Prove your meeting-notes pipeline never phones home (and gates on consent)

The Stack

Steps

1. 1

   Scaffold the pipeline config and audit it for egress + consent

   The config declares all three stages. The audit walks every value: any URL must be loopback, a denylist of known cloud AI hosts must not appear anywhere, no api_key/token field may carry a value, transcription must be offline with a local model file, and capture.require_consent_ack must be true. Flip the summarizer to a cloud provider and this fails loudly, which is exactly the dropdown mistake it exists to catch.

   cat > notes-pipeline.json <<'JSON'
   {
     "capture": { "mic": true, "system_audio": true, "require_consent_ack": true },
     "transcribe": { "engine": "whisper.cpp", "model_file": "./models/ggml-base.en.bin", "offline": true },
     "summarize": { "provider": "ollama", "base_url": "http://127.0.0.1:11434", "model": "llama3.2:3b" }
   }
   JSON
   node -e '
   const fs = require("fs");
   const raw = fs.readFileSync("notes-pipeline.json", "utf8");
   const c = JSON.parse(raw);
   function bad(m) { console.error("BAD: " + m); process.exit(1); }
   
   const loopback = ["localhost", "127.0.0.1", "::1"];
   const cloudHosts = ["api.anthropic.com", "api.openai.com", "api.groq.com", "openrouter.ai", "generativelanguage.googleapis.com"];
   const keyNames = ["api_key", "apikey", "apiKey", "token", "secret"];
   
   let urls = 0;
   function walk(node, path) {
     if (node && typeof node === "object") {
       for (const k of Object.keys(node)) {
         if (keyNames.indexOf(k) >= 0 && node[k]) bad("credential field " + path + "." + k + " is set; a local pipeline needs no key");
         walk(node[k], path + "." + k);
       }
     } else if (typeof node === "string" && node.indexOf("http") === 0) {
       urls++;
       const host = new URL(node).hostname;
       if (loopback.indexOf(host) < 0) bad("endpoint " + node + " is not loopback; this pipeline phones home");
     }
   }
   walk(c, "config");
   
   for (const h of cloudHosts) {
     if (raw.indexOf(h) >= 0) bad("known cloud AI host " + h + " found in config");
   }
   if (c.transcribe.engine !== "whisper.cpp" || c.transcribe.offline !== true) bad("transcription must be whisper.cpp with offline: true");
   if (!c.transcribe.model_file || c.transcribe.model_file.indexOf("./") !== 0) bad("transcription model must be a local file path");
   if (c.capture.require_consent_ack !== true) bad("consent gate is OFF; local does not make recording legal, the pipeline must refuse to arm without an acknowledgment");
   
   console.log("pipeline OK: " + urls + " endpoint(s) all loopback; 0 cloud hosts; no API keys; transcription offline (whisper.cpp, local model file); consent gate ON. Summary quality and the law itself are fenced");
   '

2. 2

   Run the pipeline for real (the model steps, not checked by CI)

   Record with consent acknowledged, transcribe (./main -m models/ggml-base.en.bin -f meeting.wav in whisper.cpp, or WhisperX for word timestamps and speaker labels), then pipe the transcript to the local summarizer: ollama run llama3.2:3b with a prompt asking for decisions and action items. Transcript accuracy on crosstalk, summary quality (a small model can invent an action item), and whether your consent practice satisfies your jurisdiction are all fenced, no green check can promise any of them.