Developer infrastructureInspectable controls · production APIs

Build governed AI systems on an inspectable control layer.

One entry point for Maha's production APIs, SDKs, operational boundaries, benchmarks, and security model.

Use only the capability your workflow needs: govern agent tools, compile bounded context, evaluate evidence retention, run bounded optimization jobs, or preflight consequential documents.

Production capability map

Choose the control your workflow is missing.

Agent governance

Enterprise MCP Gateway

Register tenant-owned MCP upstreams, discover tools, enforce method and tool allowlists, bound execution, and retain metadata-only audit evidence.

Context engineering

Context Compiler

Rank and deduplicate task-relevant passages into a declared context budget while preserving source IDs, passage hashes, and measurable coverage.

Evaluation

Context Pack Evaluator

Declare evidence that must survive compilation, then measure exact retention alongside token, byte, coverage, and duplicate-removal metrics.

Evidence assurance

MPS Preflight

Screen AI-assisted documents before review and produce an explicit record of claims, evidence status, uncertainty, and remaining verification work.

SDKs and integration contracts

Integrate without adopting another runtime.

x402 Observatory

Open factual checks for x402 v2 protocol and Bazaar discovery correctness, without subjective trust scoring.

Inspect public conformance

x402 Buyer Policy

Open pre-signing budgets, allowlists, approvals, replay controls, and settlement verification for autonomous buyers.

Apply buyer payment controls

Compatibility Pack contract

Exact schemas, fixed price, sample evidence report, limitations, and refund behavior for the bounded A2A + MCP machine product.

Inspect the published contract

Bounded machine decisions

Carry identity, entitlement, quota, and evidence boundaries into each request.

These operational guides connect Maha's implemented contracts to conceptual lenses from the books. The book links frame the problem; they are never treated as evidence that a machine control exists or works.

Published operational evidence

Benchmarks state the boundary, not just the number.

GPU measurements are seven-run warm solver baselines on an NVIDIA A10G. They exclude HTTP latency and container cold starts, and heuristic results do not claim global optimality.

Tensor-network QUBO / Ising

80.840 ms

Warm p95 at 256 variables and bond dimension 256; reviewed promotion threshold ≤150 ms.

Review scope and limits ↗

Weighted SE(3) registration

108.604 ms

Warm p95 at 16,384 paired points; reviewed promotion threshold ≤200 ms.

Review scope and limits ↗

MCRB-1 evidence retention

60.4%

Complete human evidence-set retention at 74.4% mean token reduction across 250 independently annotated QASPER questions.

Review scope and limits ↗

Context compiler recipe

74.34%

Executable four-chapter workload with 100% source participation and economics stated against the $0.001 x402 fee.

Review scope and limits ↗

Security and data boundary

Minimize what crosses and what remains.

Security claims are capability-specific. Review the API contract and deployment boundary before sending production, regulated, personal, or confidential data.

  • Tenant isolation: API credentials resolve to tenant-scoped resources; one tenant cannot enumerate another tenant's MCP servers.
  • Bounded execution: Request sizes, methods, tools, rates, timeouts, failure thresholds, and circuit cooldowns are server-validated.
  • Credential handling: Supported MCP upstream secrets are encrypted at rest and never returned by listing or discovery APIs.
  • Data minimization: Context compilation is transient; gateway audit records exclude tool arguments and upstream response bodies.

Evidence-grounded guides

Start from measured behavior and runnable controls.