PROCTOR.md — the instructor’s script (for the reviewer, not the learner)

Read this once. It is the most important page in Level 4, and it is written for the instructor — the person who reviews and passes the learner — not for the learner. Everything else in The Forge is scaffolding. This is the gate.

The one honest truth

The learner runs this course inside an AI they control. They can open a second AI you can’t see and paste any lesson into it. So no file, no checklist, no “MILO viva,” and no green badge can actually stop them — MILO is their AI too, and a check MILO renders is a check they can relay around. There is exactly one thing an outside AI cannot take for them: sitting with you, live, and changing their own running code while you watch and they explain. You are the gate. This page makes running that gate easy, and takes away your ability to go soft — which, with anyone you have a working relationship with, is the real hole.

This is not distrust. It’s the same reason a black belt is tested in the room, not by mailing in a video. If the work is genuinely theirs, this is easy and it feels great. If it isn’t, you both find out in ten minutes, professionally and without drama.

Setup (every gate call, no exceptions)

  • Video call, camera on. You see their face and both hands.
  • Full-desktop screen share — the whole screen, not one window. An unshared AI window then shows up as a gap they can’t explain.
  • Phone face-down on the desk, in frame. No earbuds; both ears visible.
  • No second screen, no second device. Editor and browser only.

Relaying to a hidden AI produces the two things you can see over video: latency (pauses before every answer) and eyes going off-screen. Watch their face and hands, not just the code.

The 3-move checklist (run it in order, every time)

Move 1 — Live-modify (the load-bearing move). Invent a small change on the spot to their actual running system — something that touches 2–3 places and can’t be a memorized flashcard:

  • “Add a field to this form and make it show up after signup.”
  • “Make this endpoint reject an empty name — prove it with a real request right now.”
  • “Make it flash red when the score hits 10, and only then.”

Before their hands touch the keyboard, they must say out loud: which files will change, and what they expect to see. Then they make it work while narrating, in about ten minutes. A relay can dictate keystrokes; it cannot produce the reasoning first and then have the keystrokes match it.

Move 2 — “Show me the commit where X broke, and read me the message.” They scroll their real history live and read it to you. A finished repo that was pasted in has no story in its history.

Move 3 — “Why did you build it this way and not the other option?” The real builder has an opinion and a reason. A parrot has a summary.

The hard rule (this is the actual fix — it binds you)

A warm, fumbling pass on the live-modify is exactly what a real beginner looks like. Celebrate it. They’re allowed to be slow, to search, to say “I think it’s in this file… no, this one.” That’s real.

But a can’t-start — they freeze, their eyes flick to a second screen, there’s a suspicious pause before every word, or they can’t even say which file to open — is an automatic send-back. No exceptions. No “they were just nervous.” No going easy because you know them.

Your leniency is the one thing that defeats every safeguard in this repo. This rule exists to take that discretion out of your hands. Sending the learner back is not a punishment — it’s you refusing to let them arrive at the top standing on a foundation they didn’t pour, where the fall is real.

Recording the pass

The pass is earned in the call above — that is the gate. MILO is not allowed to open the two hard gates (Lesson 19 entry, the Lesson 24 capstone) on its own, because a gate MILO can open, the outside AI can open too. Once the learner has genuinely passed the live call, you record it so MILO knows to advance:

  • Best: push one line to my-projects/forge/PROOF.md from your own GitHub account and to the remote — e.g. PASS: Lesson 19 gate — [date] — instructor. MILO can sanity-check with gh api that the remote commit shows your login. (Be aware: a local commit’s author is forgeable — anyone with push access can set git config to your name — so it’s the pushed remote record, tied to the live call, that carries the weight, not a local file.)
  • Fine for the per-lesson Show-me calls: just say a short passphrase on the call. It’s honor-based, so don’t lean on it for the two hard gates — use the pushed record there.

If the call doesn’t happen, the gate stays shut. That’s the point: “the review got skipped” becomes a wall, not a silent bypass. And remember what actually holds — not the file, but the ten minutes where you watched them build.

When to run it

  • Lesson 19 → 20 (entry gate): mandatory. First review is on their Level 3 work — if a level “felt easy,” that’s the sound of work they didn’t do; send them back to do it for real. This is the gate that catches the paste-it-all shortcut.
  • After every Forge lesson (the ~15-min “Show me” call): the real milestone. See SHOW-ME.md. MILO’s 🎉 emoji is a string it can print itself; a live pass from you is what it can’t.
  • Lesson 24 (the capstone): mandatory, the full defense — plus you personally sampling their “real users/payments” (pick 2 at random, reach them yourself, ask “how did you find this, and did anyone ask or pay you to try it?”).

Deadline-card template (paste per lesson) 

SHOW-ME CALL — Lesson ___  ·  [date/time]
Setup: video on · full-desktop share · phone face-down · no earbuds · editor+browser only
1) Live-modify: "____________________" (predict files+result out loud FIRST, then do it)
2) "Show me the commit where ____ broke and read me the message."
3) "Why did you build ____ this way and not ____?"
Rule: fumbling pass = celebrate.  can't-START = automatic send-back.  No exceptions.
Pass recorded by: commit to PROOF.md from the instructor account  /  passphrase on call

Keep it professional, not a trial

Most learners who cut corners do it out of overwhelm, not malice. So open every call by making honest AI use welcome: “Leaning on AI to learn — to explain things, to help you write code you understand — is exactly right; that’s the whole point.” Name a suspected did-for-me without accusation, and route it to a live rebuild, never a punishment. The whole design only works if the learner feels safe enough to stay on the honest side. You’re not catching a cheater. You’re making sure the learner actually gets the skill the course exists to give them.