# Bundles, layers, records

*Building a meta-model · lesson 2 of 6 · ~15 min*

## What you will learn

How model content is actually organized and written: the bundle → layer → record hierarchy, naming that carries meaning, and what makes a good record.

## The hierarchy

- A **bundle** is one semantic responsibility: "why the product exists", "system design", "quality and risk". Bundles are ordered: foundations first, so a single sequential read meets definitions before use.
- A **layer** is one question inside a bundle: "who are the stakeholders?", "what are the APIs?". A layer directory carries a `README.md` that states the questions the layer answers: the README *is* the layer's meaning declaration.
- A **record** is one independently routable statement: one requirement, one decision, one API contract, one risk. One record, one file.

The flagship domain meta-model, AISMM, fixes thirteen bundles for the domain "software product" (b0 product core through b12 economics) with ~90 layers. Your domain will have its own bundles; the mechanics are identical.

## Names that machines can read

Records follow a naming convention that carries classification in the filename:

```text
{record_kind}-{YYMMDDNNNNN}-{memo}.md
   req-26071800001-minor-parental-consent.md
   decision-26061500001-concept-decisions.md
   i18n-26060300001-languages-roster.md
```

The prefix is the record kind (requirement, decision, api, risk...), the digits are a sortable date-based ID, the memo is for humans. Because the kind is in the name, a single declared rule in the manifest classifies thousands of files (next lesson). Two hard-won details: prefixes may contain digits (`i18n`), and the ID never changes after birth, even if the memo or content does: stability of reference beats prettiness.

## What is inside a record

A record has two parts: a machine block and a human body. The machine block (front matter) carries identity (document ID, layer, product), classification (`kind_class`: is this *normative* for the product, *descriptive* of it, or a *projection* of something else?), and, crucially, **derivation**: if an agent produced this record from sources, the block says from what, by whom, how transformed, and whether an owner has validated it. The body is ordinary, well-written prose and structured data: the standard has no opinion about your writing style, only about your traceability.

The `kind_class` + `derivation` + `validation_status` triple is what lets mixed human/agent teams trust a model: you can always tell an owner-validated normative statement from an agent's unvalidated summary of a wiki page.

## What makes a record good

- **One concern.** If you cannot title it without "and", split it.
- **Routable.** Someone assigned "handle req-26071800001" needs nothing else to find their work.
- **Referenced, not repeated.** Records cite canon, evidence in `raw/`, and each other; the same fact stated twice will eventually be true once.
- **Honest about status.** Draft, unvalidated, superseded: states are metadata, not secrets.

## Key takeaways

- Bundle = responsibility, layer = question, record = one routable statement.
- The filename carries the kind; the front matter carries identity, classification and derivation.
- Traceability (who derived this from what, who validated it) is what makes agent-written content usable.

## Go deeper

- [MMAS core architecture](/spec/#02-architecture/MMAS-Core.md) · [Naming conventions](/spec/#02-architecture/Naming-Conventions.md)
- [Creating a new meta-model](/spec/#07-guides/Create-a-New-Meta-Model.md)
- AISMM, the flagship domain meta-model: [github.com/orkestron-ai/software-meta-model](https://github.com/orkestron-ai/software-meta-model)

Next: [The lossless walk](03-walk.md)