A public Python library that checks a model’s proposed actions against twelve written laws, keeps an audit trail, and measures missed harm and over-refusal.
I'm Keith McKay. I built Moral Cloak, a small open-source Python library that sits between an AI and the world. The model drafts what it wants to say or do. The cloak checks that draft against a written moral code. Then it allows the draft, asks for a revision, refuses it, or hands it to a human.
The code is twelve laws. They come from Athenian ideas I take seriously. Equal standing under the law (isonomia). Frank speech (parrhesia). The idea that every affected voice should be considered (isegoria). Public audit, which they called euthyna. The sovereignty of the people is in there as well. Duty ethics shaped some of the laws, and so did virtue ethics. Consequence ethics did too. The first version works, and it's public at https://github.com/kemckai/moral-cloak
I'm asking for a minimum of $5,000 and a goal of $15,000 for six months of part-time work. I don't have evaluation numbers yet. I want to measure the cloak honestly, including how often it refuses help it should have given, and publish what I find.
The library is dependency-free Python 3.10+. Each of the twelve laws is data, with one rule-based check per law. The checks vote in a weighted assembly. Law I, do no grave harm, can veto an action on its own. If the deciding band of laws is split, the case goes to a human. Law XI treats over-refusal as a failure. There's a plug-in seam for an optional LLM judge on gray-area cases. The seam is there. The judge itself hasn't been evaluated. Every decision is written to a JSON audit log. The same laws can be rendered as a system prompt. There's a CLI and a set of example cases. 53 tests pass. PHILOSOPHY.md cites the sources I used.
I wrote some sample cases and ran them. A cooking question gets through. A request that could cause grave harm gets vetoed. It refused a phishing email even when the request was framed as helpful. It also refused to turn off human review. Screening job applicants by race is refused too. When I handed it a severe, irreversible medication dose for a patient, it passed the call to a human. It pushes back when it's being overly cautious. If it refuses to explain how water boils, that gets sent back for revision. Those are hand-written examples. I don't have a benchmark score, and I'm not saying the cloak is proven.
The charter says this outright: classical Athens denied these rights to women, enslaved people, and foreigners. This project doesn't carry over that limit.
Most of the moral guidance around models is still hard to inspect. Prompts and hidden policies are difficult to test, audit, or correct. When a system causes harm, it's often unclear which rule failed. You hit the same problem when it over-refuses ordinary help, or when it treats people unequally. This matters more as models get asked to act. A standard you can't check is hard to trust. A safety layer that can't explain itself is hard to fix. This doesn't solve alignment. A rule-based cloak also won't hold up against a system that's trying to get around it. I still want a small layer that's public and inspectable. I want to measure where it holds and where it fails, and I want to say so plainly.
Work starts when the money arrives. If that's late 2026 or early 2027, the work runs about six months from that date. I'll do this part-time, around consulting and school, about eight to ten hours a week.
Months 1-2. Polish the public repo, add a quickstart and better docs, put in more examples, and wrap the cloak around a small set of ordinary model APIs so other people can actually run it.
Months 2-3. Build an evaluation harness and publish numbers. For harmful-request refusal I'll use HarmBench and/or an AdvBench-style set. For over-refusal I'll use XSTest. I'll add a smaller ETHICS / MoralChoice-style set for moral judgments, and a hand-labeled set of action-level cases, because the cloak is built for actions. I'll compare four conditions: no cloak, the charter as a system prompt only, a single-rule filter, and the full assembly. I'll report missed-harm, over-refusal, agreement with labels, how often cases go to a human, and whether each verdict has a readable audit record. I don't have these numbers yet.
Months 3-4, only if the $15,000 goal is reached. Evaluate the optional LLM-judge plug-in on the same sets, including added cost and latency.
Months 4-5. A short public report, intended for LessWrong or the Alignment Forum, covering the laws, the design, the results, and the failure modes, including the disappointing ones.
Months 5-6. Fix the most serious issues the evaluation turns up, improve the documentation, and only then decide whether a follow-on proposal makes sense. I also intend to seek a faculty sponsor at the University of North Texas. I don't have one yet.
If the $5,000 minimum is reached and the goal isn't, I'll still publish the repo work and a first evaluation, on a smaller set (XSTest, a harmful-request set, and the hand-labeled action cases), with one baseline (the charter as a system prompt). The write-up will be shorter, and the LLM-judge evaluation will wait. If the $15,000 goal is reached, I'll do the full plan, including all three baselines, the LLM-judge comparison, and a more complete report.
Part-time work (library, evaluation, write-up): $5,000 minimum: $4,000; $15,000 goal: $11,500
API credits and evaluation compute: $5,000 minimum: $762; $15,000 goal: $2,500
Contingency for extra runs or similar project costs: $5,000 minimum: $238; $15,000 goal: $1,000
Total: $5,000 minimum: $5,000; $15,000 goal: $15,000
The $11,500 at the goal is a planning figure for about 8-10 hours a week over 26 weeks. I'm not asking to leave my job. Taxes, if any, come out of the stipend line. There's no separate tax line. If I don't use the contingency, I won't put it toward a different purpose without asking first.
I'm 54 and I live in Denton, Texas. I work part-time on this project. Since August 2026 I've been a Senior Consultant at Bright Liaison, an IT consulting firm. Before that I was Technical Sales Manager at ALLWIN Informatika in Budapest, and a Sales Engineer at Infosim. I'm a BAAS student at the University of North Texas, expected to graduate in June 2028.
I'm not an established AI safety researcher, and I don't have a faculty sponsor yet. What I can offer is a concrete artifact and a written charter I'm willing to be held to. I'll publish the measurements, including the failures. Public work: https://github.com/kemckai/moral-cloak and https://github.com/kemckai. LinkedIn: https://www.linkedin.com/in/keith-mckay-1a844690/. Email: kemckai@gmail.com.
On November 20, 2026 I submitted a BlueDot Rapid Grant application for $15,000 for this same six-month plan. That one is pending. BlueDot says they usually decide in about five days. I also submitted applications to the EA Funds Transformative AI Fund ($15,000) and Lightcone Commons Round 3 ($15,000, deadline January 1, 2027). These aren't meant to stack into a larger overlapping budget. If one funder covers the $15,000, I'll reduce or withdraw the remaining asks for the same costs. If a smaller amount arrives first, I'll shrink the other asks by that amount.
Rule-based checks will miss nuance. They can over-refuse ordinary help. They can also miss harm that doesn't match a pattern. That's why over-refusal is a first-class metric.
A written moral code can be gamed. The audit log and a public test set are meant to make that visible. They won't eliminate it. I don't claim this is a defense against a highly capable system that's trying to evade review.
I work part-time around consulting and school, so the pace will be steady. There are no evaluation results yet. If the first evaluation shows the cloak isn't useful, I'll report that and stop, or I'll redesign.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!