open source, by Shalini Sah

Tools for systems that keep their promises.

Your README says "never." Your privacy page says "not retained." These tools find every promise like that and show you what actually keeps it.

the idea underneath

Every promise sits on a rung of the enforcement ladder.

The higher the rung, the less the promise depends on people remembering a rule. Most "private by design" and "safe AI" claims live on rung 1. Usually nobody is lying. Writing the sentence is cheap, and moving it up the ladder is real engineering work.

6Cryptography or hardwareMath or silicon. Even the operator can't break it, and clients can check.
5ArchitectureThe system has no path to break it.
4MonitoringSomeone finds out fast when it breaks.
3TestA check fails before the break ships.
2ProcessA review or approval is supposed to catch it.
1PolicyA sentence in a doc, a contract or a prompt.
↑ harder to breakcheaper to claim ↓
what you can run today

The tools.

agent skill · Apache 2.0

promise-check

Point it at a design doc, RFC, PR description, privacy whitepaper or agent system prompt. It finds every promise and tells you which rung enforces it, which are policy-only, and the next rung to reach.

Run promise-check on docs/design.md
Get promise-check
GitHub Action · stdlib Python

promise-lint

Promises drift. Someone refactors away the refund check and the README still says "never." promise-lint ties each promise to the file that enforces it and fails CI when the guard disappears.

- uses: shalinisah-dev/promise-check@v0
Add it to CI
open toolkit

Staff+ Behavioral Interview Toolkit

Prompts, evidence patterns and a story-bank structure for turning messy engineering notes into precise Staff+ interview stories, without invented facts.

Use the toolkit
does it work?

Checked blind, misses published.

The eval set has two synthetic systems with planted traps, plus Apple Private Cloud Compute and Signal's sealed sender, checked against their public write-ups. In a blind run the auditing session never saw the answer keys.

Read the results and the misses

18/18promises found
15/18verdicts right
16/18rungs right
what's next

Did the promise actually hold?

promise-check answers "what does this system promise, and what enforces it?" The next tool, replay, re-checks an agent's claimed wins against an evaluation the agent can't edit, starting with autonomous research loops. Issues and pull requests are welcome; the most useful ones are real public documents where you think the tool gets a rung wrong.