Skip to content

Studio 2.2.6 validation

Validation date: 2026-09-14. Resource values are real host measurements; synthetic PA/DPD runs establish software behavior, not physical RF performance.

Application and UI

  • Python regression: 951 tests passed in the complete run; its API snapshot comparison was rerun successfully after the new nullable job-count fields were exported. Additional cleanup-failure and request-boundary tests were then run against the final implementation. The release PR runs the complete configured CI matrix again.
  • Frontend: 234 tests in 55 files passed, with strict TypeScript and lint checks. OpenAPI generation and the committed TypeScript contract agree.
  • Real Chromium over the installed shared CDP browser: local telemetry; web layout at 1366, 390 and 360 pixels; one top-bar UTC timestamp; no old per-page alert; no horizontal overflow or toolbar collisions. Stale samples hide utilization, and navigating away stops resource polling. The web layout check used a local API adapter; production verification is recorded separately with the release.
  • Current Home and local load-page screenshots were captured from a clean workspace. Earlier unchanged signal-chain and ILC images retain their 2.2.5 provenance.
  • Security regression covers directory traversal, local-only route denial, cross-session access, private telemetry authentication/schema/body limits, expired sessions, duplicate/malformed headers, concurrent writes, heavy-work limits and low temporary storage. GPU-cleanup failure must leave health unavailable while expired tenants are still removed.

Resource overhead and GPU execution

The loopback HTTP check made 100 authenticated status reads while other Python tests were running: median 1.01 ms, P95 1.56 ms, maximum 1.87 ms, and at most 557 bytes per local response. This measures cached endpoint latency on this host, not public Internet latency or a capacity benchmark. Tests verify that reading the cache never invokes nvidia-smi; one sampler per machine runs every five seconds with a two-second GPU-command timeout. Public job-count collection is shared across viewers for five seconds.

The GPU smoke record uses the same non-root, network-free, read-only container command and resource limits as production. It runs one-epoch GRU PA training, GRU DPD training and default ILC/ILA against 32,768 synthetic pairs, with 6,452 held-out test samples. The Ideal test-feedback waveform remains separate from fitted DPD training. These are execution checks, not convergence or quality guarantees.

Advisory and release checks

Security review and scan summary document remaining findings without suppressions. Frontend and actual Python environment audits accompany the full OS/container scan.

The GitHub release carries merged source, deployed API/agent/image, Pages build and published-wheel identities, followed by a public session and CUDA workflow check. Linux/macOS/Windows default-install CI verifies dependency resolution and headless browser fallback; native-window and MPS hardware coverage remain limited to the support matrix.