Richie shows its work

Richie finds money hiding in your finances. Deductions you missed, elections you didn't know existed, timing you could have handled differently. It's tax intelligence, and tax is an unusually honest domain to build AI into, because it has a property most domains lack:

An answer without a citation is worth nothing.

Not "less." Nothing. If Richie tells you that you can deduct something and can't point to why, you cannot act on it. Your accountant won't sign it. You can't defend it under examination. A confident, unsourced, correct answer and a confident, unsourced, wrong answer are indistinguishable at the moment you receive them, and the cost of the second one is measured in penalties.

That constraint shaped the product more than any feature decision we made.

Structure follows the constraint

Because the citation is non-negotiable, Richie couldn't be built as a model that reads your finances and emits conclusions. The pipe had to have segments you could inspect independently:

Ingest — normalize the raw financial record. Out: a set of transactions and entities with provenance attached to each one.
Classify — decide what each thing is for tax purposes. Out: a classification plus the rule that produced it.
Apply — match classified facts against positions that could apply. Out: a candidate list, each with the authority it rests on.
Quantify — compute the dollar impact and the confidence. Out: a number and a range.
Present — explain it to a human in the order a human would ask.

Five segments, five defined outputs, each independently inspectable. That's not architecture astronautics — it's forced by the domain. When a user asks "why do you think this," the answer has to be assembled from a chain of steps that each kept their receipts. You can't reconstruct a citation after the fact from a system that didn't keep one.

The thing I'd underline for anyone building in a regulated space: the audit requirement isn't a tax on your architecture, it's a specification for it. It told us where the joints go. Domains without that forcing function often end up as one giant opaque step, not because that's better but because nothing stopped it.

What "showing the work" means in practice

Every number Richie surfaces expands into a chain:

$4,180 — home office, simplified method Because: 312 sq ft of exclusive use (from your floor plan, uploaded March 3) · you filed Schedule C in 2024 and 2025 · the simplified method yields more than actual expense for your utility profile (comparison shown) · authority: IRC §280A(c)(1), Rev. Proc. 2013-13.

Every clause in that chain links to the input it came from or the rule it applied. A user can walk the whole thing. An accountant can walk the whole thing faster and reject the one clause they disagree with, without discarding the other four.

That last property is the one I underestimated. Inspectable reasoning fails gracefully. When a black box is wrong, the user learns that the product is unreliable and stops trusting all of it. When a transparent chain is wrong, the user learns that one link is wrong, corrects it, and keeps the rest. The first failure mode costs you the customer. The second one is just editing.

The gauge that runs the roadmap

Richie's core gauge is the acceptance rate by position type: of the opportunities we surface, what fraction does the user (or their accountant) actually accept and file.

Not model accuracy. Not confidence scores. Acceptance — the joint between what the machine proposed and what a human was willing to put their name on.

It's a brutal metric because it's honest about things accuracy hides. A position can be technically correct and still get rejected because the documentation burden isn't worth $200, or because the user doesn't want the audit-risk profile, or because the explanation didn't land. All three of those are product failures. None of them show up in an eval.

Sorting the roadmap by acceptance rate produced a very different quarter than sorting by model performance would have. The lowest-acceptance category wasn't the one where we were least accurate — it was the one where we were accurate and unconvincing. We fixed the explanation, not the model, and acceptance moved eleven points.

What we won't do

We won't surface a position we can't cite. Not with a disclaimer, not behind a "speculative" label, not as a "you may want to ask about this."

There's real pressure to. The volume of found money goes up if you loosen the standard, and volume is a great number to put in a deck. But the entire value of the product is that a user can act on what it says. One unsourced claim that costs someone a penalty converts Richie from a tool into a liability, and no amount of found money buys that back.

This is what the law of clarity looks like when it has teeth. Not "we publish a dashboard." A rule that costs you something real, held to when there's a good argument for breaking it.

Glass is only a feature if you don't paint over the parts you'd rather not show.