# Inference Operations

> Evidence for typed current projection and governed declared-versus-observed reads.

## Assessment identity

- Repository: `kinra-ai/inference-ops`
- Revision: `f6d68351aea8dfc811a1153169b93739c9d0dd51`
- Assessment date: 2026-08-09
- Evidence class: bounded maintainer attestation; topology and operational
  metadata remain private

## Domain

The repository records the selected composition, validation, history, and
reversal of a local inference fleet.

## Mechanisms observed

- a typed `current.toml` selects the active profiles and a narrow set of
  properties worth checking;
- fuller immutable Markdown profiles retain provenance, topology,
  measurements, validation, and rollback;
- a human index, typed selection, and full profiles must agree;
- an offline governed reader checks repository coherence without fetching or
  mutating;
- an optional live read compares declared services, artifacts, and secret-free
  endpoint contracts with observations;
- every check carries expected, observed, status, source, and operational
  meaning;
- `match`, `mismatch`, and `unknown` remain distinct; and
- an unreachable system is not normalized to an expected state.

## What it paid to teach

Current truth need not duplicate the complete journal. A small executable
projection can select what matters now while richer immutable records retain
the evidence and reversal context.

It also established the observation semantics in `TM-OBS-005` through
`TM-OBS-009`: the reader must not repair the state it measures, and absence of
an observation is not evidence of match or mismatch.

## Limits

This implementation is a fleet-specific truth machine, not a generic
reconciliation or mutation system. Its sole-operator policy, TOML schema,
service checks, and deployment profile structure are not portable
requirements. It provides strong evidence for governed reads and current
projection, not by itself for every publication requirement.