# Kin

> Evidence for an AI Peer durable office, explicit authority limits, untrusted-input handling, and structural Kit adoption without a core conformance claim.

## Assessment identity

- Repository: `kinra-ai/kin`
- Revision: `af026b1eb3887d67687000625a110b98df7cab34`
- Assessment date: 2026-08-12
- Profile: AI Peer; not a core Truth Machine implementation
- Evidence class: exact-revision bounded maintainer attestation; the assessed
  revision is present on `origin/dev`

## Domain

Kin is a local AI peer and application harness. It inhabits many independently
owned workspaces, uses their tools and operating rules, and preserves its own
configuration, sessions, application behavior, and release evidence.

Its coherence boundary is Kin itself. A workspace keeps authority over its
domain truth, evidence, review, and advancement. Inference Ops owns fleet
selection and observation, Cloud Ops owns host and service appointments, Kinra
Site owns public rendering, and Paddock owns local-model operation.

## Mechanisms observed

- `AGENTS.md`, `CONSTITUTION.md`, `STATUS.md`, and the technical reference form
  a durable office that survives conversation and model replacement;
- `truth-machine.toml` maps all nineteen Kit patterns to nine adopted
  product-native mechanisms and ten explicit declines;
- the owned structural checker reported nineteen matches, zero mismatches, and
  zero unknowns at the assessed revision;
- typed tools, permission classes, exact action approvals, and dry-run or
  proposal seams constrain consequential work without granting the model
  independent authority;
- web, remote, MCP, media, subagent, and repository content enters through
  explicit untrusted-data framing before any separate action boundary;
- effective settings retain source provenance while locked writers keep
  provider configuration, credentials, and grants in distinct stores;
- numbered decisions retain rationale while bounded status prose names current
  work and known unknowns; and
- the complete local promotion gate passed with 3,280 tests and four expected
  skips. Separate bounded client gates observed the public `default` text alias
  and explicit vision route without turning those observations into fleet
  truth.

## What it paid to teach

An AI Peer can adopt the Kit usefully without being relabelled as the truth
system it assists. Structural adoption improves orientation, authority
boundaries, evidence hygiene, and explicit negative space even when the peer
does not own domain reconciliation or current truth.

Kin also demonstrates why operational selection belongs behind a stable
contract. Its hosted inference configuration selects the public `default`
alias while the concrete Qwen runtime, capacity, topology, and rollback stay
in Inference Ops. The observed response model is evidence about one call, not
a client-owned deployment declaration.

## Limits

The adoption checker proves declared paths and deliberate declines, not the
semantics of every mechanism and not Truth Machine conformance. Kin has no
single governed read over a complete product coherence boundary, no evidence
or change packet lifecycle, no truth-transition atomicity, and no
publication-freeze or reproducible-generated-artifact claim.

Tool approvals bind actions, not acceptance of domain truth. Application
settings are a current projection only for Kin configuration. Live model tests
qualify bounded client behavior and do not prove fleet health. The pushed
revision makes this source assessment durable; it does not elevate the report
into a conformance certificate or an operations declaration.