Your local Vessel.
No development setup required.
Download a portable Windows ZIP with its own Node.js runtime. Double-click a launcher to open a private browser dashboard that can show your actual locally installed Ollama models.
Start in three steps
- Download the ZIP above and choose Extract All.
- Open the folder and read
START_HERE.txt. No installer or administrator account is needed. - Double-click
Start-Local-Vessie.cmdfor the default read-only console, then press Discover my local models.
The local console opens at http://127.0.0.1:8791/ and stops when you close its command window. Ollama must already be running on 127.0.0.1:11434 to show your models.
Optional fixed trial protocols for human model comparisons
In the separately enabled trial mode, choose a public fixed governance, logic or code prompt card and click Load. This only fills the form and never executes a model. Select a locally size-reported model, then manually approve each individual prompt using the existing checkbox and browser confirmation. The server validates the exact public protocol text and token limit before adding a versioned protocol ID to the redacted receipt. The Evidence Bench can show that ID alongside chronological results, including after portable bundle export/import. No automatic tournament, model ranking, cloud call or BrainC authority is created.
Protocol cohorts: compare coverage without picking a winner
The local Evidence Bench now groups known public-protocol receipts by fixed prompt and model, showing observation counts, matching human review counts, reported median timings and missing-evidence flags. A mismatched output-token cap is excluded; old custom trial receipts stay visible in the main bench. Cohorts do not trigger new prompts, score intelligence, independently verify hardware or authorize BrainC or Genius routing. Single observations are marked as sparse evidence and imported JSON remains unauthenticated.
Optional local one-shot experiment
Once you've discovered your models, you can optionally close the read-only process and launch Start-Local-Trial.cmd. It requires typing ENABLE at startup, then selecting a locally size-reported Ollama model, checking an explicit consent box, and confirming each prompt in the browser. Cloud-tagged references and unknown locations are excluded. Limited to six attempts per hour and 128 output tokens per trial. The generated answer stays in the local browser; an optional redacted receipt reports latency and token counts without saving the prompt.
Local Ollama inference may use your CPU/GPU and can be slow; these observations aren't model admission or independent device attestations. Start the normal launcher again to return to read-only mode.
Human protocol answer references, revealed only after a response
When you run an approved public Governance, Logic or Code trial, the local console offers a separate Reveal operator answer reference action after the response. It shows a short public expected answer and three human interpretation checks. Nothing is scored or submitted automatically, and the reference does not change your optional usefulness review or the redacted evidence formats. Custom/unlabeled trials remain without an answer reference; the model response stays local.
Human quality review, never an automatic grade
After an explicitly approved local trial completes and its answer is visible, you can record your own usefulness, completeness and factual-check assessment. This creates a separate redacted human self-report, linked to the response hash, not a model-generated grade or proof of correctness. The prompt and answer are omitted. Exporting the review JSON is optional and deliberate; no model routing, learning or cloud calls are enabled.
Compare evidence locally, without running another model
The optional Local Evidence Bench accepts previously exported redacted performance receipts and human reviews through a deliberate file selection, or a click to keep the current result in memory. Compare model latency, loading, prompt-processing and generation time beside your own usefulness and verification assessment. At most 24 observations are held in browser memory; file imports are limited to 16 KiB each, never uploaded to a server. Rows are chronological, not ranked. A review matches a generated output only by its model label and SHA-256 digest; this is not proof of correctness, device identity or model suitability. Download the redacted descriptive comparison only if you click Export. You can also click Save portable bench JSON to save strictly redacted evidence for a later session; choose that file manually next time to restore your bench. This is not automatic persistence or a cryptographic signature. Portable bundle files are limited to 64 KiB; the previous descriptive-comparison JSON remains a view-only report and cannot restore the bench. Close the tab or clear the bench to discard its in-memory contents.
What stays protected
The server listens only on loopback, uses an ephemeral per-process read-only API secret, enforces an exact Host and same-origin policy, and exposes no execution in the default read-only mode. The separately enabled one-shot local trial is bounded and requires per-prompt approval; it provides no memory/agent routing endpoint.
This is a local model inventory console, not the full canonical SuperPhiVessel runtime. The hosted React website is separate and does not automatically gain access to this local dashboard. No model install, paid calls, cloud upload, background execution, or routing approval occurs. Manual local inference is available only after a separate explicit opt-in.
If Windows or your antivirus blocks a downloaded program, verify the origin and source. Do not disable device security to force it.