Core and execution
Local models, task policy and tool execution. The source foundation for a company-adapted Core.
An isolated Linux workspace for every employee. SWAPP, AI-Scientist and other agents run here; operations become records, and valuable experience improves local models.
01Manifesto
Turn human experience into a lasting company capability.
A Linux container starts for each employee; the SWAPP web application, AI-Scientist and other agents run within this isolated scope. Sessions, files and access permissions stay separate; company knowledge remains private.
Human, agent and service operations in the workspace are logged with goals, actions, outcomes, failures and corrections. Records persist independently of containers. Only permitted examples with verified value enter learning.
Architecture changes are prepared with powerful large language models: decision rationale, code changes, tests and rollback plans are reviewed together. People approve versioned releases. Daily agents run open-weight models locally.
Recognizing screens is not enough. An agent must demonstrate finding assets, interpreting data and completing permitted workflows with independent outcome verification.
Models do not grant permission. Tools and access are scoped to the employee and task. People decide, accept learning records and approve new versions for use.
AI-Scientist is the first specialist; other local agents can join through shared task, permission, logging and result contracts. Every specialist runs in an isolated scope; scientific experiments retain reproducible data, methods and results.
Selected records become tested skills, sourced RAG knowledge and fine-tuning candidates for local models. LoRA/QLoRA is evaluated only for compatible models. Measure each candidate’s contribution; raw logs are not training data by default.
Track time and resource use for every model call, experiment and training job. Evaluate quality, latency and cost together; unmeasured savings or charges are never treated as zero.
A good answer is a beginning.Verified knowledge is a traceable record.
02Architecture
AOS Core manages an isolated Linux container scope for each employee. Each workspace runs the SWAPP web application, AI-Scientist and other specialist agents. Local model services and persistent records remain in the company environment. Integrated acceptance of this target architecture is still pending.
Local models, task policy and tool execution. The source foundation for a company-adapted Core.
Assets, time series and digital-twin context. The public presentation is accessible; application code and company content are private.
The scientific specialist foundation for clustering, anomaly detection and reproducible experiments.
Company environment · Linux host
Separate sessions, files, network access and resource quotas. The same template is provisioned independently for every employee.
A model is a compute resource. An agent is an execution role with defined boundaries.
An employee workspace consists of one container or service containers within the same isolation scope. SWAPP, Core and specialists use separate process permissions; Scientist experiments run in additional sandboxes. Core is the sole automation owner of each browser session. Records and approved learning artifacts survive container recreation. Scientist’s existing broker remains the single shared-GPU allocation authority; employee containers do not create separate GPU schedulers.
The AOS source baseline uses Mapika/decider-2b. It selects from permitted options using a fresh observation, rather than acting as a free-text chat model. Decision data, training objectives and adapters are specific to S1.
The AOS source baseline uses Ternary Bonsai-2 27B for planning, sourced explanation and recovery. S2 has separate data and adapter profiles; it is distinct from the closed-source teacher line.
Each role can move to current compatible models. Unsloth training uses supported base weights and role-specific data formats. Adapter loading and benefit in the target inference engine require separate validation.
These model names identify the inspected AOS baseline, not permanent product choices. Unsloth compatibility is not assumed for Decider or ternary/GGUF Bonsai. Validate LoRA and 4-bit QLoRA support per model; unsupported models require a separate backend or an evaluated alternative. Scientist research profiles also differ from the Core S1/S2 protocol.
AI-Scientist is the first specialist. Other locally running specialist agents can join through versioned capability registration, task/result contracts, scoped tools and data, resource quotas, cancellation and operation logs. Every specialist follows workspace isolation and human oversight.
The new registered-agent runner in AOS provides a source foundation for this extension. A stable general plugin API, compatibility with every specialist and multi-employee UMAY acceptance are still pending. Open here means extensible; project licensing is a separate decision.
Architecture and contract changes are prepared with powerful large language models in a separate development workspace; rationale, affected components and test evidence are recorded together. Daily local agents do not change the architecture during operation. Company records are not automatically sent to external providers. Model identity and cost are tracked; local operation should continue without teacher access.
First test isolation, record integrity and restarts with two synthetic employee workspaces. Once private SWAPP code is delivered, validate real workflows in the same scope. Measure the learning value of selected records separately; architecture changes are prepared with powerful models and pass tests and human approval.

03Scientist
First task: from digital twin to scientific experiment.
AI-Scientist will run as the first specialist in each employee’s Linux container workspace. Clustering and anomaly experiments on permitted SWAPP digital-twin data will run in separate sandboxes; every trial, failure and result will link to the shared operation trace. Research models will be evaluated separately from Core S1/S2 roles. The new source includes CPU experiments and selected experiment memory; LLM/GPU calls are disabled in the default field-lab profile. Integration into UMAY employee workspaces requires separate acceptance.
Versioned data with asset, time, units and quality. Independent test splits by time and asset.
Explore similar behavior through clustering. Review clusters with domain knowledge; cluster labels are not ground truth.
Compare against simple baselines; measure false alerts and stability. An anomaly score is neither a fault diagnosis nor a control command.
Illustration · six synthetic points · not an ML model
1 / 6 · Question
Consider six days of history for an anonymous rotating asset. The aim is to compare possible explanations, not to intervene in a control system.
2 / 6 · Data
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
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
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
The packet carries the question, source version, method, supporting and contradicting evidence, alternatives and missing data. Uncertainty remains visible.
6 / 6 · Review
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
This change is not a fault diagnosis.
Load or measurement conditions may have changed. Mechanical condition cannot yet be distinguished.
Speed, load, sensor calibration and maintenance records are absent.
| 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.
04Experience
Log SWAPP use, agent steps, model and tool calls, experiments, failures and human corrections within each employee workspace. Select records with verified outcomes and reusable value; review them for skills, RAG and local model training.

Retrieve knowledge; keep weights unchanged.
Reviewed documents and workflow knowledge form a private vector dataset with access scope and source versions. Agents use relevant knowledge in answers or task context.
Make workflows repeatable.
Version a SWAPP workflow or agent method with preconditions, steps, authority, expected outcomes and tests. A record becomes a skill after its repeatable contribution is verified.
Separate, model-bound adapters for S1 and S2.
Reviewed examples are converted to the role’s data format. Choose LoRA or a compatible 4-bit QLoRA path for Unsloth-supported models. Evaluate the adapter against the same base on independent tasks before human-approved use.
Logging an operation does not grant permission to train on every payload. Remove secrets and unnecessary personal data; failures and corrections can also be valuable. RAG does not change weights. Fine-tuning candidates enter local operation only after independent evaluation and human approval.
05Development
The new pushes were reviewed at source level. Execution results below are dated upstream acceptance reports, not UMAY runtime tests rerun on this site.
Employee container template, access boundaries and persistent operation ledger.
From synthetic workflows to the real application; a digital-twin experiment setup.
Valuable records → skills / RAG / fine-tuning; independent local-model comparisons.
Architecture developed with powerful models, human approval and reversible releases.
Track teacher calls, coding, experiments, GPU training and local inference separately. Keep budget estimates, measured usage and reconciled invoices distinct. Include failed jobs and retries in costs.
Employee container architecture, operation records, value selection, local learning and changes with powerful models; contracts, acceptance criteria and a handoff prompt in one pack.
Download implementation pack · Markdown (Turkish)Project licenses for these repositories have not been selected; public access does not grant commercial reuse permission. Verify controlled forks and usage scope against source and model licenses. This manifesto, plan and example do not prove a working company integration.
Turn verified experience from each employee’s isolated workspace into lasting capabilities for company-owned local models and agents.
Explore development on GitHub