01LOCAL
Local-first
Sensitive context remains on the user’s device whenever possible. Changing a model or provider does not require surrendering personal memory.
NOETELIC / R.A.B.İ.A + CCOS
Personal intelligence with clear boundaries.
Design principles, dated test results and the development path.
TRUST / DESIGN DECISION
Privacy is not a feature. It is a system boundary.
Security is more than encrypting data. R.A.B.İ.A uses separate gates for what becomes memory, how much authority a tool receives and when an outcome is allowed to become shared history.
02SCOPED
Least privilege
Tools receive only the scope and time they actually need. Authority begins with the task and does not leak beyond it.
03REVERSIBLE
Reversibility
External actions are designed to be reversible whenever possible. An outcome is observed before it is treated as successful.
04TRACEABLE
Provenance
Every piece of knowledge carries origin, time and confidence. Memory, inference, research and dream remain distinct.
EVIDENCE / DEVELOPMENT STATUS
Measurement before spectacle.
Recent work focuses on local planning, separate action permissions and cancellation controls for Computer Use. Explore what was tested and what comes next.
From observation to proposal
A real local 7B model proposed steps from a controlled test window’s structure. Rebinding to a fresh observation was checked; no actual typing was performed in this trial.
Stopping during generation
Two cancellation paths were tested during real 7B generation. Partial responses stayed unpublished and the model context could be reused. This does not test stopping every computer action.
Keep actions distinct
Window-activation history was separated from typing records and linked to task history. Permission and restart checks were tested in synthetic scenarios; end-to-end computer interaction remains in development.
R.A.B.İ.A DEVELOPMENT STATUSVGP / IMPLEMENTATION AND ACCEPTANCE
Local-firstActive developmentEvidence-gated delivery
Local modelcontrolled window-planning tests
Separate permissionswindow focus and typing
Cancellationtests during model generation
4 Octdevelopment summary · 2026
The 4 October development summary, test scopes and a link to earlier measurements. No personal data; not an application download.
CAPABILITY / BOUNDARY
What can it do—and what does it not promise?
Sourced recall
Current approach: Conversation and admitted records can be used while preserving their origins.
Boundary: No guarantee of remembering every detail forever or without error.
Permissioned tools
Current approach: Operations run within task, resource and authority boundaries.
Boundary: No implicit unlimited permission for cameras, the internet or files.
Forgetting and withdrawal
Current approach: Source-linked records can be removed through controlled workflows.
Boundary: Overall acceptance for older records with missing provenance remains open. No claim of physical SSD secure erase or automatic deletion of all independent backups.
Local visual work
Current approach: Image understanding, generation and concurrent chat were tested with selected local models.
Boundary: Available features and performance depend on the chosen model and device resources.
DATED DEVELOPMENT EVIDENCE
Numbers matter with their conditions.
The results below were checked against existing development reports. Tests were not rerun for this website update. These are neither live counters nor success rates for the whole product.
Recall from the long-term archive
- Active / archived records
- 50.000 / 100.000
- Synthetic queries
- 100
- Recall@4
- 100%
- Cold-search P95
- 0,0526 ms
Short, specific lexical queries against a reopened store. Recall@4 means the expected record appears in the first four results. This is not general semantic recall or end-to-end chat latency.
Source: BSD-7 closure report and capacity measurements.Model transitions and failure states
- Deterministic lifecycle cycles
- 10 / 10
- Tauri tests
- 176 / 176
- Frontend contract/security
- 23 / 23
Cancellation, deadlines, late-result isolation and rollback states were covered. Ten cycles do not mean a large physical model was loaded onto a GPU ten times; the report distinguishes these.
Source: BSD-8 closure report.Supervised chat + image generation
- Verified PNG
- 256 × 256
- Total job time
- 7,576 s
- Chat chunks during job
- 95
- Model fallback
- 0
RTX 5070; 7B native CUDA chat + SD-Turbo Vulkan. The first run was rejected for insufficient VRAM; the second passed. The 95 chunks are not tokens and include loading/verification time. A single run is not p95 or a guarantee for every device.
Source: OLG-04 local visual pipeline report and supervised GPU receipt.4 OCTOBER 2026 / DEVELOPMENT PATH
From continuity to a capable companion.
Alongside memory, persistent tasks and permissioned file flows, a new focus: Computer Use infrastructure that understands a selected window, proposes steps and preserves action boundaries.
In validationIn progress
VGP-00–02In validation
Project context and sourced memoryVGP-03–05In validation
Fresh knowledge, user alignment and experienceVGP-06–08In progress
Persistent tasks and controlled background workVGP-09In progress
Bounded tool executionVGP-10–11In progress
Trial spaces and cross-device continuityVGP-12–14In progress
Multimodal context, skills and model alignmentVGP-14AIn progress
Computer Use: observation, planning and separate action permissionsVGP-15–16In progress
Unified interface and resource managementVGP-17–18In progress
Integration testing and release preparationR.A.B.İ.A remains a development build. Selected desktop flows have been tested; broader usage testing continues.
A CLOSER LOOK