Doc system test
Page 2 of 6 · Concept map

The code has 79 concepts, and the old docs explained only 16 of them correctly

16 covered, 5 partly, 28 wrongly and 30 never.

What we counted

A concept is one idea that a reader must understand on its own

The last three pages covered the big mechanisms. But a reader meets many smaller ideas too. What is a Cyan? What does cyanprint try do? What does the tool exit with when it fails?

We call each of these a concept. A concept is one idea that needs its own explanation. Each one gets its own doc. It can be a thing, such as the state file. It can be a step, such as dependency flattening. It can also be a command, a file format or a rule.

This page lists all of them, not just the famous ones.

How the map was built

We read the code, then swept every named item into 79 concepts

The map comes from the code, not from the old docs. The core of the command-line tool was read by hand. Three helpers then swept the rest: the old docs, the server side, and every other command-line item.

The helpers found 194 named items Measured. Each one went into one of 79 concept docs. A small item, such as one flag, sits inside the doc of the concept it belongs to.

A later, wider sweep of the servers listed 1,431 rows Measured. It counted every config key, Helm value and error type. 620 of those rows belong to CyanPrint v4. The rest fall under the registry and coordinator docs.

Then each concept was checked against silicon. Nothing was run: each mark comes from reading the old page next to the code. There are four marks:

  • CoveredSilicon explains it, and the explanation matches the code.
  • PartlySilicon has it, but leaves out a part a reader needs.
  • Covered wronglySilicon explains it, but says something the code does not do.
  • Never coveredNo silicon page explains it at all.

"Covered wrongly" is the worst mark. A missing page sends a reader to the code. A wrong page sends them off with a false picture.

The score

The old docs explained 16 of the 79 concepts correctly

Here is how all 79 concepts scored against silicon:

16 of 79
concepts silicon explains correctly
Measured
5 of 79
concepts silicon explains only in part
Measured
28 of 79
concepts silicon explains wrongly
Measured
30 of 79
concepts no silicon page explains
Measured

More concepts are explained wrongly than correctly: 28 against 16. A reader who trusts silicon gets the wrong picture more often than the right one.

Where it fails

The old docs did best where authors work, and worst in the machinery

The marks are not spread evenly. Group the concepts by bounded context, and a pattern shows.

Here is each context, split by mark. Compare authoring at the bottom with composition near the top:

Old-docs marks for the 79 concepts, by context
Number of concepts in each context
  • covered
  • covered wrongly
  • never covered
  • partly
Across contexts
2 concepts
Project lifecycle
13 concepts
Composition
8 concepts
Prompting
9 concepts
Coordinator
14 concepts
Registry
12 concepts
Authoring
21 concepts
Show as table
coveredcovered wronglynever coveredpartlyTotal
Across contexts02002
Project lifecycle147113
Composition03508
Prompting32409
Coordinator246214
Registry336012
Authoring7102221

Concept map of the trial docs tree, read from the code, 18 Nov 2026.

Authoring has the most right: 7 of its 21 concepts Measured. Composition has none right: 3 wrong and 5 never mentioned. Project lifecycle, which owns your folder, gets 1 of 13 right.

Those two weak contexts hold the ideas from earlier pages: stacking, the merge, flattening and layering. They decide which files land in your folder. Every one of the eight concepts we wrote in full was covered wrongly Measured.

Now look at the 16 that silicon got right. Most are things a user names: a template, a question, an answer, an artifact, the create command. So silicon named the parts well, but went wrong on how they work together Estimate.

All 79

You can browse all 79 concepts and filter them by context and mark

Each card below is one concept, with its one line and its old-docs mark. A card opens that concept's doc in the new docs tree. The eight marked "full doc" carry a long worked example.

Use the chips to narrow the list. Try Composition, or "never covered":

Showing 79 of 79 concepts

Across contexts 2 concepts

Project lifecycle 13 concepts

Composition 8 concepts

Prompting 9 concepts

Coordinator 14 concepts

Registry 12 concepts

Authoring 21 concepts

"Never covered" is the largest group, with 30 concepts. Many of them are about failure: session release on failure, readiness checks, exit codes. A reader needs those exactly when something breaks.

Left out

CyanPrint v4, called sulfone.lite, is left out of this map on purpose

sulfone.lite is CyanPrint v4, release 4.10.2. It is a rewrite in TypeScript. Its artifacts run as local Bun code, not in Docker containers, and it has no coordinator.

It keeps the core ideas. Resolver votes stay, now in three tiers, and so does the three-way merge with a rebuilt base. It also fixes gaps in this map: for example, deterministic state is saved.

So it is a different product, with different words. It would need its own docs tree. That tree would keep five of these six contexts, and add trust and probes. We did not build it here.