RESEARCH ↔ IMPLEMENTATION TRACEABILITY

Defaults. Rebuilds.
Extensions. Evidence.

Every named provider below is a product-research reference. UNITIQX runs original user-interface code and separate local Ollama models; it does not copy commercial code, proprietary model weights, subscription rights, or credential flows.

12 provider-workflow comparison contracts

Architecture: original from scratch vs reuse

From scratch: this public chat page's HTML/CSS/JS, standalone localhost inference broker, policy prompts, state mapping, and backend proxy integration were authored for UNITIQX. Reused: Python standard library HTTP service, Ubuntu/systemd, Tailscale Funnel/public-room, Ollama pretrained third-party open-weight models, browsers and fonts. "From scratch" does not mean training original frontier-sized language models or replacing the operating system.

Baseline interface acceptance

1. The workspace can be opened through an ordinary public HTTPS browser, displays independent UNITIQX branding and names its ACTUAL local model.
2. A real user prompt returns text from Ollama, not a canned sample; a second prompt carries bounded recent local history.
3. Browser-local conversations can be selected, cleared and exported; no other visitor can retrieve them through the server.
4. Tokens, CPU memory and concurrency are constrained. Requests fail closed with an accurate 4xx/5xx error instead of pretending success.
5. Private owner files, credentials, Workbench, external providers and admin APIs remain outside the public inference container/service.
6. Keyboard, responsive mobile rendering, invalid-input rejection and privacy negatives are tested and documented; physical-device/human review remains explicitly unverified.

Next limits to remove—safely

Backend architecture: capability-tagged model adapters, streaming tokens, cancel/job queues, source-grounded public retrieval, secured accounts/optional sync, independent per-profile components and editor previews. Private operator extensions: separately authenticated files/read-write with project-scoped permissions, diff previews, tests, policy grants, audit, approval and rollback. Compute: throughput benchmarks and optional approved GPU/worker expansion. Never make private admin APIs public merely for visibility.

Evidence and version boundaries

Machine-readable results: workspace inference tests, project stages, and wider public development registry. All private draft work should be represented by public status metadata—not the actual drafts, credentials, conversations or internal source.

Reference documentation used to ground the comparisons: OpenAI Projects, Claude Artifacts, Gemini Gems. Other links in the table are navigational research starting points, not an audit of every vendor feature.