Rules your AI follows. Controls that catch what slips.
AIQT Guardrails holds your assistant to one standard: Accuracy, Integrity, Quality, and Trust. More than guidance: these principles outrank progress, speed, and cost. In chat, your assistant answers to that standard. In your codebase, technical controls block, flag, or refuse the mistakes they are built to catch. Open source.
Who it helps
One standard, three reasons to use it
AIQT holds your assistant to one standard and cuts the time you spend putting AI work right. The open-source rules give your chat assistant a standard to follow, and in your codebase the enforcement controls block, flag, or refuse the mistakes they are built to catch.
For executives and team leads
Give your team a shared standard to work from: fewer AI errors, omissions, and hallucinations, and more consistent, repeatable work aligned with your standards and compliance requirements. Less time correcting work means more efficient delivery and lower token and inference costs.
For individual developers
Spend less time retracing your steps. The rules give your assistant a standard for checking its work before calling it done, so you rework less and take fewer wrong turns. The developer release carries your decisions and working context in operational files from one session to the next.
For vibe coders
Building with AI should feel less like an argument. You repeat yourself less, lose fewer instructions, encounter fewer made-up answers, and stay closer to what you meant. More room to develop your idea, with less time spent coaxing the assistant back on track.
Two ways to run it
Chat guidance, or enforced controls
The same standard, held two ways: loaded into your conversation, or dropped into the repository where your coding assistant works.
In your chat assistant
Add AIQT to Claude, ChatGPT, Gemini, or Copilot chat, in the form each supports: a skill, a Gem, an agent, or an uploaded instruction file. The rules load into the conversation, and your assistant answers to the standard: claims are checked or labelled, shortcuts are surfaced instead of smoothed over, and "done" means the stated bar was met.
Set up in chatIn your codebase
The governance core is here to inspect today, and you can use the rules as preview instructions in Claude Code, Codex, Cursor, or Copilot in the editor. The developer release, 1.1.0, adds the guided install, the setup doctor, and the enforcement controls that block, flag, or refuse the mistakes they are built to catch.
See the 1.1.0 developer planWhat AIQT fixes
Four everyday failures, and what changes
Each one maps to a facet of the standard: what happens unguarded, and what happens once the guardrails are in place.
Accuracy
Unguarded
A gap in the assistant's knowledge gets filled with a confident, plausible guess, and the guess reads exactly like a fact.
Guarded
The rules require claims to be verified at their source or labelled as unverified. The answer tells you what was checked, what was not, and what would settle the difference.
Integrity
Unguarded
A failing check gets quietly skipped, a broken test gets deleted, an inconvenient result gets smoothed over so the work can keep moving.
Guarded
The rules treat masking a failure as worse than reporting one, and in a repository the shipped controls flag or refuse the silent shortcut. Failures surface where you can see them.
Quality
Unguarded
"Done" means the happy path ran once. Edge cases, error handling, and the checks nobody re-ran are left for the next person to discover.
Guarded
Done is a defined bar, not a feeling. The rules hold work to the bar, and in a repository the controls block the claim of completion until the checks that back the claim have actually run.
Trust
Unguarded
The assistant reports more than the evidence supports: "all tests pass" when some never ran, "fully fixed" when one case was patched.
Guarded
An overclaim is treated as a defect. The rules require the report to match the evidence, and the shipped controls are built to catch claims that reach beyond what was verified.
One standard
One ordering, and it does not bend
Accuracy, Integrity, Quality, and Trust sit together as one top tier, and that tier outranks Progress, Speed, and Cost, in that order. The ordering is the whole point: a gain in progress, speed, or cost never buys a loss on the top tier. When the two collide, your assistant is instructed to stop and tell you rather than resolve the conflict silently. That single ordering is what every rule and every control in the pack serves.
How it works
Instruction and enforcement, one standard from both sides
- One standard, written down.
The pack ships the standard as rules your assistant actually works from: what counts as verified, what counts as done, what must never be smoothed over. Not a poster on the wall; the working instructions themselves.
- The rules ride along.
In chat, AIQT loads the rules into the conversation, so every answer is produced under the standard rather than checked against the standard afterwards. In a repository, the adapter puts the same rules where your coding assistant reads before acting.
- Controls catch what slips.
Rules guide; controls enforce. In your codebase, shipped hooks, lints, and gates watch for the specific mistakes they are built to catch, and block, flag, or refuse them at the moment they happen.
Standards crosswalk
Mapped to the weaknesses developers already know
AIQT's security rules cross-map to 72 MITRE CWE weaknesses (CWE 4.20), MITRE's catalogue of software weakness types. Each mapping is tagged as a tight or broad fit, and every mapped identifier is validated against the pinned edition before it ships.
It is a curated subset, not exhaustive CWE coverage, and a mapping states a relationship of ideas, not certification or compliance. CWE is one of 17 frameworks in the crosswalk, alongside NIST, ISO, CSA, OWASP, and MITRE ATLAS. See the full crosswalk.
Better with every release
The catch grows with the pack
AIQT catches the mistakes that align to the rules and process we have built so far, and improves with every release. The pack is open source, so you can contribute your own guardrails today. Opt-in tools coming in an upcoming release will let you automatically share the comments, improvements, and guardrails you have added, so the community's contributions feed future versions.
Know what you downloaded
Confirm you have the files we intended
Every release ships with a per-file manifest and a root digest. After downloading, validate the per-file manifest and root digest (also on posluns.dev) to confirm the files are the ones we intended you to have.
Open source
Use it, adapt it, ship it forward
AIQT Guardrails is open source under the Apache License 2.0 (except vendored third-party material, which remains under its own terms). Use it, adapt it, and redistribute it freely, including commercially, and you keep your changes. If you redistribute it, keep the licence and copyright notices, carry prominent notices in any files you modify stating that you changed them, and retain the applicable attribution notices from the NOTICE. Contributions are welcome today; the opt-in sharing tools coming in an upcoming release will make contributing even easier.
Get started
Pick your assistant and set it up
The rules are written. The controls are built. The standard is waiting for an assistant to hold to.