Statewave Guide 第 5 天：教會系統表達後，才發現它仍未完成任務

Built 27 Aug 2026 · written up 21 Oct 2026

Yesterday's question:

Can we turn verified product truth into human guidance without inventing anything new?

Yes.

At least linguistically.

Task-wise…

not so much. 😅

We added Guidance Compilation.

Instead of rendering the Product Model directly:

What the app does → What we should tell the user → How we should say it

Then we ran the same 21-feature review again.

Same instrument as yesterday: one model, one fixture, and the language and usefulness scores come from a separate model against a fixed rubric, not from people.

The language result was almost absurdly clear.

Too technical                 14      → 0
Natural-language quality      1.14/2  → 1.95/2
Clarity                       1.43/2  → 1.86/2
Malformed identifier soup     gone
Incorrect factual claims      still 0

And for the first time:

two features scored a full 3/3 usefulness.

And, also for the first time, one scored 0.

A feature with nothing verified to say, which produced a title and a question.

Round 2's spread had neither end.

And the capability split moved: measured again after the language fix, features with a verified capability scored 2.00 against 1.07 for those without. Yesterday's "barely a difference" was bad presentation, not a missing signal.

Nice.

Except the overall usefulness gate still failed.

1 / 3   median
33%     useful outputs

Why?

Because we'd taught the system how to speak…

but it kept stopping halfway through the task.

Rotate API key?

Open Settings.

That's it.

Delete client?

Open Client Detail.

Upgrade client?

Open Client Detail.

Export CSV?

Open Clients.

Perfectly readable English.

Almost impressively unhelpful. 😂

The metrics made the bug obvious.

missing_workflow    5     → 16
Actionability       1.24  → 1.14

And 10 of our 14 low-scoring examples had the same root cause.

Their verified workflow referred to:

feature:<id>

The compiler expected a graph node.

Couldn't classify it.

Quietly dropped the action.

Then proudly called:

Open Clients.

"actionable guidance."

Narrator:

it was not actionable guidance.

That gave us another rule:

Correct language is not the same as complete guidance.

So we're now separating:

Entry → Task action → Terminal action → Complete path

And another rule:

Every dropped verified action needs a reason.

No more silent disappearing acts.

The bottleneck keeps moving.

First prove truth.

Then prove ownership.

Then make it human.

Now make sure it actually finishes the job.

Still no chat UI.

And I'm increasingly happy about that.

Because a beautiful chatbot answering:

Open Settings.

when you ask how to rotate an API key…

is still just a very expensive tooltip.

Statewave Guide — Day 5.

Next question:

Can verified product knowledge become a complete user task — without giving the model any new authority?

This day in the repo: PR #5 · Previous: Day 4 — We passed the safety test. Then failed the product test

Statewave Guide is built on Statewave — an open-source, self-hosted memory runtime for AI agents. Governance and provenance, not just retrieval. Apache 2.0.

→ statewave.ai · github.com/smaramwbc/statewave

Every post in this series: The Statewave Guide Journey Index