> The full index of these docs is at [llms.txt](https://u22a8.ai/llms.txt).

# Model identifiers

A model identifier names a family, a version, a minimum accuracy and a horizon, and each part moves for its own reason.

*Applies to a8-1. Last reviewed 2026-10-22.*

A **model identifier** names the model that answers a request, such as `a8-1-1:95@20260715t140322z`. The family names the kind of judgment, the version is ours to publish, the tag is the minimum accuracy you ask for and the horizon belongs to your teaching. Each part moves for its own reason, so your data and our releases keep separate clocks.

## 1 Anatomy

A full identifier reads family, version, tag and horizon, joined by `-`, `:` and `@`:

```text
a8-1-1:95@20260715t140322z
│  │   │  └─ horizon: everything you taught up to 2026-07-15 14:03:22 UTC
│  │   └──── minimum accuracy: 95, the level a verdict is promised at
│  └──────── version: 1.1, ours to publish
└─────────── family: a8, the judgment a model makes
```

| Part              | Example            | Meaning                                                             |
| ----------------- | ------------------ | ------------------------------------------------------------------- |
| Family            | `a8`               | The kind of judgment: does the subject satisfy the stated criterion |
| Version           | `1-1`              | Major and minor, written A8 1.1 in prose                            |
| Minimum accuracy  | `95`               | The tag: a whole number from `0` to `99`, and 90 without a tag      |
| Horizon           | `20260715t140322z` | The supervision the model was fitted over, as a UTC timestamp       |
| Reference horizon | `reference`        | The horizon at zero: the reference model alone                      |

*Table 1.* The parts of a model identifier; the tag and the horizon are the parts that belong to you.

A request names a model in one of these forms:

| You send                     | It resolves to                                                             |
| ---------------------------- | -------------------------------------------------------------------------- |
| `a8`                         | The latest version, at your organization’s latest ready horizon, at 90     |
| `a8:95`                      | The same model, at a minimum accuracy of 95                                |
| `a8-1-1@20260715t140322z`    | That version at that horizon, held still, at 90                            |
| `a8-1-1:95@20260715t140322z` | That version at that horizon, held still, at 95                            |
| `a8@20260715t140322z`        | The same model as the full form, and `a8:95@20260715t140322z` with its tag |

*Table 2.* The forms a request’s `model` takes. Every response names the full identifier that answered, its level included: a request at the default level is answered as `a8-1-1:90@20260715t140322z`.

Quickstarts use the alias `a8`. Tests, and anything else that must reproduce, use the full identifier a response named, which holds the minimum accuracy still too. `GET /models` lists the alias and your horizons untagged, since every tag works with every one of them. See [Horizons](https://u22a8.ai/docs/a8-horizons.md).

The tag is the one way to set the minimum accuracy, the same in every interface. A tag outside `0` to `99` returns `400` naming `model`, and a level written in a prompt counts as part of the criterion. What the level promises is on [Accuracy and abstention](https://u22a8.ai/docs/a8-abstention.md).

## 2 What each part promises

| Part               | Moved by                                                     | Promise                                                                                                                                 |
| ------------------ | ------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------- |
| Family             | Us, rarely                                                   | A different family asks a different question. Families run side by side.                                                                |
| Version            | Us, announced                                                | A minor can move verdicts and keeps the request and response contract. A major marks a new model generation, a contract change or both. |
| Minimum accuracy   | You, on each request                                         | A verdict is right at least that often. A level is a promise of the model it tags: it moves neither the version nor the horizon.        |
| Horizon            | You, by teaching or retracting                               | The model reflects your supervision up to that moment                                                                                   |
| System fingerprint | Anything that determines a verdict, our corrections included | It changes whenever something that determines a verdict moved                                                                           |

*Table 3.* The five signals and who moves each.

Responses carry the system fingerprint. See [Reproducibility](https://u22a8.ai/docs/a8-reproducibility.md).

## 3 What moves what

| When this happens                          | Family | Version | Horizon | Fingerprint |
| ------------------------------------------ | ------ | ------- | ------- | ----------- |
| You teach or retract                       | Holds  | Holds   | Moves   | Moves       |
| We correct a defect and refit              | Holds  | Holds   | Holds   | Moves       |
| We publish a new reference model           | Holds  | Moves   | Holds   | Moves       |
| We change the request or response contract | Holds  | Moves   | Holds   | Moves       |
| We publish a different kind of judgment    | Moves  | Moves   | Holds   | Moves       |

*Table 4.* Row two is why the fingerprint exists: a correction reaches every caller, keeps every identifier as it was and announces itself in the fingerprint.

## 4 Two models, one organization

The horizon axis has a zero, so your organization has two models to call at any time:

| Model                     | Ask for                      | Holds                                                     |
| ------------------------- | ---------------------------- | --------------------------------------------------------- |
| Your organization’s model | `a8`, or one of its horizons | The reference model plus everything you taught            |
| The reference model       | `a8-1-1@reference`           | The reference model alone, as a new organization meets it |

*Table 5.* Your teaching moves your organization’s model, and the reference model stays as published.

Run the same test against both, and the difference is what your teaching changed. See [the reference model](https://u22a8.ai/docs/a8-horizons.md#the-reference-model).

## 5 Holding still

A pinned identifier holds four of the five signals: the family, the version, the minimum accuracy its tag names and your data as they stood at that horizon.

The fifth, the system fingerprint, moves. A pin holds what the model learned, and the software that serves it stays current. When we correct a defect, the correction reaches pinned callers too, so every caller gets the corrected verdict. The identifier stays the same and the system fingerprint changes. That change is the signal to re-check a stored baseline.

> **The rule in one line.** The identifier says what the model learned. The fingerprint says which build served it.

## 6 Each version has its own horizons

A horizon belongs to the version that fitted it, and you ask for it under that version. A new minor starts its own line of horizons, and your pins on the earlier minor keep answering as they did.

Moving to a new minor is automatic. We refit your teaching on the new reference model once it is published. Until that fit completes, the alias `a8` answers from the model you already have. A model you pinned stays as it is until its version retires. See [Deprecations](https://u22a8.ai/docs/deprecations.md).

## 7 Related concepts

- [Horizons](https://u22a8.ai/docs/a8-horizons.md): listing horizons, pinning one and the provisional horizon
- [Reproducibility](https://u22a8.ai/docs/a8-reproducibility.md): reading the system fingerprint
- [Teaching](https://u22a8.ai/docs/a8-teaching.md): what moves your horizon

## Next steps

- [Horizons](https://u22a8.ai/docs/a8-horizons.md): A horizon is the timestamp that names a fitted version of your model, and pinning one holds its verdicts still.
- [Reproducibility](https://u22a8.ai/docs/a8-reproducibility.md): The system fingerprint tells you when anything behind a verdict moved, and the model identifier reproduces a verdict.
- [Deprecations](https://u22a8.ai/docs/deprecations.md): The lifecycle status of each A8 version, how much notice a retirement gets and how to watch for one.
