montepath 0.1.2

Alpha agent-native Monte Carlo runtime with Rust CPU execution, Python APIs, and MCP tools

MontePath

montepath is an alpha-stage, agent-native Monte Carlo runtime with Rust-backed CPU execution, Python APIs, benchmark artifacts, and MCP-compatible agent tools.

Current Stage

This repository is ready for controlled public alpha use. It is suitable for agent-native Monte Carlo experiments, reproducible CPU workloads, and benchmark-audited method comparison. It is not yet a general production replacement for mature scientific or quantitative stacks.

• architecture and design docs are in docs/
• project-level agent instructions are in AGENTS.md
• Codex project skills are in ./.codex/skills/
• core engineering rules are in docs/repository-rules.md
• roadmap is in roadmap.md
• Rust workspace scaffolding is in crates/
• public alpha positioning is in docs/public-alpha.md
• uv and MCP install guidance is in docs/uv-and-agent-install.md

Core Principles

• architecture and design docs are source of truth
• roadmap is a living artifact and must be maintained
• test-driven development is the default workflow
• production-grade quality bar from day one
• fast and lightweight implementation choices
• benchmark claims must stay honest about workload scope and methodology

Workspace Layout

• crates/mc-schema: schema types, diagnostics, compatibility, and validation
• crates/mc-core: planner interfaces, backend contract, CPU runtime, and execution planning
• crates/mc-bench: benchmark harness and benchmark result schema
• python/montepath: Python-first configs, pricing helpers, benchmark helpers, and method recommendations
• docs/site: quickstarts, API reference, benchmark interpretation, and migration notes

Python Quickstart

With uv, after the package is published:

uv pip install montepath

Inside a uv project:

uv add montepath

From a checkout:

uv pip install -e .

from montepath import (
    EuropeanCallConfig,
    native_runtime_status,
    price_european_call,
    price_european_call_greeks,
)

cfg = EuropeanCallConfig(n_paths=20_000, n_steps=64, seed=42)
price = price_european_call(cfg)
greeks = price_european_call_greeks(cfg)

print(native_runtime_status().as_dict())
print(price.explain())
print(price.manifest)
print(greeks.greeks)

The Python helpers are dependency-free reference UX helpers. Timing claims remain tied to Rust benchmark artifacts. Use native_runtime_status() to check whether compiled Rust execution is installed in the active Python environment. Metal-enabled macOS native builds also expose strict price_*_metal bridge functions for the current GBM option family; use backend_capabilities() or select_backend(..., backend="apple_metal") to inspect availability before running them.

The Python package also exposes stable native-bridge configs and result surfaces for the Rust-only workload families: lookback, basket, American put, Bermudan put, Heston, Merton jump-diffusion, Gaussian UQ moments, arithmetic Asian MLMC/MLQMC, and European parameter sweeps. Those helpers validate configuration locally, then require the installed montepath._native module with the matching function; they do not silently fall back to slow or unsupported behavior.

Agent And MCP Usage

Installed distributions include the montepath-mcp console entry point.

After PyPI publication, agents can launch it through uvx:

uvx --from montepath montepath-mcp

From a local checkout:

uv run montepath-mcp

See docs/uv-and-agent-install.md and docs/agent-tooling.md for MCP client configuration, schemas, execution limits, and failure policy.

Agent Tool Example

from montepath import (
    agent_capabilities,
    agent_execute,
    agent_plan,
    agent_production_check,
    agent_tool_manifest,
    agent_validation_report,
)

print(agent_tool_manifest()["schema_version"])
print(agent_capabilities({})["result"]["native_runtime"])
print(agent_validation_report({})["result"]["schema_version"])

plan = agent_plan({
    "workload": "european_call",
    "config": {"n_paths": 10_000, "n_steps": 64, "seed": 42}
})

run = agent_execute({
    "workload": "european_call",
    "config": {"n_paths": 10_000, "n_steps": 64, "seed": 42}
})

preflight = agent_production_check({
    "workload": "european_call",
    "config": {"n_paths": 10_000, "n_steps": 64, "seed": 42},
    "backend": "auto"
})

print(plan["plan"])
print(preflight["result"]["validation"]["selection"])
print(run["manifest"])

Expressive API Example

use mc_core::{EuropeanCallMethod, EuropeanCallPricer, SamplingMethod};

let result = EuropeanCallPricer::new()
    .s0(100.0)
    .strike(100.0)
    .rate(0.03)
    .volatility(0.2)
    .maturity(1.0)
    .paths(100_000)
    .steps(64)
    .seed(42)
    .method(EuropeanCallMethod::StepwisePaths)
    .sampling(SamplingMethod::LatinHypercube)
    .control_variate()
    .price();

use mc_core::{EuropeanCallPricer, Greek, GreekEstimator};

let greeks = EuropeanCallPricer::new()
    .paths(100_000)
    .seed(42)
    .greeks(GreekEstimator::Pathwise);

let delta = greeks.estimate(Greek::Delta).unwrap().value;

Current Runtime Surface

The CPU runtime now exposes:

• a fair step-wise path benchmark path
• a specialized terminal-distribution fast path
• variance-reduction techniques including antithetic variates and control variates
• separate sampling selection via SamplingMethod::{Pseudorandom, RandomizedHalton, LatinHypercube, ScrambledSobol, ScrambledSobolBrownianBridge}
• arithmetic Asian multilevel Monte Carlo via ArithmeticAsianMlmcConfig and arithmetic_asian_call_price_mlmc_cpu()
• structured Greek reports via GreekReport, with bump-and-revalue support across current CPU workloads and European pathwise / likelihood-ratio estimators where valid
• a machine-readable method capability catalog via monte_carlo_method_capabilities()
• method recommendation via recommend_method() in Rust and montepath.recommend_method() in Python
• multiple workload families:
  • European call
  • arithmetic Asian call
  • down-and-out call
  • two-asset basket call
  • fixed-strike lookback call
  • American put LSM
  • Bermudan put LSM
  • Heston European call

The GPU backend layer now exposes:

• truthful delegated fallback semantics for unsupported features
• staged native CUDA and Metal artifact metadata and kernel ABI contracts
• real native Apple Metal execution for the current GBM step-wise workload family on supported Macs

The current native Apple Metal path supports:

• European call: Standard, Antithetic, ControlVariate
• arithmetic Asian call: Standard, ControlVariate
• down-and-out call: Standard, ControlVariate

Structured-sampling requests currently run truthfully on CPU reference paths rather than being silently approximated on Metal.

Native Feature Gates

mc-core exposes host-side native staging gates:

• cuda-native
• metal-native

These validate native backend boundaries, kernel metadata, and toolchain probing without requiring local GPU hardware.

The CUDA path includes a staged kernel source at:

• crates/mc-core/src/backend/kernels/european_call_stepwise_v1.cu

When cuda-native is enabled and nvcc is available, the backend attempts PTX compilation during artifact staging and records the result in native artifact metadata. Execution still falls back to the CPU reference path until native CUDA launch support lands.

The Metal path includes a staged shader source at:

• crates/mc-core/src/backend/kernels/european_call_stepwise_v1.metal

When metal-native is enabled, the backend can compile and execute the current staged Metal workload family in-process on supported macOS hosts.

Running Tests

cargo test

cargo test -p mc-core --features cuda-native

cargo test -p mc-core --features metal-native

Running Benchmarks

Fast smoke profile for local checks and benchmark-gate tests:

cargo run -p mc-bench -- --profile compact

Full profile for competitiveness artifacts:

cargo run -p mc-bench -- --output benchmarks/latest-results.json

cargo run -p mc-bench --release --features metal-native -- --output benchmarks/release-results.json

The compact profile reports a representative subset and does not overwrite benchmarks/improvement-plan.md. The full profile remains the source for tracked performance claims.

Benchmark Gates

Benchmark thresholds are documented in docs/benchmark-gates.md. Competitive benchmark policy is documented in docs/competitive-benchmark-policy.md. User-experience research and UX implementation plan is in docs/user-friendliness-research.md. Agent integration guidance is in docs/agent-integration-plan.md. Public function inventory is in docs/function-catalog.md. Technique roadmap is in docs/simulation-techniques.md. GPU testing strategy is in docs/gpu-testing-strategy.md. Market landscape notes are in docs/market-landscape.md.

Current Results

From the latest release benchmark run:

• fair step-wise Rust CPU European path: 12.416 ms
• step-wise Rust antithetic path: 27.450 ms
• step-wise Rust control-variate path: 13.233 ms
• arithmetic Asian Rust CPU path: 19.932 ms
• arithmetic Asian Rust CPU control-variate path: 18.927 ms
• arithmetic Asian Rust CPU MLMC path: 4.544 ms
• arithmetic Asian Rust CPU MLQMC path: 6.064 ms
• randomized Halton Rust CPU European path: 81.870 ms
• Latin hypercube Rust CPU European path: 67.374 ms
• scrambled Sobol Rust CPU European path: 83.418 ms
• scrambled Sobol Brownian bridge Rust CPU European path: 101.373 ms
• down-and-out Rust CPU path: 16.912 ms
• down-and-out Rust CPU control-variate path: 17.067 ms
• fixed-strike lookback Rust CPU path: 16.531 ms
• American put LSM Rust CPU path: 263.685 ms, binomial-reference abs error 0.000323
• Bermudan put LSM Rust CPU path: 138.789 ms, binomial-reference abs error 0.019098
• Heston European Rust CPU path: 25.977 ms
• basket Rust CPU pseudorandom path: 3.970 ms
• basket Rust CPU Latin hypercube path: 3.943 ms
• basket Rust CPU scrambled Sobol path: 6.976 ms
• European bump-and-revalue Delta error vs Black-Scholes: 0.000126 in 3.095 ms
• European pathwise Delta error vs Black-Scholes: 0.000281 in 1.496 ms
• European likelihood-ratio Delta error vs Black-Scholes: 0.002631 in 1.382 ms
• all-current-workload bump Greek breadth: 26 estimates in 193.818 ms
• specialized Rust terminal-distribution fast path: 0.610 ms
• native Metal European path on macOS: 1.472 ms
• native Metal European antithetic path on macOS: 0.985 ms
• native Metal European control-variate path on macOS: 0.987 ms
• native Metal arithmetic Asian path on macOS: 0.696 ms
• native Metal arithmetic Asian control-variate path on macOS: 0.703 ms
• native Metal down-and-out path on macOS: 0.789 ms
• native Metal down-and-out control-variate path on macOS: 0.752 ms
• NumPy fair CPU baseline: 76.321 ms
• Numba fair CPU baseline: 222.326 ms
• measured planner choice accuracy vs local backend winners: 87.5%

Current QMC generation scoreboard from the same release run:

• Rust scrambled Sobol normal generation: 78.868 ms
• SciPy scrambled Sobol normal generation: 115.809 ms
• Rust randomized Halton normal generation: 57.871 ms
• SciPy randomized Halton normal generation: 141.066 ms
• Rust Latin hypercube normal generation: 40.679 ms
• SciPy Latin hypercube normal generation: 195.462 ms

Current QMC pricing-quality and UQ scoreboard from the same release run:

• European scrambled Sobol stderr ratio vs pseudorandom: 1.000
• arithmetic Asian Latin hypercube stderr ratio vs pseudorandom: 1.003
• down-and-out randomized Halton stderr ratio vs pseudorandom: 0.994
• basket Latin hypercube stderr ratio vs pseudorandom: 0.997
• basket scrambled Sobol stderr ratio vs pseudorandom: 0.996
• European randomized Halton abs-error ratio vs pseudorandom/Black-Scholes: 0.035
• European Latin hypercube abs-error ratio vs pseudorandom/Black-Scholes: 0.021
• European scrambled Sobol abs-error ratio vs pseudorandom/Black-Scholes: 0.129
• European scrambled Sobol Brownian bridge abs-error ratio vs pseudorandom/Black-Scholes: 0.001
• Gaussian UQ pseudorandom abs error vs analytic mean: 0.006344
• Gaussian UQ randomized Halton abs error vs analytic mean: 0.000056
• Gaussian UQ Latin hypercube abs error vs analytic mean: 0.000039
• Gaussian UQ scrambled Sobol abs error vs analytic mean: 0.000043
• Gaussian UQ Latin hypercube abs error vs analytic variance: 0.002484

Current quality ratios from the same release run:

• European control-variate stderr ratio: 0.411
• European antithetic stderr ratio: 0.747
• arithmetic Asian control-variate stderr ratio: 0.607
• arithmetic Asian MLMC stderr ratio: 2.013
• arithmetic Asian MLQMC stderr ratio: 0.418
• arithmetic Asian MLMC abs error vs high-budget standard MC reference: 0.022778
• arithmetic Asian MLQMC abs error vs high-budget standard MC reference: 0.002080
• randomized Halton European control-variate stderr ratio: 0.411
• Latin hypercube European control-variate stderr ratio: 0.410
• down-and-out control-variate stderr ratio: 0.418
• basket randomized Halton stderr ratio vs pseudorandom: 0.996
• basket Latin hypercube stderr ratio vs pseudorandom: 0.997
• basket scrambled Sobol stderr ratio vs pseudorandom: 0.996
• native Metal European control-variate stderr ratio: 0.409
• native Metal arithmetic Asian control-variate stderr ratio: 0.609
• native Metal down-and-out control-variate stderr ratio: 0.417

Honest Status

What we can honestly claim now:

• CPU performance is strong against the available NumPy and Numba baselines on the tracked fair European workload.
• Native Apple Metal is materially faster than our CPU baseline on the tracked European, arithmetic Asian, and down-and-out workloads.
• The library has better breadth than before, with early-exercise LSM, Heston, Merton jump diffusion, two-asset basket, typed European parameter sweeps, Gaussian UQ moment checks, randomized Halton, Latin hypercube, scrambled Sobol, and Brownian-bridge path construction. Direct QMC normal generation beats the available SciPy QMC baselines on the tracked release Sobol, Halton, and Latin-hypercube rows.
• European QMC now has an analytic realized-error scoreboard against Black-Scholes, so accuracy claims for that workload are no longer limited to standard-error ratios.
• MLMC and MLQMC foundations are live for arithmetic Asian calls with per-level estimator metadata, pilot-based allocation tuning, adaptive tolerance planning, replicated Sobol scrambling, and first realized-error calibration rows.

What we should not overclaim yet:

• structured sampling generation is now competitive and pricing overhead is much lower, but full structured-pricing paths still trail the pseudorandom CPU baseline; realized-error wins are currently benchmark evidence for the European analytic-reference case, not a universal guarantee
• MLMC and MLQMC are CPU-reference only, and their tolerance planning is calibrated for the arithmetic Asian path but not yet broadly calibrated across workload families
• native CUDA execution is not implemented yet and is deferred to a later accelerator-focused version
• installed wheels now include a Rust-backed montepath._native CPU extension and production preflight helpers for supported CPU-native use; native Metal wheels and CUDA execution remain future work
• planner calibration is improving, but 87.5% measured local accuracy is not broad production-grade backend intelligence yet

Next Steps

• keep native CUDA launch deferred to the later accelerator-focused version
• continue hardware-runner work for native CUDA/Metal validation and accelerator competitor artifacts
• broaden calibrated MLMC/MLQMC and structured-sampling evidence only when backed by release artifacts