CSE Roadmap

Common tier · about 4 hours · 11 parts

Part 5 of 11 · 4 min read

What the example found

Outcomes under five policies, false completions as over-claiming or verifier error rises, and the stall-threshold trade-off.

Setup: five subgoals. 15 percent of tasks cannot get past subgoal 3. The budget is 30 steps where one applies, otherwise a hard cap of 100. There are 20,000 tasks per policy, seed 0.

How to read the policy table: each row is one of five stopping policies, run on the same task distribution.

Policy Finished Said done, was not Escalated rightly Escalated wrongly Stopped by budget or cap Mean cost Wasted cost Cost per success
agent says done 59% 40% 0 0 0.5% 21.8 10.1 36.8
agent says done + budget 49% 30% 0 0 21% 17.8 10.8 36.3
verified + budget 64% 0.9% 0 0 35% 20.3 10.6 31.7
verified + stall detector 77% 0.8% 15% 7.5% 0 21.8 6.7 28.3
verified + stall + budget 76% 0.8% 13% 7.0% 3.5% 21.6 7.0 28.4

What to notice: adding a budget on top of self-report lowers false-done from 40% to 30%. But that is only because 21% of tasks now end as a plain timeout instead. The budget is not catching false claims. It is just cutting runs off before the agent gets the chance to make one.

How to read the over-claim sensitivity: each value is the false-completion rate under the self-report policy at a different assumed over-claim rate per step.

False completions as the agent over-claims more (agent says done): 0 percent at no over-claiming, 25 percent at 0.02 per step, 41 percent at 0.05, 55 percent at 0.10.

What to notice: the false-completion rate tracks the over-claim rate almost linearly (25% at 0.02, 41% at 0.05, roughly doubling as the input roughly doubles). Self-report's failure rate is essentially a direct pass-through of however untrustworthy the underlying agent's claims happen to be.

How to read the leaky-verifier sweep: each value is the false-completion rate under verified+stall+budget at a different assumed verifier error rate.

A leaky verifier (verified + stall + budget): false completions are 0 at a verifier error of 0. They are 0.7 percent at 2 percent, 3.4 percent at 10 percent, and 9.6 percent at 30 percent.

What to notice: even a verifier wrong 30% of the time still holds false completions to 9.6%, far below self-report's 40%. A flawed verifier is still categorically safer than no verifier. It only approves the fraction of false claims it happens to miss, not every claim it sees.

How to read the stall-threshold table: each column is a different number of idle steps required before the monitor triggers a replan.

Steps without progress before a replan 2 3 5 8 12
Escalated although the task was possible 30% 15% 7% 2% 0
Escalated impossible tasks 15% 15% 13% 6% 0
Mean cost 14.2 17.6 21.5 24.5 25.7
Correct outcomes 0.70 0.84 0.89 0.81 0.71

What to notice: correct outcomes peak at threshold 5 (0.89) and fall on both sides of it. Too tight wrongly escalates possible tasks. Too loose fails to ever escalate the genuinely impossible ones. So the threshold is a real sweet spot. It is not a dial where "more cautious" or "more patient" is simply better.

How to read the impossible-task-share comparison: each pair is correct outcomes for budget-alone against budget-plus-stall-detection, at a different share of genuinely impossible tasks.

The share of impossible tasks (correct outcomes, verified + budget against verified + stall + budget): at 0 percent it is 0.75 against 0.89. At 15 percent it is 0.64 against 0.89. At 40 percent it is 0.45 against 0.88.

What to notice: budget-plus-stall-detection's correct-outcome rate barely moves as impossible tasks rise from 0% to 40% (0.89 to 0.88). Budget-alone collapses (0.75 to 0.45). Stall detection is not just an improvement. It is what makes the system's accuracy independent of how many of its tasks turn out to be unsolvable.

{{svg:loop|Goal specification, agent step, monitor, and four decisions: continue, replan, verify and stop, escalate.}} {{svg:results|Outcomes under five policies, false completions against the over-claim rate, and the effect of the stall threshold.}}

In the simulation the monitor does three things after every step. It checks whether a new subgoal was completed. If the agent claims to be done, it asks the verifier. If there has been no progress for stall_k steps it replans. After more than two replans in a row without progress it escalates. The replan counter resets when progress resumes.