hawkes.aiKey ideas
- The one idea — R1 physically can't end its turn on a cop-out. Seven mechanical layers — a 14-phrase cop-out scan, an open-plan-item check, a commit-body check — bounce the model back into the loop until the work is actually finished. Machine-mechanical, not prompted; only the operator can override.
- How it is shown — The end-turn gate catching "I'll leave the rest as an exercise" and bouncing the model back into the loop with a forcing message.
- The trap to avoid — Thinking you can just prompt a model to finish. Prompts are soft — the last act proved that. This is a hard, deterministic gate on the act of stopping.
- What it sets up — R1 proves criteria and refuses to quit early — but who checks the finished work? crucially, not the worker itself. a different brain.
This agent runtime scans every attempt to stop against a fourteen-phrase cop-out list. Land on one, and the machine bounces the model straight back into the loop.
The one idea
R1 physically can't end its turn on a cop-out. Seven mechanical layers — a 14-phrase cop-out scan, an open-plan-item check, a commit-body check — bounce the model back into the loop until the work is actually finished. Machine-mechanical, not prompted; only the operator can override.
"I'll leave the rest as an exercise for the reader." R1 physically cannot end its turn on that sentence. Before it can stop, a mechanical gate inspects the attempt, and a cop-out doesn't get through — not because we asked nicely, but because code guards the finish line. Here's how that gate is built. The problem is that models love to truncate. "Leave the rest as an exercise." "Implement the remaining functions similarly." "The pattern continues from here." Each is a smooth, confident sentence that quietly abandons half the work. It sounds like a natural stopping point — and that's what makes it dangerous. The abandonment is disguised as completion. The defense is an end-turn gate. Before the model can stop, a mechanical checkpoint intercepts the attempt — not a line in a prompt, but actual code that inspects the work before permitting any exit. The act of stopping is made conditional. The gate runs seven layers of checks; three tell the story.
How it works — the demo
The end-turn gate catching "I'll leave the rest as an exercise" and bouncing the model back into the loop with a forcing message.
Take three of the seven. A fourteen-phrase cop-out scan sweeps the output for stock abandonment lines. An open-plan-item check cross-references the SOW tree for unfinished nodes. A commit-body check reads the actual commit to confirm the claimed work is there. Not vibes about whether it feels done — hard tests for whether the job is finished. Fail any check, and the gate bounces the model back into the loop: the work isn't done, keep going. It literally cannot stop until the checks pass. And this is crucial — it's machine-mechanical, not a polite instruction. A model can rationalize its way around "please finish." It can't rationalize its way around a gate that simply won't open. The trap: thinking you can just prompt a model to finish. Prompts are soft — the whole last act proved that — and a determined truncator rationalizes right around them.
The trap to avoid
Thinking you can just prompt a model to finish. Prompts are soft — the last act proved that. This is a hard, deterministic gate on the act of stopping.
Why it matters — and what’s next
R1 proves criteria and refuses to quit early — but who checks the finished work? crucially, not the worker itself. a different brain.
This is a hard, deterministic gate on stopping: Act nine's structural-constraint idea, pointed at end-turn. One way through without passing: the human operator can override. The model never can. So R1 refuses to say "done": seven mechanical layers bounce it back until the work is finished — machine-mechanical, not prompted, only the operator overrides. It proves its criteria and won't quit early. But who checks the finished work? Not the worker itself — it grades nothing of its own. A different brain does. Next group.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:45 of narration
"I'll leave the rest as an exercise for the reader." R1 physically cannot end its turn on that sentence. Before it can stop, a mechanical gate inspects the attempt, and a cop-out doesn't get through — not because we asked nicely, but because code guards the finish line. Here's how that gate is built.
The problem is that models love to truncate. "Leave the rest as an exercise." "Implement the remaining functions similarly." "The pattern continues from here." Each is a smooth, confident sentence that quietly abandons half the work. It sounds like a natural stopping point — and that's what makes it dangerous. The abandonment is disguised as completion.
The defense is an end-turn gate. Before the model can stop, a mechanical checkpoint intercepts the attempt — not a line in a prompt, but actual code that inspects the work before permitting any exit. The act of stopping is made conditional. The gate runs seven layers of checks; three tell the story.
Take three of the seven. A fourteen-phrase cop-out scan sweeps the output for stock abandonment lines. An open-plan-item check cross-references the SOW tree for unfinished nodes. A commit-body check reads the actual commit to confirm the claimed work is there. Not vibes about whether it feels done — hard tests for whether the job is finished.
Fail any check, and the gate bounces the model back into the loop: the work isn't done, keep going. It literally cannot stop until the checks pass. And this is crucial — it's machine-mechanical, not a polite instruction. A model can rationalize its way around "please finish." It can't rationalize its way around a gate that simply won't open.
The trap: thinking you can just prompt a model to finish. Prompts are soft — the whole last act proved that — and a determined truncator rationalizes right around them. This is a hard, deterministic gate on stopping: Act nine's structural-constraint idea, pointed at end-turn. One way through without passing: the human operator can override. The model never can.
So R1 refuses to say "done": seven mechanical layers bounce it back until the work is finished — machine-mechanical, not prompted, only the operator overrides. It proves its criteria and won't quit early. But who checks the finished work? Not the worker itself — it grades nothing of its own. A different brain does. Next group.