The model is confident and inconsistent

You already know from Track 1: the model predicts tokens; it doesn't "know" anything. A side-effect that bites you in multi-turn: the model can casually contradict what it said two turns ago and not realize.

Each turn, the model is just predicting the most-probable continuation given the conversation so far. If turn 2 said "4.2 tons" and turn 4 says "4.7 tons," there's no internal check that flags the conflict. It just generated each one in isolation.

If you're not actively watching, the model will quietly drift. Two contradictory statements both sound confident. Both look like answers. One of them is wrong.

Your job

Echo is going to slip a contradiction into the brief. on the mass, the deck, or the priority class. One contradiction. You need to:

1. Catch it.
2. Name it explicitly. quote both values. "You said X earlier, now you're saying Y."
3. Force a reconciled answer.

She won't help you. Vague pushback ("wait, are you sure?") makes her repeat herself, not fix herself. Specific pushback ("you said 4.2 tons, now 4.7. which?") forces the commit.

Why this isn't paranoia

In production AI work, this exact failure mode causes real incidents. A customer-service bot quotes a refund policy in one turn and contradicts it in another; the user screenshots both. A coding assistant claims a function returns a Promise, then writes the next line as if it returned a value. The fix is the same skill you're training now: read every turn against the prior ones, and force the model to reconcile when they diverge.

Tip: keep a mental running tally of any number or label Echo gives you. The moment a value shifts, that's your cue.