# StarloOS 0.9.0 — verification and product boundary

Date: 2026-10-22. Environment: Linux x64, Node.js v22.16.0. This report applies to the 0.9.0 local Workbench and matching downloadable application. It is not an independent security audit, Windows certification, or a claim of technological novelty.

## Verified in this build environment

- **345 Node tests passed.** This includes 41 new Workbench tests and 304 retained tests, updated only where release version expectations changed. The new tests execute actual local algorithms and actual loopback HTTP APIs; the older language-model/provider integration tests use fixtures.
- **18 isolated-browser checks passed.** The UI submitted real CSV, document comparison and classifier jobs to the actual local server; trained a model; exported/imported its training card; recorded a human review; and downloaded a ZIP. The mobile layout fitted the viewport and no page-script exception occurred.
- **31 fresh-process HTTP/acceptance checks passed.** A newly started local runtime calculated a real CSV sum of 625 and group totals, trained/evaluated a document classifier, routed new documents with an abstention, compared two documents, verified an exported bundle, rejected an altered artifact, shut down, and restarted with the accepted result and trained model intact. All five website download payload hashes matched their published manifests.
- The full website build generates software downloads from source plus pinned release inputs. All HTML links and application route targets are checked. Runtime modules are syntax-checked recursively.
- The Windows x64 executable is compiled from the included application using clang/lld. Its PE type and embedded source manifest are verified. It remains unsigned. Building it is not executing it on Windows.
- The original motion engines and selected original UI assets remain byte-identical to the preceding release. The archived whitepaper Markdown and PDF remain unchanged. The new entry experience, workbench, guides, download versions and release metadata are intentional changes.

## Browser-method qualification

Native browser navigation is blocked by this environment's policy. No policy was changed. Browser checks render the actual HTML/CSS/JS in an isolated page and bridge requests to the actual authenticated localhost server. SHA-256 in that insecure isolated context uses a Python digest bridge; the same verifier is separately tested with real Node WebCrypto. These checks do not establish that native Windows browser launch, a Vercel origin, or every browser/OS combination works.

## What is operational without an AI account

1. CSV profile, duplicates/missing-value counts, numeric statistics, and optional grouped summaries.
2. Supplied-document line comparison and exact change artifacts.
3. A TF-IDF centroid document classifier learned from labelled examples, explicit separate holdout evaluation, category proposals, overlap terms, and abstention rules.
4. Durable local jobs, frozen model snapshots, human review states, individual file exports, and portable receipt verification.

These operations are executed, not canned model responses. The classifier is narrow machine learning, not a pretrained foundation model. The two-case held-out acceptance example is an engineering check, not a credible generalization benchmark or commercial accuracy claim. Human review remains necessary. CSV calculations use floating point, not an accounting ledger.

## What is NOT established

- Native Windows installation, initial runtime download, shortcuts, native shutdown or uninstallation.
- Publisher signing, independent penetration testing, platform distribution approval, or complete commercial production readiness.
- The optional Unix bootstrap's external download path or native macOS execution.
- A bundled generative model, funded cloud call, external Ollama tool loop, real remote MCP/search/Docker/RPC operation, or end-to-end Vercel deployment.
- Useful work for arbitrary tasks with no model configuration.
- Bootable Windows/Linux replacement, kernel or device drivers, new frontier AI method, AGI, full StarNet parity, or universal automation.
- Token deployment, crypto checkout, staking, trading, income, paid users or returns. STARLO controls remain pre-launch.

## Privacy and limits

The new built-in workflows operate in the local process without network calls. Their state uses owner-restricted files and an explicit bounded queue. Optional advanced model requests send supplied context to the chosen provider. Receipts check internal file integrity and event consistency; they are not signatures and cannot prevent someone rewriting an entire bundle. See docs/WORKBENCH.md for limits, data locations and safe removal.

## Reproduce

```
npm test
npm run build
npm start
node scripts/verify-job.mjs /path/to/extracted-mission
python3 installer/build-windows.py
```

Use the full web source package for `npm test`/`npm run build`: it includes pinned release inputs needed to reconstruct the downloads. The smaller matching desktop-source archive intentionally excludes those binary inputs; it can run `npm start`, direct Workbench tests, and the Windows source compiler.

The release evidence contains sanitized checks and a real tutorial output bundle. It does not contain owner credentials, model keys, machine ready files, or private database state. No online deployment was changed during this release.
