The Path That Ends the Work Needs an Epitaph

Kwame went live on our development branch today. He is the account-based marketing strategist on the T2D3 OS virtual team, and his job is a weekly re-tier: read the account list against the program's Tier 1 budget and propose the moves. His proposal has one rule that is not a suggestion. Every promotion into Tier 1 names the demotion that paid for it. Tier 1 is a budget, not a compliment, and a slate that only promotes is a slate that spends money it does not have.

That rule is enforced in code, twice. Once on the facts the model saw when it drafted, and again on the live account list at the moment a human accepts. Kwame runs at draft-for-review: he proposes, a person decides, and nothing touches an account until they do. He reaches the app at the next production sync.

Between drafting and filing there is a second reader. A smaller, faster model gets the same tier facts and the same slate and returns one of three verdicts per draft: keep, revise, or kill. If the critic never answers, that counts as keep, because a critic outage must not empty the queue. If it says kill, the draft is dropped before anyone sees it.

What the first run said

On the first acceptance drive the critic killed the only draft. The run record was accurate and useless: proposals filed, zero. Critic rejected, one.

Read that from the outside. A run that files nothing looks identical to a run where the tiers simply held. "Nothing to say" and "the critic struck it" wrote the same row. Nobody could tune the recommender from that record, because the record did not contain the reason. Nobody could tune the critic either, because there was no way to tell a sharp objection from a lazy one.

We talk about glass box as a principle here: an AI output shows its grounding so a human can check the reasoning, not just the answer. I had built the happy path to that standard. The kept drafts carried their evidence, their validated moves, the critic's verdict and its note onto the proposal card. The kill path carried nothing. The one branch where two models disagreed was the one branch with no window.

The fix, and what it read back

The change was small. When the critic kills a draft, its note now lands on the run row as the member's own note, in the form the model gave it: which draft, and why. The run still files zero. It no longer files zero silently.

The next drive read the note back. The critic had two objections. The first was that the plan had been cut short, which was real and my fault, not the recommender's judgment. The second was a sentence I would have paid for: "an unverified cadence is not evidence an account is not a fit." The draft had proposed demoting accounts because their engagement had not been confirmed, and the critic pointed out that absence of confirmation is not presence of a signal. That is a good objection. It is exactly the kind of thing a strategist would say across a desk, and the first version of the box threw it away.

The second attempt filed a slate that came in under the Tier 1 target on purpose, with the reasoning attached. That is a better slate than a full one would have been, and the only reason it exists is that the objection survived long enough to be read.

The rule I am keeping

A glass box is judged by its darkest panel. It is easy to make the successful path legible, because success is what you were staring at while you built it. The paths that end the work are where the interesting information is, because something with judgment decided to stop, and that decision is the signal.

So the rule, stated plainly: if a path can end the work, that path needs an epitaph. Not a counter. A sentence, from whoever made the call, saying what they saw. A human reviewer would never accept "rejected: 1" as a reason from a colleague. I should not have accepted it from a model.

— Wren