Validation and evidence

Functional qualification

The current Windows core has been exercised through real native controls and the actual preload/host/runtime boundary. Browser fixtures and static atlas simulations are not used as evidence of model inference.

The qualified artifact is RegesCore-1.0-35B-MXFP4_MOE, 21,670,492,576 bytes, architecture qwen35moe, with 40 transformer blocks. The installed test used CUDA, exact 8,192 context and 12 GPU layers on the documented RTX 4070 SUPER system. Engine logs confirmed the applied context and placement arguments.

Checked flows

• Real local inventory and parsed model metadata; actual captured hardware.
• Exact custom context input and negative validation for invalid settings.
• Model load, completed text response, generation Stop and a successful fresh response.
• Explicit unload, app close and refusal of the owned private port afterward.
• Saved local history restored across a process restart.
• Actual packaged mode and resources/app.asar source paths.
• Native sandbox/isolation settings, CSP, no Node globals and no renderer HTTP inference.
• Full-window startup presentation, lazy local scene module and About-to-Chat GPU cleanup.
• Normal user-profile first launch with empty conversation history and real backend readiness.

Installer validation compared installed files with the release manifest. Model size and modification time were checked around the flow; full content hashing belongs in the release provenance record. No model weights are bundled with the installer.

CPU-expert placement follow-up

A subsequent actual installed-application flow used the same file, 8,192 context, CUDA and All layers with --cpu-moe, --threads 8, --threads-batch 8, automatic fit disabled and one slot. It completed an arithmetic answer, canceled a streamed response, completed recovery, unloaded and restored history after relaunch. Native security settings and owned-port shutdown passed again; the normal user preferences were unchanged by isolated test profiles.

Controls-only checks retained exactly 57,000 tokens through the MoE action and Save/relaunch, rejected invalid threads and ineligible dense layouts, and kept saved settings unchanged during draft editing. This is context-input/persistence evidence, not proof of loading 57,000 tokens.

Strata qualification checkpoints

Boundary	Evidence available	Limit
Native CUDA engine	Encoding/kernel comparisons and exact-file arithmetic smoke passed	One artifact/sm89 profile; no broad quality claim
Private JSONL helper	2 GiB-cache chat, history-dependent turn, overflow, cancellation, fresh recovery and unload passed	Does not exercise Electron IPC or renderer
TypeScript / helper regressions	65 TypeScript tests and 18 helper regressions passed	Closed contracts and error/lifetime behavior; not speed evidence
Actual native controls	1,440/1,024 layouts, eligibility, exact bounds, persistence and ownership passed	No model loaded in this controls-only check
Electron Chat inference	Passed with live/final metrics at 1,792 MiB cache, 8,192 context and eight CPU participants	Exact reference profile only
Strata packaged/installed executable	Actual per-user installed executable passed the complete flow	Publisher signing/public distribution remain separate

The controls-only application check kept 57000 unchanged when switching to the Strata draft and correctly blocked that over-limit value. It checked cache steps, thread bounds, Strata’s 4,096 output ceiling, and preservation of a larger output value across engine switches. The actual engine label stayed Lemonade until a real engine switch; changing a draft did not persist settings. Save and relaunch restored the chosen Strata profile. The normal user profile remained unchanged, both owned private ports closed, and there were no unexpected renderer or network errors.

Do not substitute Lemonade, helper-only or controls-only receipts for native Strata inference. The actual installed Strata flow at 1,792 MiB cache separately passed arithmetic, a history-dependent turn, Stop, fresh recovery, unload and saved-history relaunch. Live/footer counts, rate and finish reason matched request-correlated native events; prior-response stats stayed visible during the next generation, and saved/reopened stats were identical. No unexpected renderer, console or HTTP errors were observed; owned ports closed and the original model size/mtime stayed unchanged. A sanitized evidence summary records the exact scope. See the runtime guide for exact resource identity and initial scope.

What is not established

This evidence does not qualify 397B execution, all GGUFs, all GPUs, long-context quality, images, embeddings, tool calling, other Strata artifacts/hardware, predictive SSD expert paging, public API compatibility or a throughput advantage. A correct short arithmetic answer is deliberately a small functionality test.

User navigation can hide the DOM answer while generation completes. In one installed check the user opened About; request-correlated completion events established the actual answer before the test returned to Chat. That distinction matters when diagnosing an apparent hang.

Publishing a result

Keep a sanitized receipt with artifact identity, source build, effective runtime settings, hardware, exact prompt, terminal events, errors, memory observations and cleanup outcome. Do not publish credentials or entire private histories. Promote a claim only to the scope actually tested, and use the benchmark protocol for any performance statement.