The Video Assistant · Lesson 1 — The Check ← Course

The check — how a model decides.

An LLM — the AI behind tools like ChatGPT — works like a football VAR.

In football, when a goal is scored but something looks off, the referee sends it to the VAR — the video assistant referee — who studies the replays. Did the move begin with an attacker offside? Did the whole ball cross the line? Did it brush an arm in the build-up? Was there a foul? To decide, the VAR weighs several clues at once: the semi-automated offside line, goal-line technology, the handball check, and the build-up replay.

No single clue settles it. The VAR weighs them all together and always reaches a call — GOAL or NO GOAL — confident when the clues agree, only just leaning one way when they don't. (A real VAR can leave it to the on-field decision; our model can't — a language model always answers.)

A large language model (or LLM) reaches its decisions the same way — weighing everything it learned in training and committing to a call. Its call happens to be a single word: given the text so far, it puts odds on every word that could come next and commits to one, then repeats for the word after that. The same instinct as the VAR, made one word at a time.

Check the three goals and watch the VAR weigh the evidence — then see what happens when the clues disagree.

Send a goal to the VAR. Watch the review weigh the evidence.
offside line
Evidence · the VAR weighs all of it
Offside —·
Goal-line—·
Handball —·
Build-up —·
Read by a VAR shaped by seasons of past football — that experience is baked in, not looked up.
AWAITING DECISION

You hand over the inputs. The VAR decides.

You don't write the weighing — you assemble the evidence and ask. The VAR weighs it with an eye shaped by seasons of football, and hands back a call, along with the odds it put on that call. 

# Hand the VAR the evidence — you don't code the weighing.
decision = var.review(
    offside="level — onside",
    goal_line="ball fully over",
    handball="no arm contact",
    build_up="clean move",
)

# decision.call → "GOAL"   ·   decision.confidence → 0.96

Weigh everything, commit to one call.

Three finishes, three different bundles of evidence. The offside line alone didn't decide it — the VAR weighed it against the goal-line read, the handball check, and the build-up, and always committed: a confident call when they agreed, a barely-there one when they didn't. It never left the decision unmade.

An LLM chooses its next word the same way. It weighs the words so far, puts odds on every candidate, and commits to one — sure when the odds pile onto a front-runner, only just when they're split. Then it repeats, word after word.

Quick check · make the call
The offside, goal-line and build-up reads all conflict on a tight one — the odds are nearly even. What does the model do?
Next: the odds →
▸ Go deeper — how the VAR actually decides optional

It's a distribution, not a single verdict

When the VAR renders a call, it isn't picking from a fixed menu of two answers. Underneath, the model produces a probability for every possible next token, and the "decision" you see is simply the most likely one. The confidence number in this lesson — 0.96 for the clean strike, 0.54 for the tight one — is exactly that: the share of probability sitting on the top option. A tall, lonely spike means the model is sure; a near-even split across options means it isn't, which is why the tight call comes back as a barely-there call rather than a confident one.

Weighed against training, not looked up

Where do those probabilities come from? From patterns the model absorbed across an enormous amount of text during training — the VAR's accumulated experience, baked into its weights. It is not consulting a database at decision time, and it has no specific past match it can replay. (Looking things up live — pulling the actual archive footage — is a separate capability called retrieval, and it gets its own lesson later.) This is the difference between judgement shaped by experience and checking the records.

Why "probabilistic" matters

Two consequences follow, and both become later lessons. Because the answer is sampled from a distribution, the same check can produce different calls depending on how greedily you pick — that's the temperature dial. And because the top option is merely the most probable, not the verified truth, a confident-looking answer can still be wrong — that's the root of hallucination.