mcpField
B
B2IGE Verify

B2IGE Verify

Local
github.com/b2ige2-a11yDeveloper Tools0Updated 6w ago

About B2IGE Verify

B2IGE Verify is local, deterministic verification infrastructure for AI-written software, aimed at developers and coding agents. It runs programs and records evidence for bounded results, with separate products for hidden tests, behavior comparisons, and local side-effect checks; agents do not grade their own work.

Key features

  1. 01Run programs and record verification evidence
  2. 02Return bounded PASS, FAIL, and other results
  3. 03BlindTest runs sealed hidden tests
  4. 04BehaviorSeal compares reference and candidate behavior
  5. 05SideEffect Proof checks committed local SQLite changes

Use cases

  1. 01Check AI-generated software against hidden tests
  2. 02Compare program behavior before and after a change
  3. 03Verify local data changes under retries and faults

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

No install method listed yet

We haven’t collected a package or remote endpoint for this server. Follow the instructions in the project README.

Repository
Packages
mcpbb2ige-verify-0.1.0.mcpb

Tools

No tool list yet

Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.

Details

From the project README

B2IGE Verify

Don't trust "done". Prove it.

B2IGE Verify is deterministic verification infrastructure for AI-written software. It runs real programs under declared contracts, records evidence, and produces a bounded PASS/FAIL/INCONCLUSIVE/ERROR result without asking the coding agent to grade its own work. BlindTest makes the boundary concrete: hidden tests exercise the produced program without exposing the suite or oracle to the agent.

Latest GitHub Release Candidate checks Apache-2.0 Rust

Latest release

B2IGE Verify v0.3.0 is the current public release. The final release record documents its exact-main qualification, platform boundaries, and public asset set.

PlatformNative archive
macOS Apple Siliconb2ige-0.3.0-aarch64-apple-darwin.tar.gz
macOS Intelb2ige-0.3.0-x86_64-apple-darwin.tar.gz
Linux x86_64b2ige-0.3.0-x86_64-unknown-linux-gnu.tar.gz

Also download SHA256SUMS and verify the matching archive before extracting. See installation details for prerequisites and offline validation.

Quick install

Choose the archive for the host, then verify, extract, and put bin on PATH:

B2IGE_RELEASE_URL=https://github.com/b2ige2-a11y/b2ige-verify/releases/download/v0.3.0
B2IGE_ARCHIVE=b2ige-0.3.0-aarch64-apple-darwin.tar.gz
curl -fL -O "$B2IGE_RELEASE_URL/$B2IGE_ARCHIVE"
curl -fL -O "$B2IGE_RELEASE_URL/SHA256SUMS"

if command -v shasum >/dev/null 2>&1; then
  shasum -a 256 "$B2IGE_ARCHIVE"     # compare with SHA256SUMS
else
  sha256sum "$B2IGE_ARCHIVE"         # compare with SHA256SUMS
fi
tar -xzf "$B2IGE_ARCHIVE"
cd "${B2IGE_ARCHIVE%.tar.gz}"
export PATH="$PWD/bin:$PATH"
b2ige --version

The archive is ready for the adoption flow or the product-specific commands. Docker is required for BlindTest; missing prerequisites remain a readiness or non-PASS result.

30-second BlindTest demo

BlindTest asks: does the produced program actually work? It runs visible checks against a correct and a buggy implementation, then executes sealed hidden cases in an attested Docker boundary. The correct implementation passes, the buggy implementation fails, and the sanitized Agent view is checked for private-value leakage.

From a source checkout or the source archive, with Docker available:

docker pull node:24.18.1-bookworm-slim
scripts/demo-blindtest.sh

The demo performs real execution and records evidence; its output is not canned terminal text. See the BlindTest example for the bounded secrecy boundary and the threat model for its limits.

Three verification products

ProductQuestionWhat it checks
BlindTestDoes it actually work?Sealed hidden tests against an agent-produced target inside the declared Docker boundary.
BehaviorSealDid it change?Trusted reference/candidate behavior under equivalent deterministic experiments.
SideEffect ProofDid it actually happen?Committed local SQLite effects under retries and fault schedules, not inferred attempts.

Easy Adoption

inspect → init → prepare → trust approve → doctor → verify ID

prepare is non-authoritative: it discovers inputs and writes a reviewable draft. Trust approval remains an interactive human/trusted-controller decision. doctor reports readiness, not verification. verify ID resolves the approved identity and produces the real, evidence-backed result. Read the adoption guide for the full boundary.

Measured benchmark

These are committed measurements on the bounded B2IGE Verify Bench v1 corpus, not a claim of exhaustive correctness or a replacement for product evidence.

B2IGE Verify Bench v1 / 31 explicit cases
Measurement (on the benchmark corpus)Observed
Known bugs detected16 / 16
False PASS0
False FAIL0
Verified reproduction16 / 16
Hidden leakage observed0
Agent private-value leakage0
Mutation adequacy: known benchmark mutants2 / 2

All numbers above are on the benchmark corpus only. Bounded testing cannot establish complete correctness.

See the measured snapshot and benchmark methodology.

Platform support

PlatformSupport boundary
macOS Apple SiliconNative release
macOS IntelNative release
Linux x86_64Native release; Docker and full benchmark CI
Windows x64Source/build plus bounded CLI smoke only; not VERIFIED_NATIVE
Linux arm64Deferred

macOS binaries are unsigned and unnotarized. See installation and platform details for Docker, source-build, and archive boundaries.

What the result means

  • PASS / 0: required evidence is complete within the declared executed scope, with no violation observed.
  • FAIL / 1: a violation is supported by evidence.
  • INCONCLUSIVE / 2: required evidence or execution coverage is incomplete.
  • ERROR / 3: the verifier, configuration, isolation setup, or stored evidence failed.

Missing evidence cannot become PASS. Models do not assign verdicts. A result is bounded by the declared experiment, evidence, and platform scope.

Trust boundary and limitations

  • Hashes provide integrity, not publisher authentication.
  • Same-user host access is outside the BlindTest secrecy boundary.
  • Bounded verification is not exhaustive proof.
  • The internal v0.3.0 holdout was AI-operated, not independent human or external validation.
  • Genuine external-user V110-D evidence is deferred; internal or synthetic evidence is not presented as external adoption.

Read the threat model, contracts, and evidence model before relying on a result.

Documentation

Use the documentation index for getting started, product guides, trust and architecture, benchmarks, release history, and developer resources. The false PASS report accepts only synthetic or sanitized details; never upload hidden suites, private canaries, raw human evidence, or credentials.

Contributing

Start with CONTRIBUTING.md, the architecture, and the security policy. Changes to verifier authority, evidence, contracts, or schemas require deliberate review; missing evidence must never be converted into success.

License

B2IGE Verify is licensed under Apache-2.0. No telemetry, hosted service, or paid API is required. Product names and trademarks are covered separately by TRADEMARKS.md.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.