Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
Decision classifiers
Models → Decision Classifiers groups closed-set decision systems separately
from chat models. Each classifier has a shareable URL and remains discoverable
when disabled. The old /models/llms/jev URL redirects to
/models/decision-classifiers/jev; its command-palette identity remains stable.
Jev
Jev retains installation, manual scoring, trained-head adoption, integration switches, per-source policies and agreement counters. Its stored feature ID, API endpoints and defaults are unchanged. Disabling Jev integrations does not disable its explicit manual experiments. No migration is necessary.
Jev opens Try a decision, with input and evidence alongside each other on
wide screens. Integrations, Results, Training, and Setup have
shareable URLs under /models/decision-classifiers/jev/. Local view switches
retain drafts, the last score and active work; reloading does not persist text
drafts. Legacy /models/llms/jev/:taskView links redirect to the corresponding
view. Installation, scoring, training and adoption still require explicit actions.
Laya-MLX
/models/decision-classifiers/laya-mlx offers native Apple Silicon experiments
using the 322M multilingual checkpoint. Windows and Intel hosts can discover it
but cannot install or run it; Jev remains a separate choice.
- Install a native Apple Silicon Python 3.11+ on macOS 14+.
- Click Install Laya-MLX to create a dedicated environment, install the pinned runtime, download approximately 650 MB of weights, and verify loading. Additional disk space is needed for runtime dependencies.
- Enable experiments, supply a question, premise and 2–12 distinct options, then click Run experiment. Installation and enablement are separate.
The runtime Git revision and multilingual weight revision are pinned in
server/lib/layaMlx.js. Setup downloads only on an explicit action. Status
reads local installation metadata and file presence; it never loads a model.
Scoring runs offline, passes the premise through stdin, and releases the process
and model after each request. Only one experiment or installation can run at a
time. A request has a two-minute timeout and disconnect cancellation. Disable
blocks subsequent experiments; it does not cancel one already in flight.
The checkpoint has a 1,024-token combined budget, with a smaller question/options budget. PortOS rejects over-budget input and option labels that upstream would truncate, rather than classifying a partial premise. The UI reports normalized choice probabilities, entropy confidence, winner/runner-up margin and total elapsed time including process/model startup. These are not Jev entailment probabilities or upstream warm-inference latency measurements. Ties always abstain, including at a zero margin floor.
Laya is currently a manual experiment adapter. Selecting or enabling it does not replace Jev for triage, issue replies, scope checks or any other automation. Its thresholds need task-specific evaluation before production adoption; no speed or accuracy superiority is claimed by this integration. Additional classifiers can add their own panel/adapter and declared route without changing Jev's persisted policy format or pretending all scoring semantics are identical.
Storage and privacy
The enable flag uses the existing machine-local instance-feature store. Runtime,
model and installation marker live under data/python/laya-mlx/, in the existing
rebuildable Python-runtime asset tree; they are not native app records or new
JSON collections. No new database schema, seed, migration or federation payload
is introduced. The anchored /python/laya-mlx/ backup exclusion covers this rebuildable
installation. Inputs/results are memory-only and never enter logs, counters,
files, issue reports or federation. Errors returned by the service are bounded
codes, not subprocess output or paths.