diff --git a/agent-os/product/README.md b/agent-os/product/README.md
index d8da9a59..4ae6b720 100644
--- a/agent-os/product/README.md
+++ b/agent-os/product/README.md
@@ -2,6 +2,7 @@
 
 This directory contains artifacts produced by the **Agent OS product planning workflow** (mission, roadmap, tech stack, naming).
 
+- [names-compose-and-module.md](names-compose-and-module.md) — **start here for naming**: the settled names (Prisma Compose, Module, Prisma App), the philosophy, the rubrics, why each candidate won or lost, and the lessons.
 - [naming.md](naming.md) — how we name by user value; the piece names (Prisma Compose, Module, Prisma Data; "Prisma App" = the artifact) and the compose-from-blocks distribution model.
 - [vocabulary-tests.md](vocabulary-tests.md) — the spoken-sentence rubrics: unit nouns (System → Module, ADR-0025) and referential product names (Prisma App → Prisma Compose, ADR-0026).
 
diff --git a/agent-os/product/names-compose-and-module.md b/agent-os/product/names-compose-and-module.md
new file mode 100644
index 00000000..96596a32
--- /dev/null
+++ b/agent-os/product/names-compose-and-module.md
@@ -0,0 +1,167 @@
+# The names: Prisma Compose and Module
+
+How the framework and its unit of composition got their names — the philosophy,
+the selection rubrics, the candidates that lost, and the lessons worth keeping.
+The binding decisions are [ADR-0026](../../docs/design/90-decisions/ADR-0026-name-the-framework-prisma-compose.md)
+(Compose) and [ADR-0025](../../docs/design/90-decisions/ADR-0025-name-the-unit-of-composition-module.md)
+(Module); the reusable rubrics live in [vocabulary-tests.md](vocabulary-tests.md).
+This document is the story that connects them.
+
+## The names at a glance
+
+> You build a **Prisma App** by composing **Modules** with **Prisma Compose**, and
+> deploy it to Prisma Cloud.
+
+| Name | What it names | Register |
+|---|---|---|
+| **Prisma Compose** | the framework — the tool you work *in* | milieu (referential token) |
+| **Module** / `module()` | the unit of composition — what you plug together | vocabulary noun |
+| **Prisma App** | the artifact — the thing a user builds and deploys | value word (the hero noun) |
+| **Service** | one deployed compute unit inside a Module | vocabulary noun |
+| **Extension** | what you slot into `prisma-compose.config.ts` | vocabulary noun |
+| **package** | the artifact npm hosts | npm's word — never renamed |
+| **Topology** | the graph the framework infers | machinery — users never say it |
+
+Surfaces follow the framework name everywhere a user meets it: packages
+`@prisma/compose` and `@prisma/compose-prisma-cloud`, CLI `prisma-compose`, config
+`prisma-compose.config.ts` ([ADR-0027](../../docs/design/90-decisions/ADR-0027-two-packages-compose-and-compose-prisma-cloud.md)).
+
+## The philosophy
+
+Every name here is chosen by asking **what job the word actually has to do**, and
+the jobs differ by register. Three principles, each learned by shipping a wrong
+name first:
+
+**1. Products are named for user value — but a name must refer as well as
+describe.** "Name the value, not the machinery" gave the family its role names
+(Compute, Data, Postgres) and gave the artifact its hero noun: people say "my
+app," never "my topology." But a name has two jobs — *describe* (what is this?)
+and *refer* (which thing do you mean?) — and value-naming only measures the first.
+The framework is not a component mentioned occasionally; it is a **milieu**
+developers live inside and talk *about* constantly. Milieu names must be
+referential tokens — which is why every durable framework name (Rails, Django,
+Vite) names mechanism, exactly what the value rule forbids for components.
+
+**2. Vocabulary nouns are not product names.** To a user, a container noun is
+*all* machinery — the value lives in the capability (cron, auth), never in the
+box. So the unit noun is chosen by a different rubric entirely: sentences said
+aloud, at every grain, in every frame (authoring, composing, installing,
+publishing).
+
+**3. Identity belongs to the registry, not the unit noun.** Branded unit nouns
+(Hex; later Prism, Shard, Facet as candidates) fail by construction — each gets
+explained via the common word it displaced, and taxes every composition sentence
+to buy an ecosystem signal that a registry name carries better. The registry name
+is deferred until the registry exists; Prism sits on its shortlist.
+
+## The rubrics
+
+Three test batteries, applied by register (full versions in
+[vocabulary-tests.md](vocabulary-tests.md)):
+
+**Product value tests** (components): would the user say "my ___"? does it predict
+the tooling? does it name the goal, not the tax? does it keep the family legible?
+
+**Referential battery** (products people talk about, above all the framework):
+- *Workbench frame:* "I'm working on this feature in ___" — the daily sentence.
+- *Artifact-overlap flag:* a tool named after its output word tends to fail the
+  frames above, because the output word saturates every conversation around the
+  tool. A red flag that says run the frames — not a rule that decides by itself.
+- *Bare-token test:* people drop prefixes; does the name survive unprefixed?
+- *Identity frames:* "Intro to ___", "___ 2.0", "does ___ support X?"
+
+**Unit-noun spoken tests** (vocabulary): grain sweep ("the auth ___" *and* "a ___
+providing one cron job"), repetition (three in one sentence, invisibly), install
+("install the cron ___"), publish, instance/artifact duality, adjacency with
+Service/App/Extension, the gloss test (if the explanation contains a better word,
+the gloss wins), and a prior-art check against how developers actually talk.
+
+## Why "Module"
+
+The unit was **Hex**, then **System**, then Module — each failure teaching one of
+the rules above.
+
+**Hex** (a nod to hexagonal architecture) put ecosystem identity on the unit noun:
+uniquely ours, wrong at first reading (a color, hexadecimal, a curse), a jargon
+tax on every newcomer. **System** fixed the naive reading and failed in speech the
+moment the first *small* shared unit existed: "install the cron system" is a
+category error — English "system" has a size floor and an ownership register (a
+system is something you *operate*, not something you install). "This system uses
+the cron system and an image-resizer system" is barely sayable. The deciding
+claim behind it — "nobody says 'my auth module'" — was simply false in the
+ecosystems nearest ours (Nest, Angular, Terraform, Go), and our own README
+betrayed the choice by introducing it as "a component — a System."
+
+**Module** passes every spoken test. No size floor or ceiling ("the auth module,"
+"a module providing one cron job," "your app is the outermost module"). Natural
+in the consuming frames, because npm packages already *are* modules colloquially —
+which supplies instance/artifact duality for free. Its prior art means exactly
+our meaning: a Nest or Angular module is a composable boundary-owning unit wired
+by dependency injection; a Terraform module is a reusable composition with typed
+inputs and outputs. Nest's convention that the root module *is* the app
+corroborates our model — the App is simply the outermost Module. The one real
+adjacency, ES modules, is inter-register: file-grain vs composition-grain, and no
+spoken sentence confuses them.
+
+Runner-up: **component** — the word our own README glossed with, disqualified
+because a Module *contains* React components; "the storefront component" is
+ambiguous inside a single conversation.
+
+## Why "Prisma Compose"
+
+The framework was **MakerKit**, then **Prisma App**, then Compose — same lesson,
+other register.
+
+**MakerKit** was a standalone-mascot name that sat outside the family. **Prisma
+App** scored perfectly on the value tests — it *is* the value word — and failed
+every referential frame the first week the team lived with it: "I'm working on
+this feature in App" refers to nothing. The mechanism: "App" is the user's
+artifact word — the most common noun in every conversation around the tool — so
+the workbench and bare-token frames had nothing to grab. The fix kept the value
+word where it was always right (**Prisma App is the artifact**) and gave the tool
+a referential token.
+
+**Compose** won a graded rubric over Construct, Alloy, Assemble, Wire, Forge,
+Stack, Prism, "Prisma Framework," and a nineteen-word synonym sweep — the only
+candidate with no failing grade. It is also the semantically exact verb: build,
+construct, make, and forge describe fabricating from raw material; *compose*
+means assembling finished parts into a whole — functions, music, apps — which is
+why "composability" was already the word for the property the framework sells.
+It restores the family's own logic (the framework's role in the product table was
+always "compose," so name and role now agree, as they do for Compute), and it
+completes the vocabulary chain in one sentence: *you compose Modules with Prisma
+Compose into your Prisma App.*
+
+Near misses: **Construct** (fabrication semantics; a third com/con- token; "a
+construct" blurs with Module), **Assemble** (already a pipeline stage —
+`assemble` — and the whole cannot share a name with its part), **Alloy** (best
+metaphor, metaphor tax). Accepted trade-offs, eyes open: the **Compute/Compose**
+in-family adjacency (frames never overlap — things run *on* Compute, are built
+*with* Compose), the shared token with Docker/Jetpack Compose (evidence it
+carries a prefix well), and an SEO fight with docker-compose-with-Prisma content
+(content contention is winnable; grammatical jamming — "Prisma app" — is not).
+
+## Lessons worth keeping
+
+- **Test names in speech, not on the page.** Every failure here read fine in
+  writing and died in a spoken sentence. Capital letters are silent.
+- **Test at the smallest grain.** Composition math puts most units below any hero
+  example; "the auth system" proved nothing about "the cron system."
+- **Naming a tool after its output invites referential failure.** The output word
+  saturates the conversation; treat the overlap as a red flag and run the frames.
+- **The gloss test is a tell.** If you introduce a name by reaching for another
+  word ("a component — a System"), the gloss is the better name.
+- **Check "nobody says X" against reality.** The deciding claim for System was
+  false in four neighboring ecosystems.
+- **Put identity where it's said occasionally, not constantly.** Branded nouns
+  belong on destinations (the registry), not on words used three times per
+  sentence.
+
+## Related
+
+- [ADR-0025](../../docs/design/90-decisions/ADR-0025-name-the-unit-of-composition-module.md) — Module: the decision and full alternatives record.
+- [ADR-0026](../../docs/design/90-decisions/ADR-0026-name-the-framework-prisma-compose.md) — Prisma Compose: the decision and full alternatives record.
+- [ADR-0027](../../docs/design/90-decisions/ADR-0027-two-packages-compose-and-compose-prisma-cloud.md) — the two public packages carrying the names.
+- [vocabulary-tests.md](vocabulary-tests.md) — the rubrics with both failure records.
+- [naming.md](naming.md) — the product family and register model.
+- [naming-proposal.md](naming-proposal.md) — the original proposal that opened the discussion (historical).
diff --git a/agent-os/product/vocabulary-tests.md b/agent-os/product/vocabulary-tests.md
index 0045d7c8..9f1606f3 100644
--- a/agent-os/product/vocabulary-tests.md
+++ b/agent-os/product/vocabulary-tests.md
@@ -94,10 +94,10 @@ referential frames:
 1. **The workbench frame.** "I'm working on this feature in ___." Said by a
    contributor, understood by a stranger. The daily sentence; if it fails, the team
    invents shorthand and the name is already dead internally.
-2. **The artifact-collision rule** (hard rule, not a scored test). The name must not
-   be the word for what users produce with the tool. A tool named after its output
-   can never be referred to separately from the output — every mention of the tool
-   parses as a mention of the artifact, and no prefix can rescue it.
+2. **The artifact-overlap flag.** A tool named after its output word tends to fail
+   the other frames, because the output word saturates every conversation around
+   the tool. Not a rule that decides by itself — a red flag that says run the
+   frames and weight them heavily.
 3. **The bare-token test.** People drop prefixes; does the name survive unprefixed?
    ("Rails" yes; "App" no.)
 4. **Identity frames.** "Intro to ___" as a talk title, "___ 2.0", "does ___
@@ -115,7 +115,7 @@ which is the proper register for a milieu, not a violation of the value rule.
 | Test | Result |
 |---|---|
 | Workbench frame | Fails: "I'm working on this feature in App" refers to nothing — the sentence the team needed all week and could not say. |
-| Artifact collision | Fails the hard rule: the tool was named after its output. "App" is the user's artifact word. |
+| Artifact overlap | "App" is the user's artifact word — ambient in every sentence about the work — which is why the frames above had nothing to grab. |
 | Bare token | Fails: "app" is ambient noise in every conversation about software. |
 | Identity frames | Fails: "App 2.0" and "Intro to Prisma App" cannot pick out the product. |