# Conformance and Qualification

Conformance answers whether an implementation follows a protocol contract.
Qualification answers whether a particular build and deployment works with its
real dependencies under the required conditions. AIP treats them as related but
different evidence.

## Evidence Levels

| Level | What it proves | What it does not prove |
|---|---|---|
| Unit test | One code unit behaves as asserted | Cross-process interoperability |
| Schema fixture | JSON shape matches a versioned schema | Runtime policy or side effects |
| Core conformance | Native envelopes and invariants are implemented | External provider behavior |
| Profile conformance | A compatibility mapping follows its version matrix | Every independent client works |
| Connector frozen conformance | Advertised connector operations follow the SDK contract | Real provider availability or semantics |
| Integration test | Repository components work together | Deployment identity, network, or provider behavior |
| Isolated-live qualification | Exact build works with a pinned real product in an isolated topology | Arbitrary production topology or load |
| Production campaign | Deployment survives required security, restart, fault, and load scenarios | Future untested versions |

## Native Core Checks

Run the built-in fixture suite:

```sh
cargo run -p aipctl -- conformance run
```

Validate a specific envelope:

```sh
cargo run -p aipctl -- conformance run --envelope envelope.json
```

Export machine-readable schemas:

```sh
cargo run -p aipctl -- schema export ./exported-schemas
```

Release checks must verify the generated output against the versioned files
under `schemas/aip`.

## MCP Compatibility

Run against Streamable HTTP:

```sh
cargo run -p aipctl -- mcp conformance \
  --url http://127.0.0.1:18080/mcp
```

Run against stdio:

```sh
cargo run -p aipctl -- mcp conformance \
  --command aipd \
  --arg=--mcp-stdio \
  --arg=--service-id \
  --arg=agent:aipd:mcp-conformance
```

The complete release matrix covers stable MCP versions `2024-11-05`,
`2025-03-26`, `2025-06-18`, and `2025-11-25`, method-family negative cases,
session ownership, cancellation races, replay, OAuth boundaries, and independent
client/server implementations.

## Connector Conformance

A frozen connector suite must test every behavior the manifest advertises:

- exact capability and schema admission;
- trusted execution context;
- successful invocation and typed error mapping;
- streaming ordering and terminal behavior;
- cancellation confirmation;
- idempotency collision behavior;
- approval resume;
- plan, commit, compensation, and uncertain-outcome reconciliation;
- webhook authenticity and replay;
- callback and event emission;
- restart recovery.

Unsupported behavior should fail explicitly and must not be advertised.

## Retained Evidence

An isolated-live report should contain no credentials and should identify:

- source commit and clean/dirty status;
- source-tree digest;
- built image digest;
- connector and upstream revisions;
- configuration class without secrets;
- start and finish times;
- scenario ids and outcomes;
- protocol message types and action ids where safe;
- restart, replay, approval, and provider evidence;
- the exact command used to reproduce the campaign.

Retained evidence must be copied from the raw run artifact without manual
rewriting. A byte-for-byte comparison is preferable.

## Repository Release Gate

```sh
cargo fmt --all --check
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo test --workspace --all-features
cargo test --workspace --no-default-features
cargo doc --workspace --all-features --no-deps
cargo run -p xtask -- release-check
```

Optional audit, dependency-policy, fuzz, semver, benchmark, load, and chaos
gates are required when the release policy or changed surface calls for them.

## Status Language

Use `Done fully` only when every evidence class required by the item exists and
passes. Code completion does not satisfy a requirement that explicitly calls
for external interoperability, multi-process deployment, load, fault, or
credentialed product evidence.