# StarloOS Evidence Desk 1.0 — Verification and release boundaries

Date: **22 October 2026**. Base: StarloOS Horizon website / 0.9.0 runtime. New feature edition: evidence-desk-1.0-2026-10-22.

## Results

| Check | Result | What the result establishes |
|---|---|---|
| Node automated suite | **429/429 passed** | The 383 inherited tests plus 46 new deterministic, storage, signing, API and runtime-integration tests pass. |
| Browser-local interface | **27/27 passed** | Actual UI source: create file, add versions, compare, require an exact quote, record human review, generate reports, export, verify, show NOT LIVE controls and request-only download, responsive layouts, motion/pause. |
| Connected interface | **12/12 passed** | Actual local session/API connection: explicit local-to-server copy, no implicit acceptance transfer, source approval, fixture capture provenance, bounded AI consent, queued state and actual credit reservation. |
| Native local HTTP | **29/29 passed** | Research/guide/app routes, asset MIME, authentication boundary, security headers, PDF bytes, no public software download archives. |
| Fresh server persistence | Passed, included in Node suite | Actual file store survives a fresh server and store instance, including source bytes and human acceptance. |
| Agent runtime handoff | Passed, included in Node suite | Fixture model choices drive real file tools and 12 approvals through the existing four-agent runtime. 16 fixture model turns, 4 files, returned reservation and unreviewed final acceptance. |
| Offline signature/content verification | Passed, included in Node suite | Native Ed25519 test keys, trusted-vs-untrusted fingerprint distinction, altered bytes and wrong fingerprint are rejected. |
| External site fetch | **Not verified** | A real attempted request to ethereum.org failed with DNS `ECONNREFUSED` in this environment. |

The model integration test's illustrative **0.16 internal-credit** charge and token counts are fixture-defined test data, not a provider bill, commercial price, or benchmark. The test generates `brief.md`, `sources.csv`, `unknowns.md` and the inherited `final.md`. It does not demonstrate a real model's reasoning quality.

## Browser test method — important

Native Chromium navigation returned `ERR_BLOCKED_BY_ADMINISTRATOR`. No policy was disabled or bypassed. The checks rendered the supplied HTML/CSS/JS in an isolated page, loaded the modules/assets from the source files, and bridged requests to a real local HTTP server. A SHA-256 bridge was used because that isolated page lacks the browser's secure-context WebCrypto API; native WebCrypto behavior was tested separately in Node.

The renderer's IndexedDB was unavailable, so the **memory-only fallback** was exercised. Native on-origin IndexedDB persistence, browser service-worker behavior, real-domain sessions, and hardware GPU performance are not certified by these tests. Native HTTP auth/CSRF and durable server persistence were tested separately. Screenshot examples show the actual UI but fictional/pasted test material, not market research or live customer work.

## Security and research regression coverage

- Private project ownership and existing session/CSRF controls.
- Explicit current revision on edits; concurrent stale writes rejected.
- Exact quotation required before attaching a claim to source evidence.
- Old-snapshot references and stale/source-threshold gates.
- Human acceptance invalidated after material changes; imported data requires fresh review.
- Questions remain questions rather than accepted facts.
- Exact public HTTPS consent, same-origin redirect boundary, private-network protections and bounded capture rate.
- Hash manifest matches exported file bytes; duplicate descriptor tricks and modified files fail verification.
- Exported CSV cells escaped against spreadsheet formula injection.
- Permanent project-create and capture receipts with distinct request identities.
- Optional AI launch idempotency, existing spending reservation and file-approval controls.
- Unsigned manifests do not identify an issuer; embedded signing keys are not implicitly trusted.
- Credentials stay server-side. No key is bundled, and signing-key generation is never run by the build.

This is targeted regression coverage, **not an independent security audit or proof of prompt-injection immunity**.

## Preserved functionality

Comparison against the delivered Horizon archive confirms unchanged bytes for the agent runtime/provider/tools/domain/quality code; command-center application logic; rotating Earth renderer, visual controls and styles; software-download popup and token notices; original whitepaper PDF/Markdown; and the pinned Windows installer candidate input. All inherited command-center control IDs remain present. See `evidence/current/preservation.json`.

The new API routes are narrow additions to the existing authenticated handler, signing configuration and capture-origin checks. `docs/EVIDENCE-CHANGES.json` records explicit exceptions to historical visual-only preservation tests; original historical manifests were not rewritten.

## Distribution and setup

The ZIP is a website/source distribution with `package.json` and `vercel.json` at its root. Build settings: Other, Node 22.x, `npm run build`, output `public`. Build checks the legacy release inputs but removes generated software archives from public output. The only software-download metadata says `upon-request`, and no request submission endpoint is implied. Whitepaper PDFs and generated research-result exports remain downloadable.

**Website edition != rebuilt Windows release.** Pinned 0.9.0 installer/source inputs are unchanged. They are not a new desktop build containing this edition and remain unsigned/unverified on Windows. The owner can run this latest source locally, but the public site does not provide its archive automatically.

Browser-local research requires no keys: paste, compare, review and export. Server collection/persistence requires configured private storage and login. Optional model-directed work requires a funded provider and the existing execution/dispatcher setup. Cloud unattended execution is not magically enabled by the website.

## Unverified or not implemented

No live Vercel deployment, real Redis connection, public website read success, funded model call, remote MCP execution, native Windows installation or actual token transaction was verified. No market-data feed, automatic daily monitor, cryptographic proof of source truth, new foundation model, consumer onboarding service or commercial acceptance rate is claimed.

STARLO purchases, live token-to-credit payments, mainnet receipts, refunds, creator compensation, sustainable treasury settlement, liquidity and staking are **not active**. The DEX Screener, Buy and CA controls stay NOT LIVE. No wallet is controlled; no on-chain write occurs. Adding a signature and research gates does not solve product-market fit or prove token-holder value.

## Reproduction

```sh
npm ci --ignore-scripts --no-audit --no-fund
npm test
npm run build
node tests/evidence-static-acceptance.mjs
```

Browser evidence harnesses are `tests/evidence_browser.py` and `tests/evidence_connected_browser.py`. They require Python Playwright and Chromium and use the isolated method explained above; the connected harness requires the expressly test-only server in `tests/serve-evidence-fixture.mjs`. Never deploy that fixture as the production service.

## Fresh-archive check

The final source archive is extracted to a fresh directory and tested independently. A successful record is stored in the separately delivered verification attachment and release evidence. No unverified deployment claim is inferred from a local build.
