Horizons

A horizon identifies the supervision a model was fitted over: a timestamp that reads “this model reflects everything you taught up to here”. Every fit of what you teach mints one. The alias a8 answers from your latest ready horizon, and a request that names a horizon holds the model still.

§1Reading a horizon

A horizon is a compact UTC timestamp. 20260715t140322z is 2026-07-15, 14:03:22 UTC: the moment of the newest change to what the model learned from. It tells you exactly which of your teaches a model holds.

Teaching and retracting both change what the model learned from, so either one moves the horizon. The earlier horizon stays as it was and stays pinnable, so pinning it is a second way to undo a retraction.

Horizons sort as strings, oldest first. A horizon belongs to the version that fitted it, and a request names it in the model field after an @:

// annotated: the model field in a request
// your organization's model, moving as you teach
"model": "a8"

// one horizon of it, held still
"model": "a8-1-1@20260715t140322z"

§2Listing your horizons

GET /models lists the models your key can call: the alias a8, every horizon your organization has fitted, newest first, then the reference model of the newest version. The reference model of an earlier version stays pinnable by its identifier.

curl https://api.u22a8.ai/eval/v1/models \
  -H "Authorization: Bearer $U22A8_API_KEY"

import os

from openai import OpenAI

client = OpenAI(
    base_url="https://api.u22a8.ai/eval/v1",
    api_key=os.environ["U22A8_API_KEY"],
)

for model in client.models.list():
    extra = model.model_extra or {}
    print(model.id, extra.get("status"), extra.get("active"))

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.u22a8.ai/eval/v1",
  apiKey: process.env.U22A8_API_KEY,
});

for await (const model of client.models.list()) {
  const { status, active } = model as typeof model & {
    status: string;
    active: boolean;
  };
  console.log(model.id, status, active);
}

The list while a fit is pending:

// annotated, trimmed: each entry also carries created, owned_by,
// lifecycle and shutdown_date
{
  "object": "list",
  "data": [
    {
      "id": "a8", // the alias, answering from the active horizon
      "object": "model"
    },
    {
      "id": "a8-1-1@20260716t091140z",
      "object": "model",
      "status": "provisional", // 14 teaches recorded, fit to follow
      "active": false
    },
    {
      "id": "a8-1-1@20260715t140322z",
      "object": "model",
      "status": "ready", // what a8 answers from
      "active": true
    },
    {
      "id": "a8-1-1@20260711t082915z",
      "object": "model",
      "status": "ready", // last week's model, still pinnable
      "active": false
    },
    {
      "id": "a8-1-1@reference",
      "object": "model",
      "status": "ready", // the reference model alone
      "active": false
    }
  ]
}

Two fields tell you when a model is in service, and two more when it retires:

§3Pinning

A pin holds a verdict still. A pinned horizon answers from the exact reference model it was fitted over, so later teaching and later releases move the alias a8 and leave the pin where it was.

A pin holds what the model learned, and the software that serves it stays current. A correction we ship reaches pinned callers too. The system fingerprint tells you when that happened. See Model identifiers.

§3.1Every response names its horizon

Every response’s model field carries the exact identifier that answered, including when the request asked for a8. Send that value back as model, and the same request returns the same verdict:

// annotated
// the request asked for
"model": "a8"

// the response answered as; keep this beside the verdict
"model": "a8-1-1@20260715t140322z"

Keep the response’s model beside anything you need to re-derive. Before your first fit, the reference model answers, and the field names it.

§3.2When to pin up front

• A regression suite or a published benchmark. Pin, so a re-run measures your system as it stood.
• A scored result you stand behind later. Pin, and record the identifier beside the result.
• Production judging that keeps improving. Stay on a8, and watch the system fingerprint to learn when the model moved.

§4The reference model

Beside your organization’s model, you always have the reference model: A8 as we publish it. A new organization starts on it, and it stays available however much you teach.

// annotated
// your organization's model: everything you taught
"model": "a8"

// the reference model: A8 as we publish it
"model": "a8-1-1@reference"

reference sits at zero on the horizon axis. A horizon names the supervision a model was fitted over, and this one names ours alone, so it is a word, dated by the version it belongs to.

The reference model serves three uses:

• Measuring what your teaching changed. Run the same test against both models. The difference is your teaching, isolated.
• A fallback. Where teaching has taken your model off course, the reference model answers as published.
• A second opinion on a case where your own conventions decide the verdict.

Each reference model keeps its name. An improved reference model is a new minor version: a8-1-2@reference names it, and a8-1-1@reference keeps naming the earlier one. Everything you pinned on the earlier minor answers as it did.

Your organization’s model moves forward on its own. We refit your teaching on the new reference model once it is published, and the alias a8 moves to the new minor when that fit completes. Until then it answers from the model you have, with every teach in it. The new horizon appears in GET /models, and the system fingerprint changes with it.

§5The provisional horizon

After you teach, the models list shows the identifier the next fit mints, with status: "provisional". It confirms the teach was recorded, and it gives you the identifier to ask for once the fit completes.

A request for a provisional identifier returns 503 with Retry-After: 30 until the fit completes. The OpenAI and Anthropic SDKs retry a 503 on their own, so a pinned request waits for its own model.

§6Errors

§7Related concepts

• Model identifiers: the parts of an identifier and what moves each
• Teaching: what moves the horizon
• Reproducibility: the response field that says something moved