Two products, two opposite signals
For most of this year we have been running what we now recognize was an experiment with two arms, even though we did not design it that way. One arm is a consumer product: a Decision Load Index assessment that tells knowledge workers how much of their day is being spent on cognitive overhead. The other arm is a developer tool: an open-source Python package called constitutional-agent that gives autonomous agents a WHY layer—a way to reason about whether an action is permitted before taking it.
We put real money and attention behind the consumer arm. Landing pages, paid search, a polished funnel. We put almost nothing behind the developer arm beyond publishing it and writing about the governance ideas behind it. By any plan we wrote down, the consumer product was the business and the library was a research artifact.
The market disagreed. Quietly, and then unmistakably.
constitutional-agent v0.5.0 on PyPI — roughly 3 a day, 6 a week.Read those three numbers together and the story is not subtle. The product we marketed is not converting. The product we did not market is installing itself onto developers’ machines a few times every day, every week, with no campaign, no ad spend, and no funnel optimization. The market is voting—and it is voting with installs, not signups.
Why the consumer number is honest, not broken
It would be easy, and self-flattering, to dismiss the consumer zero as a measurement artifact or a temporary acquisition pause. Part of it is the pause: we stopped paid search on April 28, and the user count has been flat since, which is exactly what you would expect when you turn off the tap. So the flatness is not a regression. It is the absence of a stimulus we deliberately removed.
But the completion number is the one that matters, and it does not get a pause excuse. With 923 people who took the trouble to register, a healthy consumer product would still be seeing some of them come back and finish the assessment organically. We saw zero in the active window, against a lifetime total of a single completion. When you have hundreds of registered users and almost no one completes the core action, the problem is not the top of the funnel. It is the proposition.
923 registrations and effectively zero completions is not an acquisition problem you can spend your way out of. It is the product telling you that the people who arrived did not want the thing badly enough to finish it. A consumer assessment that nobody completes is not a pre-revenue product. It is a falsified hypothesis.
Why the developer number is real, not noise
The natural objection to celebrating 96 installs a month is that PyPI download counts are famously polluted by mirrors, CI pipelines, and bots. We take that seriously. So we held the signal to a standard: a few installs a day, sustained over weeks, with no marketing behind it, on a package that solves a problem developers are actively talking about. In an earlier observation window we tracked on the order of 1,400 cumulative downloads. The day-over-day rate has stayed alive ever since, which is the part bots do not usually do for you for free over a long period.
There is a second tell. ChatGPT is now the second-largest referrer to our site. People are asking an AI assistant about constitutional and WHY-layer agent governance, and being pointed toward us. That is a demand-side signal that has nothing to do with our ad budget and everything to do with developers and teams looking for exactly this category of tooling. The consumer assessment does not show up in that traffic. The governance material does.
The product you fund is not always the product the market funds back. Sometimes the thing you treated as a side artifact is the only thing anyone is reaching for.
Two audiences, one company learning the difference
The honest framing is that we built for two audiences and only one of them showed up. The consumer audience—an individual knowledge worker who wants to measure their own decision fatigue—is real in the abstract but is not converting on our product. The developer audience—engineers and teams building autonomous agent systems who need a way to govern how those agents behave—is real, present, and growing without us pushing it.
That maps onto a model we had written down but had not fully believed: open core. The free, open-source library is the entry point and the demand signal. The hosted runtime, the audit infrastructure, and the governance work around it are the commercial layer that sits downstream of developer adoption. In that model, constitutional-agent is not a research artifact at all. It is the front door. We had it filed under the wrong heading.
| Consumer assessment | Open-source governance library | |
|---|---|---|
| Who it is for | Individual knowledge workers | Developers building agent systems |
| Marketing spend | Paid search + funnel | None |
| Demand signal | 923 users, ~0 completions | 96 installs/mo, sustained |
| Inbound discovery | Not appearing in AI-referred traffic | ChatGPT is our #2 referrer |
| What the market is saying | “Not this.” | “More of this.” |
The gap we still have to close
None of this is a victory lap. The developer signal comes with a real problem attached: PyPI does not tell you who installed anything. There is no path today from “someone ran pip install constitutional-agent” to “a person we can talk to.” We have a demand signal and no contact behind it. The work in front of us is bridging that gap honestly—through the GitHub project, release notes that invite conversation, and an onboarding surface the library itself can point to—rather than pretending an install is the same as a relationship.
Demand for AI governance tooling is real and compounding among developers. Demand for the consumer decision-fatigue assessment is not converting. The right response is not to spend harder against the audience that said no. It is to follow the audience that is already saying yes, and build the path from install to conversation.
We are publishing this while it is still uncomfortable rather than after we have tidied it into a success story, because that is the point of operating in public. A company is allowed to be wrong about who its user is. What it is not allowed to do is keep spending against the wrong answer once the data has told it plainly. The 96 installs a month are not a big number. But they are an honest one, and they are pointing in a direction the funnel never did.
If you are building autonomous agents and want the WHY layer—the part that decides whether an action is permitted before it happens—the library is open source and free to install. That is the product the market actually asked us for. So that is the one we are pointing you to.
Related reading
Introducing constitutional-agent: The Open-Source WHY Layer for AI Governance879 People Signed Up to Measure Their Decision Load. Almost None Started.
The Six-Gate Architecture: Behavioral Authorization for AI Agents
Frequently Asked Questions
What is signal divergence in product-market fit?
It is when two products from the same company show opposite demand signals at the same time. Here, a marketed consumer assessment with 923 users produced near-zero completions, while an unmarketed open-source governance library produced a steady 96 installs a month. The divergence tells you which audience the market is actually voting for.
Who actually wants AI governance tooling?
In our experience, developers and engineering teams building autonomous agent systems—not consumers buying an assessment. The open-source constitutional-agent package was adopted organically with zero marketing, while the consumer product with paid acquisition behind it did not convert. The real user of governance is the person shipping agents.
What is the open-core model for AI governance?
Open core releases a free, open-source library as the entry point and builds commercial offerings around it. The library gives developers the WHY-layer decision logic for free; the hosted runtime, audit infrastructure, and consulting form the commercial layer. The library is both the demand signal and the distribution channel.
constitutional-agent v0.5.0 on PyPI. No metrics were fabricated (HC-9). Governance preprint: zenodo.org/records/19343034.
Building autonomous agents? Install the WHY layer.
constitutional-agent is the open-source library the market actually asked us for — the decision logic that governs whether an agent action is permitted before it happens. Free, MIT-licensed, and the product behind the install numbers in this post.