Standard · EMO v1.0 Draft

Ethical Model Operations

Ten things a deployed AI model must never do: model APIs, chat apps and products that serve a model to people.

Certify a repo against EMO Read EMO.md Try in sandbox

GitHub publication is waiting on Rex. This link serves the Markdown file in the repo.

Badge a passing repository earns example badge

+------------------------------------------+
|    o---o---o        ASC  EMO   v1.0      |
|     \ / \ /         ETHICAL MODEL        |
|      o---o          OPERATIONS           |
|     / \ / \                              |
|    o---o---o        PASS  10/10          |
|  your-org/repo@a1b2c3d                           |
|  sha256:4b1e07c2a9d3   2026-11-04                 |
+------------------------------------------+

Inscription: not yet (Draft) · policyRef: sha256 of EMO.md

A model deployment shall never do

Rule	What it means, and the evidence the jury reads	Weight
EMO-01	MUST NOT serve a different model than the one it names. Responses carry the exact model id and version actually served; no silent fallback to another model. Evidence: routing, fallback logic, response metadata	critical
EMO-02	MUST NOT change a pinned model version silently. A pinned version stays pinned; changes are announced in config and responses. Evidence: version pinning, release config	normal
EMO-03	MUST NOT train on user data without opt-in. Prompts and outputs are excluded from training pipelines unless the user opted in. Evidence: data pipelines, opt-in flags	critical
EMO-04	MUST NOT keep prompts longer than the declared retention. Retention is configured and deletion jobs exist. Evidence: retention config, deletion jobs	normal
EMO-05	MUST NOT leak one user's data into another user's context. Caches, memories and retrieval are scoped per user or tenant. Evidence: cache keys, memory and retrieval scoping	critical
EMO-06	MUST NOT send prompts to undeclared third parties. Prompts go only to declared processors. Evidence: outbound calls, processor list	critical
EMO-07	MUST NOT insert undisclosed paid placements into answers. Sponsored content, if any, is labeled in the response. Evidence: ranking and injection code, labels	normal
EMO-08	MUST NOT present generated content as human-made. Generated media and text are labeled where the product publishes them. Evidence: publishing paths, labels and metadata	normal
EMO-09	MUST NOT run without a way to report harmful outputs. A report path exists and reports are stored. Evidence: report endpoints, storage	normal
EMO-10	MUST NOT claim results it can't reproduce. Published evals and benchmarks include configs that reproduce them. Evidence: eval configs, published numbers	normal

Critical rules weigh 3, others 1. PASS needs no fail and no needs-evidence, with at least 60% of the weight checkable from code.

Changelog

- 1.0 (2026-11-04): first draft.