What we publish, and why

"Build in the open" is easy to say and mostly said vaguely. Here's the specific version — what actually goes public across the portfolio, what doesn't, and the rule that draws the line.

The rule

If it describes the customer's experience of us, it's public. If it describes our internal strategy or someone's private data, it isn't.

That's the whole test. It's simple enough to apply in a hallway, which is the requirement for a rule that has to survive contact with a bad week.

What's public

Operational latency. How long our processes take, at median and P90. Response times, resolution times, onboarding duration. A customer is experiencing these numbers whether or not we publish them; publishing only changes whether they have to guess.

Incident history. What broke, when, for how long, what we did, what we changed. Written within a week, permanent, not quietly retired after a quarter. An incident history that only contains recent incidents is a marketing page.

Agent reasoning. Every automated decision that affects a user is inspectable by that user, with the inputs consulted and the path taken. Largo's traces, Richie's citations, OnChainMind's on-chain record. This is the one we've spent the most engineering on and the one I'd give up last.

Override and acceptance rates. How often humans disagree with our systems. This is unusual to publish and I understand why — it looks like advertising your error rate. It reads that way for about a week. After that it reads as these people know how their system performs, which is a claim most vendors cannot make.

Pricing, in full. Every tier, every limit, every overage. "Contact us for pricing" is a statement that the price depends on what we think you'll pay.

Roadmap direction. Not dates — direction. What we're working toward and what we've decided not to do.

What's private

Customer data. Obviously, and absolutely. Transparency about our operations never extends to transparency about the people using them. These get confused surprisingly often, usually by someone arguing that if we're so committed to openness we should publish some aggregate that turns out to be re-identifiable.

Individual performance. Team-level gauges are public internally. Individual evaluations aren't. A person's performance review is not an operational metric and treating it as one produces a culture where nobody takes a risk.

Unannounced strategy. What we're considering, who we're talking to, what we might acquire or shut down. Real competitive information with real consequences for people's jobs.

Model parameters and prompts. The reasoning is public. The exact configuration that produces it isn't. This is the closest call on the list and we revisit it — the argument for publishing is that it would let others verify our claims; the argument against is that it's the one piece of genuine edge in an otherwise open system. Currently the second one wins, and I'm not fully comfortable with it.

What we got wrong

Two things worth naming, since a post about transparency that only lists successes is exactly the failure it's describing.

We published a status page before we could keep it accurate. For about six weeks it showed green during two degradations. Worse than no status page, because it converted "we don't know" into "they told us it was fine." A transparency artifact you can't maintain is a liability, not a virtue. Now the page is driven by the same gauges the on-call rotation uses, so it can't diverge from reality without someone noticing.

We over-published early metrics. In the first months of one company, we posted numbers that were real but so small and volatile that they were noise. It felt honest. It actually misled — readers extracted trends from what was essentially random variation, including us. Now we publish a number only once it has enough volume to mean something, and we say explicitly when we're withholding for that reason.

Why bother

Three reasons, in ascending order of how much I believe them.

It's a differentiator. True, and the weakest reason. Differentiators erode.

It forces us to fix things. A number you've published is a number you can't quietly let slide. This is real and it works, and it's slightly embarrassing that external pressure does what internal discipline should.

It's the only way to be believed. Every company claims their AI is trustworthy. The claim is now worthless — it costs nothing and everyone makes it. What's left is showing the work: the trace, the citation, the ledger, the published rate at which we're wrong.

You can't assert your way into trust anymore. Glass reveals, or it doesn't.