A model project should survive changes in UI, trainer, and compute.

Portable project manifest

The manifest is intended to record the minimum information required to reproduce and evaluate an experiment without embedding datasets, model weights, checkpoints, or secrets in the configuration.

schema_version: 1
project_id: invoice-extraction
task: structured-output

dataset:
  revision: ds_003
  provenance: recorded
  splits: [train, validation, holdout, stress]

model:
  source: open-model-repository
  revision: pinned

experiment:
  backend: mlx
  method: lora
  seed: recorded

evaluation:
  baseline: required
  holdout: protected

package:
  formats: [mlx, lora-adapter]
  include: [experiment-card, model-card, provenance]

What belongs in the project contract

• Stable project identity and schema version.
• Dataset source, revision, split, provenance, license, and integrity references.
• Base-model source, immutable revision, tokenizer, chat template, and license metadata.
• Trainer backend, method, parameters, seed, environment, and target compute description.
• Baseline, validation, protected holdout, stress cases, task metrics, and regression thresholds.
• Checkpoint, package, experiment-card, model-card, and known-limitation references.

What does not belong

Provider credentials, access tokens, signing keys, raw secret values, private machine paths that cannot be remapped, or mutable “latest” model references.

Future version and compatibility API

The first planned tensor.cx/v1 surface is deliberately small: version and compatibility information required for the native Studio, local engine, CLI, and later cloud workers to decide whether they can safely exchange a project or job manifest.

{
  "api_version": "v1",
  "studio_version": "development",
  "engine_protocol": "development",
  "compatible": false,
  "status": "not-publicly-available"
}

Contract principles

• Versioned: schema and protocol versions are explicit; silent reinterpretation is not allowed.
• Portable: local, Multi-Mac, managed cloud, and BYOC should consume the same logical experiment.
• Content-addressed: important model, dataset, and artifact inputs should be pinned and verifiable.
• Secret-free: credentials are resolved from the execution environment, never copied into portable manifests.
• Evidence-linked: a package refers back to the eval, hardware, provenance, and configuration that justified it.
• Fail explicit: an unknown or incompatible version becomes a visible incompatibility, not a best-effort run.