SUPERPHIVESSEL LOCAL CONSOLE FOR WINDOWS
======================================
Portable Ollama dashboard with read-only DEFAULT and optional manually
enabled local inference trials. No Git or Node installation needed.
The ZIP includes its own Node.js runtime (with its license).

DOWNLOAD AFTER #50 MERGES AND PAGES DEPLOYS:
https://michaelwave369.github.io/SuperPhiVessel/local/
  -> SuperPhiVessel-Local-Console-Windows.zip

1. Download this ZIP from the project's official GitHub Pages URL.
2. Right-click -> Extract All. DO NOT run directly inside ZIP.
3. Review the source files if you like.
4. Double-click Start-Local-Vessie.cmd as your normal Windows user.
   This remains READ ONLY. It cannot run any model prompts.
5. Your default browser should open http://127.0.0.1:8791/
   If not, copy that local address into the browser yourself.
6. Press "Discover my local models".
7. Stop the local server by pressing Ctrl+C in the console or closing it.

The dashboard gets models ONLY from an existing local Ollama instance
at http://127.0.0.1:11434. If Ollama is not installed or isn't running,
you will see an unavailable/empty report. No model gets installed.
Optional Ollama installer: https://ollama.com/download/windows

OPTIONAL PHISERVE MODEL DESK (SEPARATE OPT-IN, NOT READ ONLY):
- Start and authenticate PhiServe locally first. Keep its private token in
  a user-owned file, not a page, source file, command-line arg or URL.
- Stop the normal Vessie console before changing modes.
- Double-click Start-PhiServe-Desk.cmd, type ENABLE, then provide the
  PRIVATE TOKEN FILE PATH when prompted. The launcher never asks for
  the actual secret token text.
- In the PhiServe Model Desk click Discover local PhiServe models, select
  a registered GGUF alias, write a prompt and check the one-shot approval.
- Click Generate one approved draft. The model runs on this computer
  only through the authenticated PhiServe service at 127.0.0.1:11435.
- The answer is an UNTRUSTED text draft checked against its redacted
  SHA-256 receipt. It cannot run tools, modify memory, approve routing,
  access other agents, or release canonical Vessie runtime changes.
- Up to six attempted approved drafts per hour, never concurrently.
  If the service is busy, unauthorized or unavailable, it fails CLOSED.
- The read-only launcher still cannot execute any model prompt.
- The standard Ollama trial launcher remains a separate option.
- The downloaded portable ZIP now keeps its Node source dependencies
  under local-console/ and vessie-phiserve-draft-v0.1/.
- See packages/vessie-local-console/PHISERVE_MODEL_DESK_GUIDE.md in
  the source repository for security notes and operator field tests.

OPTIONAL EXPERIMENTAL LOCAL INFERENCE (NOT READ ONLY):
- Optional fixed comparison cards: after discovery, select a public
  Governance, Logic or Code protocol and click Load selected protocol.
  This only fills the prompt; it NEVER runs a model.
- To compare models, keep the public protocol prompt and selected
  token cap exactly the same, then deliberately select and approve
  each eligible local model one at a time. No automatic multi-model
  runs, cloud calls or routing approvals.
- If you edit the loaded prompt or token cap, a protocol-labeled
  execution is blocked; re-load the protocol or select Custom.
- Approved fixed-protocol receipts record a public versioned
  protocol_id. This ID is not proof of model quality or hardware
  execution. Custom and older receipts remain protocol-unlabeled.
- Stop the read-only console first.
- Double-click Start-Local-Trial.cmd, read its warning, and enter
  the exact word ENABLE. No administrative access needed.
- Open the local dashboard, click Discover, choose a model whose
  local weight size Ollama reported. CLOUD REF / UNKNOWN entries
  are intentionally ineligible.
- Enter one short prompt (up to 2,000 characters), choose 64 or 128
  maximum generated tokens, check the approval box and confirm
  in the separate browser confirmation. Each prompt needs a new
  approval. No automatic or repeated runs are scheduled.
- Ollama may take up to 90 seconds and use CPU/GPU resources.
  Larger models may fail, load slowly, or offload to system RAM.
- Review the generated answer and the evidence receipt in the
  browser. The receipt intentionally excludes the prompt and
  generated answer. It includes model name, latency and Ollama-
  reported token timing/counts (not independently attested).
- Newer receipts also include reported prompt-evaluation duration,
  an allowlisted stop reason and whether the requested token cap
  was reached. A reached cap is NOT proof that the answer was cut
  short, nor that the answer was correct.
- Phase timing is for review only. The trial runner uses keep_alive:0
  to unload weights after the request. Do not treat separate
  approved trials as a continuous warm-cache speed benchmark.
- Export the redacted performance receipt only if you explicitly choose to.
- For a completed public fixed-protocol response, the Human Review
  panel offers Reveal operator answer reference. Click it after
  reading the actual model response to see a short expected answer
  and three manual inspection checks. Reference visibility resets
  on a new trial. Custom/unlabeled responses have no answer key.
  Revealing does NOT auto-grade or change your recorded rating.
- Optionally, after a completed response, record a HUMAN answer review:
  usefulness 1-5 or not assessed, whether the answer appears complete,
  and whether YOU separately checked factual claims. This is a human
  self-report, not a factual audit, autonomous grader or authority grant.
- The review's redacted JSON links back to the output SHA-256 but never
  includes the prompt or model answer. You must explicitly click
  Export redacted human review to save it locally. A new trial or
  inventory refresh clears the current review association.
- Previous receipts without an answer cannot be retroactively rated
  from token timings or hashes alone.
- This trial feature limits execution to six attempts per hour per
  launched process. No cloud references, prompt history, memory
  writes, background agents, model routing or paid API authority.
- Opt-in is per-process, not a persistent setting. Restart the
  default launcher to return to no-execution read-only mode.
- Ollama model metadata is advisory and not a perfect guarantee
  of origin. Check what you run and avoid sensitive prompts.

LOCAL EVIDENCE BENCH (AVAILABLE EVEN IN DEFAULT READ-ONLY MODE):
- The Protocol Cohorts table below the main bench groups existing
  fixed-protocol evidence by public protocol + model and presents
  receipt/review counts, median reported timings, and missing-evidence
  flags. Choose a public protocol from the display filter if useful.
- This view NEVER runs a model, ranks candidates, or grants BrainC
  authority. Single observations are explicitly sparse evidence.
- Old custom receipts stay in the main bench only. A purported fixed
  protocol with a mismatched token cap is not counted as comparable.
- All imported JSON remains unauthenticated and human ratings remain
  subjective; medians do not establish repeatability or GPU fit.
- This bench compares redacted performance receipts with human self-
  reports. It NEVER runs a model, creates a rank or approves routing.
- After a completed trial, click Keep current performance receipt.
  After recording a human review, click Keep current human review.
  Neither result enters the comparison until you click those buttons.
- For old receipts or comparing across different model sessions, select
  exported redacted JSON files in the file input. Click Read selected
  local files. Up to 24 distinct entries, at most 16 KiB per file.
- The first exported performance-only receipt can be imported even
  if its answer was not retained. It will show No review; no answer
  quality is invented or guessed from token throughput.
- Click Save portable bench JSON to export the actual, strictly
  redacted observation set for explicit restore in another session.
  Keep this file privately: model IDs, timestamps and output hashes
  may still be sensitive even with prompt and answer text omitted.
- To restore later, reopen Vessie (read-only mode is sufficient),
  choose the portable JSON file and click Read selected local files.
  Up to 64 KiB per bundle; up to 16 KiB per standalone receipt.
  Any invalid batch is refused without partially replacing the bench.
- The Export descriptive comparison JSON is VIEW-ONLY and cannot be
  reimported as a portable bench. They serve different purposes.
- The bench matches a review only on model name plus output SHA-256.
  A matching review remains a human assertion, not verified truth.
- Explicitly click Export redacted comparison to download a JSON
  summary. No prompt or answer content is included.
- Click Clear this bench or close/refresh the page to discard the
  in-memory comparison. No persistent storage, server uploads,
  automatic rankings, or model routing exists.

SECURITY AND SCOPE:
- Bundled Node runtime is prepackaged by the GitHub Actions Windows
  build, not downloaded or installed when you start the app.
- Node license is included as LICENSE.node.txt.
- Starts an HTTP server bound ONLY to 127.0.0.1 port 8791.
- Never opens a LAN or public network listener.
- Never sends local model names to GitHub Pages or cloud APIs.
- A random read-only API secret is stored in the running process only,
  not in files or shared browser storage.
- The DEFAULT read-only launcher exposes no prompts, inference or
  execution. The separate operator-enabled trial launcher can send
  one short prompt at a time to an eligible local Ollama model;
  no cloud models, agents, memory reads, software installs, paid
  API calls, CA installation or governance changes are enabled.
- Does not connect the HOSTED Vessie React page to your PC.
  That remains a separate HTTPS R2 pilot and a different security scope.
- Your local browser displays an independent minimal Vessie console.
  This is not the full canonical 13 MB PhiVessel runtime.
- No startup persistence: stop the process and the server disappears.
- Do not run it as administrator or disable antivirus warnings to
  force execution. Only run files whose origin you trust.

If port 8791 is occupied, stop the other local process before retrying.
Do not forward ports, alter proxy/firewall settings or expose the console.
