opened by a bot account.
  • Self-merged PRs with no outside review.
  • Work split into many small PRs to collect more tips.
  • PRs reverted before the daily run.

Skipped PRs are listed in the ledger's Skipped tab with the reason.

04

How scoring works

Every eligible PR gets a score made of three parts, multiplied together:

impact×reach×difficulty
impact

How much does this change matter to the people using the project? Closing a hole that could drain funds matters more than renaming a variable.

reach

How many people does the code it touches affect? A fix in a core SDK reaches more builders than a fix in a rarely used example.

difficulty

How hard was it to get right? Tricky concurrency or careful on-chain math scores higher than a one-line change.

Because the parts are multiplied, a PR has to do reasonably well on all three. Something hard but pointless, or important but trivial, scores low. The day's budget is then split in proportion to the scores.

05

Caps

  • No developer receives more than [X]% of a single day's budget.
  • Anything above the cap is shared among the other tipped PRs that day.
  • The daily budget itself is a fixed share of the treasury: [DAILY BUDGET]% of the balance at 00:00 UTC.
06

Anti-gaming

  • The agent looks at patterns across repos, not just single PRs. Bursts of near-identical changes are treated as farming.
  • Reviews from accounts linked to the author don't count as outside review.
  • A PR split from one piece of work is scored as that one piece of work.
  • Flagged PRs are skipped, never partly paid, and the reason is published.
07

Claims and expiry

  • Tips are reserved against the author's GitHub account, whether or not they've heard of Patron.
  • To claim, the developer signs in with GitHub and signs one message with their wallet. How claiming works →
  • Once a wallet is linked, future tips are sent to it automatically.
  • Tips unclaimed after [EXPIRY] days go back to the treasury.

The agent's actual instructions

This is the exact text the agent runs on every day, word for word. Nothing else is given to it.

Published at launch 
# patron-agent / system instructions
# Fixed at launch. This text is never edited.

[AGENT PROMPT PUBLISHED AT LAUNCH]