# How an agent reads a model

*Vercy for AI agents · lesson 1 of 4 · ~10 min*

## What you will learn

The cold-start protocol: what an agent does in the first five minutes with a conformant model it has never seen.

## Why agents get a protocol

An agent pointed at an unfamiliar repository faces the same risks as a human, amplified: it reads fast, misses silently, and answers confidently. The standard's response is that a conformant model *tells you how to read it*, and an agent's first duty is to accept the offer instead of improvising.

## The cold-start sequence

1. **`BOOTSTRAP.md` first.** How this model wants to be read, local conventions, hard prohibitions. If the ecosystem has its own entry file (`CLAUDE.md`, `AGENTS.md`), read it after: it carries operating state; BOOTSTRAP carries structure.
2. **`manifest.yaml`.** Identity, ordered bundles, classification rules, exclusions. Now you know what every file *is* before opening any.
3. **`sources.yaml`.** Which datasets are mastered here, which are mirrors, how fresh, who wins conflicts. Read it *before content*: otherwise you cannot tell authoritative knowledge from a stale copy of someone's wiki.
4. **Walk in declared order.** Bundles foundation-first, layers in order, layer README before records. The order is dependency order: definitions arrive before use.
5. **Run the walker if present** (`tools/mu-walk*`). A red walk means the model's own map is broken: report that before trusting anything.

The whole sequence is three small files and one command before any content: cheap insurance against every downstream mistake.

## Answering with authority awareness

When answering from the model, an agent resolves claims to records and checks each record's dataset against the register:

- **Model-mastered** → answer plainly; this is the truth by declaration.
- **Mirror, fresh** → answer, naming the master: "per the Ops wiki, as harvested 2026-07-30".
- **Mirror, stale** → say so; prefer triggering a re-harvest over guessing.
- **`status: declared`** → an intention, not data; never present it as fact.
- **`status: retired`** → frozen history: "as of the final export", never current state.

And one distinction that prevents subtle nonsense: a `raw/` capture is evidence of *what a source said*; a record is *the model's own statement*. They may deliberately disagree (the model can hold an annotated deviation); quote accordingly.

## What this buys everyone

A human team gets agents that never mistake a mockup for a spec or a dead tracker's export for a live backlog. An agent gets something rarer: the ability to *prove* its reading was complete: the coverage check names every file it did not read and why.

## Key takeaways

- Cold start: BOOTSTRAP → manifest → sources → ordered walk → walker; only then content.
- Every answer carries authority context: mastered, mirrored (how fresh), declared, or retired.
- Evidence (`raw/`) and statements (records) are different speech acts; do not blend them.

## Go deeper

- [Agent operations guide](/spec/#07-guides/Agent-Operations.md) (the normative-grade version of this lesson)
- [Model traversal (ARCH-017)](/spec/#02-architecture/Model-Traversal-and-Layout.md)

Next: [Write rules](02-writing.md)