Research slower than the product, but further ahead
Our research runs along four axes: execution verification, distributed intelligence, computer use and domain data. MANPASIK is the independent future-research track among them.
Problem → prototype → validation → product
Research topics start from failures observed in the field. A structure that cannot explain why the agent was wrong does not become a product.
Every study runs with a reproducible experiment log, and unverified claims are not published on this site.
Problem definition
Find where automation breaks, through field interviews and failure logs.
Prototype
Build an execution that works end to end on a small scope first.
Validation
Test with limited users on real work and classify failure types.
Publication
Publish method and limits together. If there is no result, we say so.
Technical development goals
Consolidate the execution engine
Merge the plan–execute–verify loop into one engine and classify failure types.
Unified verification framework
Standardize per-task verification rules and rollback procedures.
Own execution engine
Build an execution and verification layer with less dependence on external models.
Mid- and long-term directions
Verification-first agents
How an agent can verify its own execution: failure-mode taxonomy and recovery strategies for the plan–execute–verify loop.
Distributed intelligence
Division of roles, cross-verification and task routing — the core question of the MANPASIK track.
General computer use
UI action policies that stay stable through site changes and unexpected states.
Domain data infrastructure
Collection structures that preserve the quality and lineage of public, hospitality and enterprise data.
MANPASIK — Distributed Intelligence Core
MANPASIK tests a hypothesis: that several intelligences with different roles, verifying each other, produce more reliable outcomes than one very large model. A planner, an executor, a verifier and a memory keeper falsify one another's work.
Everything on this page is an idea and a design hypothesis. It is not a finished product or validated performance, and it is neither quantum computing nor a commercial distributed system.
What this is not
- Not a quantum computer or quantum advantage technology
- Not a commercial product or service
- No validated performance figures or client cases
Research questions
How is reliability improved over a single model?
By designing role separation and cross-falsification, and by giving the verifier a different context from the executor. There are no experimental results yet.
How are cost and latency controlled?
Task routing and local or edge inference placement are hypotheses under simulation design.
When does it become a product?
Undetermined. It will be reviewed separately if research results meet productization conditions.
Are results published?
Not yet. There are design documents and a simulation plan; results will be published when consolidated.
Technical whitepaper (in preparation)
We are preparing a technical whitepaper on agent architecture and verification. It is a draft, and no publication date is fixed.
Problem definition
Where automation fails, and the cost of verification.
Architecture
The six-layer structure and the separation of plan, execute and verify.
Verification method
Per-task verification rules and failure classification.
Cases
Hospitality, public data and back-office prototypes.
Limits and next steps
Current implementation scope and open problems.
We do not present unverified results as achievements. Papers, patents, certifications and awards are published only when held; none are listed today.