gpd-compute

A no-network, sandboxed numeric oracle exposed as an MCP server. It evaluates agent-authored numeric expressions / evaluators at caller-supplied sample points inside a restricted subprocess (no network, no filesystem writes, bounded time and memory) and returns the per-point values plus a stable content hash of the executed source.

It is a pure evaluation oracle: it has no awareness of any claim or "right answer". It exists so a verification system can compare two independently constructed evaluators numbers-vs-numbers rather than trusting an LLM's prose. It is the optional gpd-compute MCP server bridged by Get Physics Done.

Guarantees

• No network. Sockets are disabled in the sandbox; the evaluator source is AST-validated to forbid imports before it ever runs.
• No API keys, no external services. Everything runs locally.
• Bounded. Hard wall-clock timeout plus CPU / memory / file-size limits.
• Deterministic. The same source + points + precision yields the same values; the content hash is taken over AST-normalized source.

Tools

• evaluate_expression(submission, sample_points, precision) — evaluate and return values + evaluator_hash.
• evaluator_hash(submission) — hash only (no execution), to re-confirm a hash.
• probe_capability(restatement) — cheap check of whether v1 can build an oracle for a quantity class (returns no_oracle_yet for classes like interacting-QFT path integrals or GR-tensor canonicalization).
• describe_runtime() — static self-description (asserts no_network: true).

Run

pip install gpd-compute
gpd-compute            # starts the MCP server over stdio

Scope (v1)

Reliable: closed-form expressions, definite integrals / series, special-function identities (gamma, Bessel, hypergeometric, zeta, elliptic), asymptotic / limiting behavior via a sample-point ladder, and dimensionless ratios. Classes without a soundly constructible numeric evaluator return an explicit no_oracle_yet signal rather than a fabricated number.

License: MIT.