Aegis: Closed-Loop Intelligence Engine
Ground behavior, improve it, and defend every ship decision with evidence.
Mode
Eval-first
Release
Gate-aware
Reports
Shareable
Access
Accounts enabled
Public siteWorkspace settings
Shell policy
The workspace chrome does not inject sample benchmark rows, synthetic scores, or decorative regression traces. Live evidence belongs in the closed loop, research runs, review queue, and release train after a real workspace is populated.
Closed Loop
Import traces, run the strict loop, and open the dossier.
Research Runs
Measure benchmark deltas and investigate candidate behavior.
Review Queue
Attach ownership, severity, and operator judgment.
Release Train
Persist gate state beside the same artifact lineage.
Launch-grade proof should be grounded in persisted artifacts, not shell placeholders.
surface
purpose
required
owner
dataset
fixed benchmark contract
yes
research
comparison
baseline vs candidate delta
yes
operator
review
annotated release judgment
yes
human
promotion
gate outcome + lineage
yes
release

Training Lab

Use this after an eval has already identified a real failure mode worth fixing. VERL plus execution_mode=real is the launch path; deterministic or simulated runs are useful for local lab work, but they are not launch-proof evidence.

Start a training experiment

Use your org or workspace slug for team runs. For solo use, keep personal.

Use verl for the real GPU-backed training path. standard remains a lab backend for deterministic local experiments and is not launch-proof evidence.

No training jobs yet.

Begin with an eval run, then use the form above to create a training experiment if the evidence supports it.

LoRA Adapter Management

Loading adapters...