agenttx proof --validator '["npm","test"]' -- codex exec "fix the failing test without weakening it"

• Proof-carrying AI changes — AgentTX derives acceptance from the command, validators, and bound evidence, then commits or rolls back.
• Reproducible MCP failures — ResiliReplay reproduces bounded recovery behavior and turns failures into regression evidence.
• Scientific transitions turned into tests — PhaseProbe preserves bounded qualitative transitions as replayable regression fixtures.

The three-case Proof Gallery is regenerated in CI from real commands. AEB tools are independently usable; include only the evidence engines a task actually needs.

The tools exchange versioned files or invoke released CLIs at process boundaries; none requires the rest of the portfolio. The smoke repository is intentionally not a user-facing product.

See the scored front-door decision, positioning records, and public claim/evidence matrix.

Every command below resolves an immutable public version. Registry publication is stated explicitly; GitHub release assets are used where registry ownership is not configured.

# AgentTX 0.3.0 from npm
npx --yes agenttx@0.3.0 proof -- your-agent-command

# ResiliReplay 0.7.1 from npm
npx --yes resilireplay@0.7.1 mcp demo --keep

# PhaseProbe 0.3.0 from PyPI
python -m pip install "phaseprobe==0.3.0" "pytest==8.4.1"
phaseprobe scan --example logistic
phaseprobe generate-test .phaseprobe/runs/<run-id>/replay.json
phaseprobe verify-evidence tests/generated/logistic_map-pytest-evidence.json
python -m pytest -q tests/generated

# Verifaxis 0.2.0 from its public GitHub release wheel (not PyPI)
python -m pip install "https://github.com/aliengineering-byte/verifaxis/releases/download/v0.2.0/verifaxis-0.2.0-py3-none-any.whl"
verifaxis demo --evidence-output demo-evidence.json
verifaxis verify-evidence demo-evidence.json

# GaugeMesh 0.2.2 immutable GHCR digest comes from its Registry/release metadata
docker run --rm -i ghcr.io/aliengineering-byte/gaugemesh@sha256:537c7e37a4fd97d216f5f8de655363e0647213a59a1d37063899fc843f8b8999 mcp-stdio

Checksums and release notes: GaugeMesh 0.2.2, AgentTX 0.3.0, Verifaxis 0.2.0, and PhaseProbe 0.3.0.

With Node.js 22 or 24:

npx --yes resilireplay@0.7.1 mcp demo --keep

The credential-free local demo runs a clean MCP call, reproduces one deterministic failure, verifies bounded recovery with no duplicate effect, generates and executes a regression, and keeps the evidence under .resilireplay/demo/. See the five-minute MCP guide and the independent Action smoke.

• Agent reliability: GaugeMesh routes while preserving declared constraints; AgentTX contains repository edits; Verifaxis evaluates explicit claims; ResiliReplay tests MCP recovery. Use only the stages the task actually needs.
• Scientific reliability: PhaseProbe brackets a qualitative transition and exports a pytest regression. Verifaxis or ResiliReplay belongs here only when a real verifier or failure-replay boundary is present.

• These tools provide evidence for their documented scopes; they are not security certifications.
• AgentTX does not sandbox a child process or reverse external side effects.
• GaugeMesh is a GitHub Release/GHCR/MCP Registry developer preview and is not on crates.io. Verifaxis remains a research prototype distributed as a GitHub prerelease wheel, not through PyPI.
• AgentTX 0.3.0 is available from npm and GitHub; its Action verifies proof packs but is not an OS sandbox.
• PhaseProbe requires the user to define a scientifically meaningful observable and predicate.
• ResiliReplay requires explicit ownership and approval before exercising remote or effectful tools.

Found a reproducibility failure, confusing quickstart, or missing integration? Open an issue in the affected repository. Successful integrations can be reported through each project's adoption or compatibility issue path. No AEB tool sends telemetry or stars a repository on a user's behalf.