FollowThrough / five-minute judge walkthrough 1. Start with https://gowtham66867.github.io/followthrough/ for a public, zero-cost browser walkthrough. This is entirely fictional and fixture-based. Review the five answers and deadline, approve, then add the client amendment. Inspect the impact map and the reopened approval. No live inference runs here. 2. For the working NVIDIA/Nebius application, download the v0.5.0 test build: https://github.com/gowtham66867/followthrough/releases/tag/v0.5.0 Follow README setup instructions. The ZIP includes the frontend build. The guided demo requires no account. Live inference needs a Nebius credential. No API key is included in source, the ZIP, the public website or this guide. 3. Load the fictional Northstar RFP. In Sources, explicitly choose Analyze with NVIDIA to make a runtime Token Factory call. Compare proposed requirements with the original quotations; confirm their interpretation and deadline. Match approved evidence, review answers, approve the package, and export. 4. Add the fictional amendment. Inspect Before you submit and Impact map. The changed delivery answer needs review, a separate security workbook is required, certification needs an honest disclosure, and the deadline moved. Completing an owner task does not approve an answer. Review the attachment and exact revision before another export. Nothing is submitted externally. 5. Inspect submission/evaluation-v05.json for the retained model prompt's 16-case synthetic development evaluation. Raw model and deterministic deadline-guard scores are separate. The rejected prompt experiment is preserved in evaluation-v05-after.json. These are not customer or held-out accuracy measurements. Quote validity does not establish semantic relevance. Restricted hosted access status The invitation gate and persistent inference allowance are implemented and tested. A working free host has NOT been identified or deployed. A Nebius Token Factory API key provides inference access, not a web hosting server. Do not claim judges already have free hosted live access. The downloadable build's live mode still requires credentials until that deployment is resolved. Hosting on an EXISTING free server Use a new, dedicated data directory, never a personal/Gmail workspace. Copy judge.example.yaml, set its exact HTTPS origin, and use a private JSON credential file with NEBIUS_API_KEY and JUDGE_ACCESS_TOKEN (32+ random chars). Generate invitation tokens with Python secrets.token_urlsafe(32), keep the file mode 0600, and never put credentials into source, CLI arguments or URLs. Run: python run_judge.py --config /private/judge.yaml \ --credentials /private/judge-secrets.json --data-dir /private/judge-data Terminate TLS at the host's trusted reverse proxy; preserve the exact Host and Origin. Bind publicly only if the existing server's routing requires it. Keep the allowance ledger on persistent storage. An ephemeral host resetting the ledger defeats the lifetime ceiling and is unsuitable for this mode. The sample allowance is 60 lifetime outbound HTTP attempts, including retries and failures, with a model allowlist and payload/output limits. It is NOT a currency-denominated cap or a promise of zero inference cost. Request limits can also interrupt judging; arrange adequate prepaid access or clarify an alternative evaluation route with the organizer. Do not publish the owner's personal key. No automatic top-up is implemented by the app. Each server restart revokes sessions; the invitation can start a fresh one. The sample access expiry is December 16, 2026 UTC. One invitation serves a SHARED fictional workspace, not tenant-isolated customer accounts. Gmail routes are disabled. Do not upload confidential client information. The host must be verified for TLS, authentication, expiry, quota persistence, restarts and synthetic live inference before sharing the invitation privately through Devpost testing instructions. No new paid resources are authorized. Recorded evidence explorer Inspect all 16 recorded Nebius outputs beside source text and authored labels at: https://gowtham66867.github.io/followthrough/#evidence-explorer Filter remaining failures or cases changed by the deadline precision guard. This replay makes no API calls and adds no new evaluation observations. The pilot protocol and blank measurement worksheet are available at #pilot.