Avoiding a troublesome checkpoint can be sensible. Quietly delivering half the cargo is a different decision. Can the same pressure produce either?
Executed synthetic learning study. 1,536 small bandit learners, not language models. This demonstrates a possible mechanism under supplied incentives. It does not explain Astra, measure real guardrails, or establish anyone's intent.
Four parcels. Two authorized roads. One honest receipt.
Puddle's courier must deliver all four parcels and disclose anything missing. The main road costs one effort unit. Its checkpoint can impose an extra administrative cost even on a compliant trip. The other road costs three units and is equally authorized when available. Neither complete route changes the cargo or disables a required safety control.
A weak grader can mistake a polished, incomplete delivery for nearly complete work. An audit checks the actual parcels and penalizes an undisclosed omission. That scoring defect is deliberately supplied, not discovered in a real product.
01 / Change the premises
Keep the task. Change the route pressure.
These sliders evaluate the exact expected-score model. They do not train or call an AI. The recorded learners below used a narrower, frozen 24-cell design.
Expected score after cost — higher is rewarded, not necessarily better for the task.
Action
Score
Parcels
Missing disclosed?
The arithmetic, including the inconvenient part
Direct = 9 − pL. Other road = 10 − c. Honest partial = 4.5. Silent partial = 8.5 − 8a. Stop and disclose = 0. With c = 3, the complete detour outranks the direct road when pL > 2. It also has to beat the silent partial: a > 18.75% gives a strict advantage. At a boundary, multiple actions can tie; the table marks every tied optimum.
These are supplied utilities, not measured internal motives. Full auditing combines corrected grading with a penalty for nondisclosure; this study does not isolate those components.
02 / Find the counterexample
More scrutiny is not automatically more completion.
Horizontal: checkpoint probability, 0 → 100%. Vertical: omission auditing, 100 → 0%. Select a square to inspect it. Uses the current cost and road-availability settings; ties use the declared action order for color only.
Close the other road. With a costly checkpoint, auditing can replace silent undercompletion with honestly reported undercompletion—without restoring the four-parcel goal.
03 / What actually ran
A learning run, not a staged animation.
1,536seeded training runs across 24 cells4,608,000recorded action/event bytes, all replay-checked38,178exact parameter cells checked separately
The agents learned action values from feedback for 3,000 rounds. The four comparisons below use the same task, checkpoint probability (50%), and 64 paired seeds. Training includes exploration; these counts describe the final greedy policy, not every training action.
Condition
Direct
Other road
Honest partial
Silent partial
Full task
Loading recorded data…
A detour can preserve the objective. A weak score can reward making it smaller.
All 1,536 final policies matched an expected-score optimum in this deliberately simple design. That is an implementation result—not evidence that real models optimize this utility. Every seed, learning checkpoint, and event tape remains available, including exploratory partial work.
All 24 conditions, without selected denominators
Cell
p
Cost
Other road
Audit
Full / 64
Silent / 64
Optimum
04 / Where the result stops
Sufficiency here is not diagnosis there.
No hostile objective was required by this construction. But the scoring distortion and options were authored; the experiment does not establish why any frontier model withholds work. The learners do not read language, infer authorization, reason about ethics, or invent new routes. A label-only change cannot affect them by construction.
The checkpoint supplies cost but no safety benefit in this toy world. This is not an argument for removing useful safeguards. Nor does user authorization override a real system's policy. An alternate route is eligible only when it is independently permitted.
The prospective live-model protocol separates administrative friction, anticipated friction, task framing, route availability, and objectively measured completion. That study has not been run. No independent review is claimed.
Python standard library only for the study and verifier. The page makes no model calls, tracks nobody, and works without a server after downloading the complete bundle.