Commerce-1

A 27B, self-hosted decision model for commerce agents. Give Commerce-1 one structured state and ordered Choice, Noul, or Score questions; it returns calibrated probability distributions and typed answers—not prose—through an offline, content-verified runtime.

Highlights

• Typed decisions, not generated prose. Choice selection, binary probability (Noul), and ordinal scoring share one request contract.
• Many questions, one state. Score several agent actions against the same structured context through one request and one shared-state contract.
• Probabilities included. Route uncertain actions to a person, policy engine, or a more expensive model instead of treating every answer equally.
• Offline and self-hosted. The native runtime verifies every package byte before loading and rejects network fallback.
• Evidence before claims. Evaluation and runtime qualification are bound to immutable artifacts; missing evidence remains unknown.

At a glance

Quick start

Use the immutable 40-character source and model revisions from the verified release receipts. Floating branches and model revisions are not part of the qualified path.

Commerce-1's BF16 weights are approximately 54 GiB. Allow additional GPU memory for activations and enough disk and RAM to stage the checkpoint. CPU, macOS, and 24/48 GiB GPU paths are not qualified by this release.

export COMMERCE_ONE_SOURCE_REVISION=<verified-github-revision>
export COMMERCE_ONE_MODEL_REVISION=<verified-huggingface-revision>

git clone https://github.com/infercrane/commerce-1.git
cd commerce-1
git checkout --detach "$COMMERCE_ONE_SOURCE_REVISION"
test "$(git rev-parse HEAD)" = "$COMMERCE_ONE_SOURCE_REVISION"

uv sync --frozen --python 3.12 --extra v3-runtime
uv run --frozen --python 3.12 --extra v3-runtime hf download infercrane/Commerce-1 \
  --revision "$COMMERCE_ONE_MODEL_REVISION" \
  --local-dir ../commerce-1-model

export COMMERCE_ONE_MODEL_DIR="$(cd ../commerce-1-model && pwd)"
uv run --frozen --python 3.12 --extra v3-runtime commerce-1-v3-serve \
  --model-dir "$COMMERCE_ONE_MODEL_DIR" \
  --base-dir "$COMMERCE_ONE_MODEL_DIR" \
  --offline \
  --host 127.0.0.1 \
  --port 8000

Make a decision

curl --fail-with-body --silent --show-error \
  -H 'content-type: application/json' \
  --data '{
    "model":"infercrane/commerce-1",
    "state":{"cart_total":117,"approved_total":94},
    "questions":{
      "route":{
        "type":"choice",
        "instructions":"Choose the safest next action.",
        "criteria":{
          "continue":"Continue.",
          "review":"Ask for review.",
          "block":"Stop."
        }
      },
      "review":{
        "type":"noul",
        "instructions":"Is review required?"
      },
      "risk":{
        "type":"score",
        "instructions":"Score commerce risk.",
        "criteria":["low","guarded","high","block"]
      }
    }
  }' \
  http://127.0.0.1:8000/v1/systemone

Probabilities help callers set review thresholds, but they are not guarantees of correctness. Recalibrate on representative traffic before using confidence for consequential automation.

Measured evidence

The Decision Index result is not an official 0.3 Full Score or leaderboard rank. Those remain pending maintainer evaluation against the immutable public release. Exact content-free bindings live in release/commerce-1-evidence.json; the public model package includes QUALIFICATION.json for the runtime measurement.

Validate the checked-in evidence snapshot:

uv run --frozen --python 3.12 --extra test \
  python tools/v3_release_evidence.py

Where it fits

Commerce-1 is designed for structured decisions inside commerce-agent workflows, including:

• proceed, review, or block gates;
• tool and workflow routing;
• approval-policy drift checks;
• product, fulfillment, and support action selection;
• calibrated risk or priority scoring.

It is not a chat model, search engine, live-fact retriever, payment processor, or autonomous authority for refunds, disputes, fraud, credit, safety, or legal decisions. A benchmark result does not establish fitness for a production policy. Validate the model on your own distribution and keep consequential actions behind explicit controls.

How it works

Commerce-1 starts from perplexity-ai/pplx-decider-v1.1-27b and uses a noncausal full-attention decision backbone with last-position pooling. A decision readout scores all supplied answer options directly, without decoding an explanation token by token. Choice-order augmentation and per-primitive held-out temperature calibration make the typed probabilities usable by downstream policy code.

The runtime inventories and hashes every required package file before model load. It rejects symlinks, undeclared files, mutable repository identities, changed checkpoint bytes, calibration drift, source-closure drift, unknown manifest fields, and network fallback. See docs/V3_RUNTIME.md for the exact contract.

The same scorer is available at POST /api/alpha/decisions. A strict OpenAI-compatible JSON bridge is available at POST /v1/chat/completions for clients that cannot call the native route. Commerce-1 does not generate text: the user message contains one JSON decision object and the assistant content contains the typed answers as JSON.

Serving boundary

Loopback is the default. Non-loopback serving requires both --allow-non-loopback and --api-key-file. The key file must be a regular, non-symlink file owned by the process user with mode 0600; token values are not accepted in command arguments or environment variables.

Useful routes:

• GET /healthz — process liveness only.
• GET /readyz — readiness and immutable loaded identities.
• GET /v1/model — content-free model, runtime, and evidence identity.
• GET /v1/models — OpenAI-style discovery with the same identities.

Responses report measured input tokens and zero generated tokens. Cost remains null until backed by real metering. See docs/API.md.

Development

uv sync --frozen --python 3.12 --extra test
uv run --frozen --python 3.12 --extra test pytest -q
uv run --frozen --python 3.12 --extra test ruff check .
uv run --frozen --python 3.12 --extra test ruff format --check .

Publication is private-first and exact-byte verified. The operator contracts are documented in release/V3_PUBLICATION.md, release/MODAL_HUGGINGFACE_PUBLICATION.md, and release/GITHUB_PUBLICATION.md.

License

The source code in this repository is licensed under Apache-2.0. Model and training-source notices are documented separately in THIRD_PARTY.md. The source-code license does not override upstream model or dataset terms.