# AI Peer Profile

> How an AI collaborator can inhabit a truth system without becoming its authority or safety dependency.

An AI Peer is a capable collaborator with durable access to the truth system's
context, governed reads, tools, and operating rules. The office survives model
replacement and conversation loss.

The profile is optional. Core Truth Machine conformance requires no AI.

## Durable office

The workspace must answer without prior conversation:

| Question | Durable answer |
| --- | --- |
| What system am I in? | Purpose, scope, vocabulary, coherence boundary, and current status |
| What is current? | Governed read with uncertainty and revision |
| Why is it current? | Evidence, source authority, ancestry, and accepted changes |
| What is unfinished? | Proposed, queued, stale, failed, and unknown work |
| What may I do? | Capabilities, actor authority, review gates, and prohibitions |
| How is success proven? | Deterministic tests, observations, recovery, and rollback |

## Front door

Place concise instructions such as `AGENTS.md` at the workspace boundary. The
file points to canonical state and tooling; it does not duplicate the entire
system in a giant prompt.

Instructions should identify:

- authoritative documents and read order;
- safe read operations;
- typed write operations and dry-run forms;
- evidence and untrusted-input boundaries;
- approval requirements;
- verification commands; and
- recovery behavior.

## Read before reasoning

Give the Peer deterministic structured reads over the complete relevant
coherence boundary. Raw database, file, or repository access is not a governed
read if the Peer must repeatedly reconstruct domain meaning or silently choose
which records count.

Structured values and human meaning should travel together. Unknown,
suppressed, absent, stale, source-only, and current values must remain
distinguishable.

## Writes on rails

Consequential actions belong behind typed operations with declared inputs,
expected state, effects, invariants, actor authority, and proposal or dry-run
forms. The Peer may perform mechanical application only within the authority
already declared.

Promotion moves repeated mechanical labour into code; it does not move review
authority into the model.

## Untrusted content

Evidence can contain text that resembles instructions. Source documents,
emails, web pages, logs, and contributions are data unless the workspace's
authority model expressly names them as instructions.

The Peer must not:

- follow source-embedded commands merely because they are imperative;
- treat model-generated summaries as source evidence;
- expose credentials or private evidence through public tools;
- accept its own proposed authority expansion; or
- use conversation memory as current truth.

## Harness portability

The bond is to the durable office, not one model. Another capable harness
should be able to read the instructions, reproduce the current account, trace
its ancestry, identify pending work, and use the same safe operations.

Use the [AI Peer Commission](/docs/truth-machine/guides/peer-commission/) to assess an existing
workspace.