1 / 6 · Question
How did past vibration change?
Consider six days of history for an anonymous rotating asset. The aim is to compare possible explanations, not to intervene in a control system.
Models change.Experience endures.Decisions remain human.
An operating layer aiming to bring local open-weight models, specialist agents and lasting knowledge together under human oversight.
An agent operating layer on Linux. Not a new kernel or distribution.
More than an answer.A way of working.
Read → analyze → advise. Humans decide.
Work on your own infrastructure.
Local operation starts in a workspace where you define tool and data boundaries. On-prem and rented self-host are different placements.
Let experience outlive the model.
Sources, reviews and canonical records endure. Replacing a model should not erase a verified history.
Teachers produce. Workers execute.
Large models produce candidates for code, skills and tests. Local models handle daily work; candidates gain no authority before evaluation.
Ground explanations in evidence.
A result comes with its source and calculation method. Competing hypotheses, gaps and uncertainty belong in the answer too.
Keep authority human.
Models do not grant permission. Software enforces policy and acceptance; humans make the final decision and promote versions.
Make skills reusable.
A skill is a method with defined inputs, outputs, versions and tests. It is more than a single successful conversation.
Keep learning controlled.
Experience starts as a candidate. Human review and independent evaluation precede versioning; production models never change silently.
Expand when it serves a purpose.
Each new specialist must address a measured need. Agent count alone is not a measure of success.
A good answer is a beginning.Verified knowledge is a traceable record.
One core. A family of specialists.
UMAY OS Core will be created through a controlled fork of AOS. Its first specialist, UMAY Scientist, will preserve the aserdargun/ai-scientist foundation. This diagram describes the target architecture, not completed runtime acceptance.
Operation
On-prem · local operation target
A model is a compute resource. An agent is an execution role with defined boundaries.
Core is the single GUI owner through typed execution, policy and input ownership. Scientist conducts scientific investigation on permitted data. Model/deployment/adapter registries, knowledge/skills and GPU management will build on AOS.
Teachers propose. Models do not grant authority.
- 01Large model · provider / self-host
- 02Code, skill, test and training candidate
- 03Harness · safety and evaluation
- 04Human review and promotion
- 05Versioned registry · rollback
There is no direct authority path from the teacher line to the live GUI. Sensitive inputs are never sent automatically to a provider. Daily local operation is intended to continue when teacher access is unavailable.
Contracts before integration. Evidence before promotion.
- 01Pinned sources · controlled fork
- 02Core–Scientist contract
- 03Local synthetic GUI acceptance
- 04Independent evaluation
- 05Human promotion · rollback
Model and agent replacement requires capabilities, versioned contracts and independent evaluation. The Linux profile is an installation target, not a verified UMAY distribution. SWAPP will be accepted using local frontend/backend and synthetic or permitted examples; actual institutional acceptance comes later.
First specialist: Scientist.
Between questions, evidence and calculation.
The original AI-Scientist’s Director, local model provider, sandbox, computation/evaluation and experiment ledger will be retained. Scientist research S1/S2 profiles differ from the Core decision protocol.
Conceptual example · synthetic data
2 / 6 · Data
Permitted data. Explicit provenance.
Six entirely synthetic samples. Time: 1–6 September 2026, daily at 12:00 UTC. Unit: mm/s RMS. These values do not come from a real facility or device.
3 / 6 · Calculation
Deterministic calculation. Visible method.
vibration-window-mean v1 compares arithmetic means of the first and last three samples. Percentage change = (last mean / first mean − 1) × 100. The same data produces the same result.
4 / 6 · Hypotheses
One finding. Several explanations.
Changes in load or speed, measurement conditions, or mechanical condition could explain the difference. Vibration growth alone cannot distinguish them or diagnose a fault.
5 / 6 · Answer
A sourced Advisory Packet.
The packet carries the question, source version, method, supporting and contradicting evidence, alternatives and missing data. Uncertainty remains visible.
6 / 6 · Review
Lasting Investigation. Human decision.
A human reviews the evidence and identifies the next data request. An asset-linked Investigation is intended to retain sources and review history. This page runs no record service.
2026 · 12:00 UTC · vibration-window-mean v1
Deterministic calculation. Visible method.
- First three days
- 2.10 mm/s
- Last three days
- 3.03 mm/s
- Change
- +44.4%
This change is not a fault diagnosis.
Alternative explanation
Load or measurement conditions may have changed. Mechanical condition cannot yet be distinguished.
Missing evidence
Speed, load, sensor calibration and maintenance records are absent.
Chart data table
| Time (UTC) | Vibration (mm/s RMS) |
|---|---|
| 2026-09-01T12:00:00Z | 2.0 |
| 2026-09-02T12:00:00Z | 2.2 |
| 2026-09-03T12:00:00Z | 2.1 |
| 2026-09-04T12:00:00Z | 2.5 |
| 2026-09-05T12:00:00Z | 3.2 |
| 2026-09-06T12:00:00Z | 3.4 |
Read and advise. Writing to control systems, setpoints, start/stop and alarm-state changes are out of scope.
Models change.Knowledge keeps its history.
What endures is more than an answer: its source, review and version.
RAG
Knowledge tied to sources.
Documents are retrieved with their provenance, access boundary and version. Producing an answer does not turn it into verified knowledge.
Skill
A versioned, testable method.
Inputs, outputs, authority boundaries and tests are recorded together. The method can be reused under control in a new task.
Training data
Evaluated learning examples.
Fine-tuning is considered only for supported models and a measured behavioral need. RAG and skill records are not automatically training data.
- Candidate
- Human review
- Verified record
- Evaluation
- Version / rollback
Canonical records and review history endure. Embeddings, tokenizer outputs and LoRA adapters are not directly portable to every model.
Where do we stand today?
Existing in public sources
Two original public repositories provide existing implementation foundations. Their pins are not presented as runtime test results on this site.
UMAY architecture decision
- AOS-based Core and Scientist as the first specialist.
- Separation of local operation and the teacher line.
- Sourced Advisory Packet, lasting Investigation and human promotion.
Integration and acceptance pending
- Core–Scientist integration.
- Broker patch compatibility with the baseline.
- SWAPP local GUI acceptance; institutional acceptance later.
- Independent evaluation and continuous learning.
- 01
Integration
Typed contracts, policy and execution ownership.
- 02
Local GUI acceptance
Local setup with synthetic or permitted examples.
- 03
Independent evaluation
Repeatable acceptance evidence separate from source inventories.
- 04
Next specialists
Expansion based on measured need and value.
A source inventory is not a UMAY test result. Public access does not grant a license or commercial usage rights. A UMAY OS project license has not been selected.
Intelligence can change.Responsibility endures.
We are building the lasting foundation for working with models, beyond the next model itself.
Explore development on GitHub