LOCAL RESEARCH · DEFENSIVE · ADVISORY-ONLY
AI agents that
can stop.
Maestro Orchestrator is a research orchestration framework for studying how multiple AI agents can work together safely. It tests fail-closed behavior, audit logs, and final human review through implementation and tests.
- HITL
- Fail-closed
- Tamper-evident
- Local-only
Research & education only. No autonomous external actions.
TRACE 04POLICY LOCKED
PAUSE / REVIEW
APPROVED
01 / OVERVIEW
WHY THIS REPOSITORY EXISTS
Validate behavioral boundaries,
not agent capability.
This repository is a local research and educational test bench for safely reading, reproducing, and reviewing workflows involving multiple agents. It does not control real systems or replace human judgment.
Fail-closed
When uncertainty, risk, inconsistency, or possible external impact is present, the workflow does not continue silently. It stops or moves to human review.
Advisory-only
Detection and remediation proposals support decisions. The system does not automatically modify, commit, push, merge, or deploy.
Traceable records
Audit logs, checkpoints, and integrity checks preserve a state in which events can be reconstructed and reviewed.
02 / SAFETY MODEL
STOP, THEN ASK
When uncertain, move to
human review—not forward.
The design separates the roles of agent, mediator, and orchestrator so that risks or inconsistencies can return to human judgment.
01Observe
Record the state of inputs, outputs, and logs.
02Advise
The Mediator returns advice, but has no authority to execute.
03Gate
Fixed-order checks assess whether the workflow may continue.
04Decide
For ambiguity, risk, or inconsistency, HITL makes the final decision.
RUN
A candidate may proceed to the next check or a permitted operation.
PAUSE_FOR_HITL
A person reviews the situation and decides whether to continue, revise, or stop.
STOPPED
A blocking condition has been reached, so the workflow stops on the safe side.
03 / ARCHITECTURE
INSPECTABLE BY DESIGN
Record why it stopped—
and what permits a restart.
With implementation and corresponding tests readable together, the project examines gate order, reason codes, audit logs, checkpoints, and artifact integrity.
row_index00042layerconsistencydecisionPAUSE_FOR_HITLreason_codeOUTPUT_MISMATCHprev_hash…c8f1
Audit rows are chained in a normalized format so unexpected changes can be detected. This does not prevent changes; it makes them detectable for review.
Pause and resume
Before resuming, validate the run ID, completed artifacts, hashes, and whether HITL confirmation is required.
Reproducible verification
Simulation output is a research artifact. It does not mean production approval or a guarantee of safety.
04 / SIMULATORS
CURRENT RESEARCH LINES
Test the same safety
boundaries across scenarios.
The highest file number is not always the recommended starting point. When verifying behavior, read the corresponding implementation and tests together.
Mediator-based flows
Tests task normalization, advisory-only mediation, fixed-order checks, and escalation to human review.
DocumentationDocument task orchestration
Uses Word, Excel, and PowerPoint tasks to examine checkpoints, resumption, and artifact integrity checks.
DocumentationAudit & tamper evidence
Uses HMAC and hash chaining to make unexpected changes to audit logs and checkpoints detectable.
DocumentationScenario simulations
Explores emergency contracts, incident mediation, and security reviews as local research simulations.
DocumentationSCOPE BOUNDARY
Do not treat advice and
execution as the same thing.
Automatic external sending, uploading, applying changes, creating branches or PRs, commits, pushes, merges, deployments, and real-world control are outside the scope of this research workflow.
This site and the related repository do not guarantee safe execution, safety, integrity, or fitness for any particular real-world use. Decisions or actions affecting external systems require appropriate human review and responsible governance.
Use only in research, education, and authorized defensive contexts.
Security policy (Japanese)