Capability Model & Limits

This page lists what Valence does today and what it does not do on purpose. Every item is traced to the code. For the line between advisory and authoritative output, see Trust, Provenance & Authority Boundaries.

What works today

• A curated registry of 14 Domains and 46 Meshes (~112.4 MB), with 27 of the Meshes in 5 Collections. One canonical manifest.toml lists them. Mesh bytes come on demand from a Hugging Face mirror.
• Deterministic registry tooling: validation, structural gates, diff, normalization (normalized-v1), content-UID identity, and byte-deterministic synthetic generation ([gen]).
• A change-controlled review layer: a deterministic scope verdict against an approved region, CRS fail-closed, a tamper-evident ruleset (behavioral golden battery), ed25519 DSSE attestation, and an offline, deterministic review-bundle verifier.
• A planning workspace that imports precomputed selection artifacts and ranks them deterministically. Every ranking component stays visible, and one human approval gate decides.
• Candidate generation through the proprietary engine, for the owner only. It runs only when the engine is installed through the owner-only path and every material confirmed constraint maps through the deterministic compiler. Without the engine, generation fails closed. The public install never includes it.
• For a generated artifact, the adapter attests the link between recipe and output; no cryptographic binding exists. Valence recomputes the compiled recipe digest and compares it to the digest the artifact echoes back.
• Mandatory provenance for real-world meshes: validate fails closed without it.
• Byte-for-byte offline reproduction of normalization, scope verdicts, bundle digests and signatures.
• Two authoritative human gates: contribution approval and region-proposal approval. Each needs an actor and a reason, and each emits an audit event. Candidate proposals are advisory. Only an approved region crosses the authority boundary and supports enforcement-grade scope verdicts.
• Ranking that orders candidates for review. Ranking makes no claim of scientific optimality, simulation accuracy, or automatic selection.

What is intentionally unsupported

• No runtime LLM. Nothing in the package, the demos, or the region-proposal subsystem calls a model at run time. Planning is deterministic.
• No runtime network for the canonical demos. Each canonical demo runs offline. It reads committed synthetic fixtures and recomputes digests locally.
• The base wheel ships no mesh files. Loading a mesh needs the [hf] extra, which fetches bytes from the Hugging Face mirror on demand. A local checkout also works. The base install pulls chilmesh, which brings numpy, scipy and matplotlib.
• The synthetic-mesh generator needs [gen]. Without it, the CLI prints an install message and exits non-zero. It does not crash with a traceback.
• .2dm boundary extraction is not implemented. Boundary metrics return None with a warning. .14 (ADCIRC fort.14) is fully supported. A boundary-only file is still valid to register (kind: "boundary").
• The validation worker runs off Vercel. Vercel Hobby functions stop at about 10 s. Browse and upload fit, because byte transfers skip the backend. The validation worker takes longer by design, so it runs in CI or locally.
• Scope verdicts are advisory and never block. The L1 and L2 scope verdicts and the signed attestation appear as advisories. Human approval is the single gate. Approval is still possible when the verdict was skipped. Passing the automated gates does not publish a mesh.
• Supported interpreters: Python 3.10 to 3.12, Node 20 or later. Newer Python versions are untested.
• Postgres-fidelity backend tests skip without a database. The default backend suite runs on in-memory SQLite.
• Unsupported material constraints block generation. Generation fails closed on them and never ignores one without notice. Valence has no natural-language-to-mesh capability.