systemone.yaml

Why a manifest

Storage does not make a model adoptable. Finding the right one does, and that needs metadata a machine can filter on: what the model decides, how well calibrated it is, how fast it responds, what it was built from.

systemone.yaml is deliberately small enough that authors write it by hand, and closed where being closed is the point.

A complete example

spec_version: "0.1"
model: support-router
namespace: biplov
category: system-one
architecture: laya
base_model: acme/laya-base
license: apache-2.0
summary: Routes inbound support tickets to the right queue.

capabilities:
  - choice
  - route

tags:
  - routing
  - support

runtime:
  framework: pytorch
  entrypoint: support_router:load
  python: ">=3.10"
  hardware: cpu

artifacts:
  - kind: huggingface
    uri: https://huggingface.co/biplov/support-router/resolve/main/model.safetensors
    filename: model.safetensors
    sha256: 0000000000000000000000000000000000000000000000000000000000000000

evaluation:
  suite: s1-decision-bench
  suite_version: "0.1"
  decision_accuracy: 0.943
  calibration_error: 0.031
  valid_action_rate: 0.999
  median_latency_ms: 4.7
  p95_latency_ms: 9.1

Fields

Field	Required	Notes
spec_version	yes	Currently "0.1". An unknown version is rejected rather than guessed at.
model	yes	Repository name. Lowercase, 2–64 characters.
namespace	yes	Your username or an organization you own.
category	yes	Always system-one.
architecture	yes	laya, jev, or another identifier. Free text until the field set is proven.
capabilities	yes	One or more of choice, score, rank, classify, extract, route.
runtime	yes	framework required; entrypoint, python, hardware optional.
license	no	A curated SPDX subset, or other with license_name.
base_model	no	namespace/name this was fine-tuned from.
summary	no	One line, at most 280 characters.
tags	no	Up to 20, lowercase.
artifacts	no	Where the files live. Never uploads.
evaluation	no	Strongly recommended. See below.

Unknown fields are rejected, not ignored — a typo should fail loudly rather than silently drop the value you meant to set.

Capabilities are closed

capabilities is an enum, not free text. This is the field that makes the registry filterable, and a field that accepts anything is a field nobody can filter on. Six values cover the decision primitives:

• choice — pick one option from a candidate set
• route — choice, where the options are destinations
• classify — assign a category from a fixed taxonomy
• score — return a number on a continuum
• rank — order a set of items
• extract — pull typed values out of an input

Evaluation is what earns trust

Accuracy alone does not describe a decision model.

• decision_accuracy — how often it is right.
• calibration_error — whether its confidence means anything. A model that reports 0.95 on cases it gets right 70% of the time makes every threshold you set from it wrong.
• valid_action_rate — how often it picks an option that actually exists.
• median_latency_ms / p95_latency_ms — whether it can sit in a request path.

Publishing calibration beside accuracy is the single thing that makes your model easier to adopt than an otherwise identical one.

Validating

The API validates and returns every problem at once, rather than one per attempt:

curl -X POST https://api.systemonemodels.tech/v1/manifest/validate \
  -H 'content-type: text/plain' \
  --data-binary @systemone.yaml

{
  "valid": false,
  "issues": [
    { "path": "license", "message": "unknown license 'apache2'", "kind": "schema" },
    { "path": "capabilities", "message": "List should have at least 1 item", "kind": "schema" }
  ]
}

The machine-readable schema is at /v1/manifest/spec.