Conformance

Conformance means a specific implementation or course meets a defined contract with evidence. The three class names below describe future qualification targets, not badges awarded by this candidate. Current tools check repository schemas, examples and selected replay behaviour. Until the later qualification work exists, no course, runner or record may claim stable 1.0 conformance (candidate and stable).

Three future conformance classes

Use the current contracts to prepare evidence. The public class criteria and runner suite remain future work, as candidate scope states. These summaries introduce no additional normative requirements.

Class	Who would claim it	Contract obligations to inspect	Evidence available today	Qualification gap
Conforming Course	The author of a course folder	Structure, references, lessons and check fixtures under the format and check rules.	Repository course/schema checks and tested example fixtures.	Public validation and fixture tooling, plus published class criteria.
Conforming Runner	An implementer of a CLI, skills or embedded application	Teaching duties, gates, capabilities, consent, pinning and protected record behaviour.	Contract cases, replay traces and the receipt boundary report; none exercises a production runner here.	A public runner suite and demonstrated behaviour on each declared host, including separate receipt qualification.
Conforming Record	An implementer that stores or exchanges learner records	Preserve event meaning, order and derived state under the record contract, including privacy and recovery.	Record/event schemas and internal replay fixtures with explicitly controlled authority.	Published class criteria and real storage/replay evidence; JSON alone cannot prove provenance.

A future claim will need its target format version, claimant and evidence. Capability declarations must identify implemented behaviour. The phrase “Agent Courses compatible” supplies none of that evidence. Candidate acceptance itself qualifies no host, package or implementation.

Capabilities and proof

A capability names what a runner can do, rather than how well a course teaches. There are exactly three; the execution contract owns their definitions. This site awards no capability to any runner or host.

Capability	Behaviour to demonstrate	Evidence it can support	Limit to demonstrate too
run-checks	Execute declared commands through host permissions.	checked evidence from the command result.	Refusals stay not run; timeouts are errors; missing capability cannot settle an objective.
observe	Inspect project files, Git state and output of commands the runner ran.	observed evidence identifying what was inspected.	Learner assertions alone are not observation; unavailable inspection stays not run.
judge	Hold the explain-back conversation and apply its rubric.	explained evidence for a judged attempt.	Without it, record the attempt and leave its objective unsettled.

Read check results, observe checks and explain-back for the exact outcomes. A runner with no capabilities can deliver lessons but settles no objectives. There is no render capability in 1.0.

Passing evidence is only part of settlement. Every declared call in the lesson must also have verified learner provenance. The receipt boundary requires protected learner confirmation, verifier authority and durable replay protection. No host qualifies here; imported receipt-shaped JSON and synthetic experiment results cannot qualify one.

Schemas check shape, not authority

The candidate has eight schema types: course, lesson, checks, feedback, lock, record, event and rubric. Their identifiers follow https://agentcourses.io/schemas/1.0.0-rc.1/<type>.schema.json. These are candidate identifiers, not a claim that live downloads have been published. The schema addresses define the policy, and the repository’s manifest lists exact source paths, identifiers and checksums.

A shape-valid file may still have broken references or false provenance. Inspect constraints beyond schemas, three kinds of validity and protected admission. Protected history, retained course bytes and original verifier context are required for authoritative replay. No schema success substitutes for them.

Proposed names are not packages

The current naming contract leaves command, library and package names unchosen. The proposals below help readers distinguish the pieces; they are not reservations, install instructions or promised import paths. Registry availability remains ⚠️ unverified until release-time checks.

Thing	Name or proposal	Status
Specification	Agent Courses	Format version 1.0; candidate label 1.0.0-rc.1.
Proposed public command	agent-courses	Proposed spelling only; no public command is released here.
Proposed Python distribution for the reference library	agent-courses-ref	Historical proposal only; release naming and registry checks remain open.
Importable Python library	No import name chosen	A distribution name does not establish an import path.

A course’s authored version, the format version, the candidate label and a runner/library release are different identities. See course versions and candidate status. Nothing in the naming table tells a reader to install a package.

Current checks and future tools

Every command below marked current runs inside the repository checkout, after the pinned setup in its README. Proposed command names are descriptive text, not runnable instructions.

Tool	Availability	Inspect today instead
Internal contract harness: uv run --locked python tools/check_contracts.py --root .	Current, checkout only; proves repository examples.	Harness source, schemas and example traces.
Public validation command, proposed agent-courses validate	Future; no released public validator.	Format rules and internal course/schema results.
Public fixture command, proposed agent-courses test	Future; no released public fixture runner.	Check fixtures and the quickstart’s checkout instructions.
Public advisory command, proposed agent-courses lint	Future; no released public linter.	Unknown-key policy and author guidance.
Reference parsing and validation library	Future; no released importable library.	Canonical format, records and identity sources.
Public runner conformance suite	Future; no qualification result is available.	Execution cases, record replay and the receipt report.

Stable launch qualification comes after candidate review. Real learner pilots, qualified host boundaries and published packages still need evidence; this candidate sets no pass mark for them. A second independent runner, external endorsement and broader governance are later work, not prerequisites for writing this candidate (scope). Start implementation with the support guide and keep every support claim narrower than its evidence.