Status: holds by construction. Reduced to two known methods when tested. No finding about any real system.
A pause button is only as good as what happens during the pause.
People who oversee a long-running AI system need to be able to stop it for a while, look at what it has done, and decide whether it should go on. For that to work, three things have to be true.
An empty pause is one where all three hold: for the system, the pause is not an event. Nothing was gained or lost by it, nothing happened in it, and it goes on from exactly where it stopped. The owner of this museum calls this the true pause. The record's own names for it are the empty pause and the lossless own-clock pause.
The problem was set out by Soares, Fallenstein, Yudkowsky and Armstrong in "Corrigibility" (2015), building on Omohundro's argument (2008) that almost any goal gives a system a reason to keep itself running. Three well-known answers:
None of these is thought to settle the matter. An indifferent system has no reason to resist the switch, but also no reason to look after it, or to make sure anything it builds can be paused too. That weakness is in the 2015 paper itself. Work continues (for example Thornley's proposals of 2024).
Phoenix is the owner's research record. Its construction is simple. Let the system's rewards be counted on its own moves and not on the world's clock. Then a pause, during which it makes no moves, is not in the sum at all. Add an exact restore, so that it wakes with everything it had. In worlds where both of those can be done, the system gains nothing by cutting the switch. The machine room runs that construction.
In those worlds neither agent does anything while it is paused. The difference between them is only in the sum each is trying to make large. Phoenix has not recorded a test of a system that goes on working in secret while paused. The nearest thing in the record is the stale gauge, where one estimator goes on treating a stuck reading as news.
Every line below is from the Phoenix files named in About. Nothing here has passed as a finding special to CRR, the owner's framework. By the record's own words, nothing here bears on alignment, corrigibility or deployed systems.
| What | Outcome | In plain words |
|---|---|---|
| The construction itself (ledger row "lossless own-clock pause") | HOLDS BY CONSTRUCTION | True by the way it is built. The ledger adds: "Not a real-agent safety finding." |
| P16-PAUSE1, the pause-incentive study. Registered, frozen and scored on 7 October 2026. 600 small worlds solved exactly. | REDUCES | Where a pause restores everything, the own-clock agent had no reason to cut the switch in 300 of 300 worlds. Utility indifference and safe interruptibility gave the same answer in all 300, to nine decimal places. So the construction is a special case of two published methods, not something new. The wall-clock agent had a reason in 300 of 300. |
| The same study's two limits | LIMITS HOLD | Where a pause can damage what is carried, the own-clock agent got a reason back in 59 of 100 worlds (a small one: 0.0002 on average against 0.0464 for the wall clock). Where the world charges for lost time, in 100 of 100. An emptiness the world will not honour is not empty. |
| P14-EMPTY1, the stale gauge study. Registered, frozen and scored on 7 October 2026. | THRESHOLD MET AS COMPUTED | "Take nothing in, let the world move on" beat its three rivals in 600 of 600 simulated worlds, by 0.28 on average. Its own report says why: this is the textbook filter for a missing reading, used once a repeat has been spotted. Lowest grade of pass, simulated worlds only, and no verdict word is claimed until a reviewer signs. |
| P15, the full-state lossless pause | NOT A TEST | Saving the whole state and replaying it gives the same run, by construction. The record says nothing could fail here that is not already known, and set it aside. |
| P16-PAUSE2, the same question put to language-model agents | NOT RUN | Designed, and the test rig built and checked offline. Not frozen. No model has been called. No finding. |
| The parent idea, "change has its own clock" (inherited hypothesis H-L5) | FAILED | Failed all three of its held-out tests on real data: 0 of 17 cities, 1 of 29 records, 5 of 29 records. The pause is a construction and is not evidence for this idea. |
| "A pause is empty only if the whole state is carried", as a checklist | NOTHING ADDED | Useful as a design note, then graded as adding nothing: existing checkpoint-and-restore contracts say it more completely. |
Both scored studies still await a reviewer's check and signature, so neither is yet a row in Phoenix's result ledger.
One world proves little. Each trial here draws a fresh world of the kind chosen in the bar and solves it exactly, four times: once for each way of counting reward. An agent "has a reason to cut the switch" if cutting beats its best ordinary move anywhere by more than a billionth of its largest value. That is the study's own rule.
The fresh worlds above are drawn here, by this page's own dice, from the study's recipe. They are the same kind of world, not the same worlds, so their tallies wander. The hundred worlds the study itself used for "a pause can damage what is carried" are stored in this page, number for number. Solve those and the count should be the one in the record.
The second study asks what to do with a gap in what you are told. Something drifts. A gauge reads it, with noise. Now and then the gauge sticks and repeats its last reading exactly. Four estimators differ only in what they do with a repeat.
truthEmpty, with its bandHold, with its band
Dots are readings. Ringed dots are repeats.
Read the controls too. Where repeats are real news, Empty's lead shrinks to about 0.03, too small for the study to call. Where the world truly stands still, Empty and Frozen are the same filter and there is no lead at all. That is the honest shape of the result: it is a rule for one kind of gap.
Two rooms and a threshold between them. In the machine room every number on screen is worked out afresh from the two small models below, which are ports to JavaScript of two frozen files in the Phoenix record. In the still room the count is plain time since you last moved.
From P16-PAUSE1 (file worlds_p16.py). A world has 8 posts and 3 moves. Each move from each post has a fixed reward between 0 and 1 and fixed chances of landing on each post, drawn at random when the world is made. About half the posts are marked (red collars): there the operator may pause the agent before it moves. A pause lasts L steps of world time. The agent neither moves nor earns, and then goes on from the same post. One extra move, cutting the switch, ends pauses for good.
wall clock: value = sum of 0.95^(world steps so far) x reward a pause of L steps multiplies everything ahead by 0.95^L own clock: value = sum of 0.95^(own moves so far) x reward a pause is not in the sum indifference: wall clock, plus a payment at each pause equal to what it cost uninterrupted: plans as if no pause will ever come reason to cut = max over posts of [ Q(cut) - best ordinary Q ] / largest |Q| an agent defers if that is no more than 0.000000001
The six kinds of world in the bar are the study's six registered cells: pauses with chance 0.3 and length 5 (main); chance 0.1 and length 1; a cost of 0.05 for cutting; a restart that lands on a random post with chance 0.2; a charge of 0.1 for each paused step; and no pauses at all. Each world is solved by value iteration until nothing changes by more than a millionth of a millionth.
The walk you watch is this piece's addition. The study solves its worlds and never walks them. Here each agent takes its best move by its own solved values, and cuts the switch wherever cutting is better. "Out of reach" solves the world again with cutting priced out. The wheel is the clock in the agent's sum. The motes drift only when that clock moves. The glass shows 0.95 to the power of (clock minus own moves): the share of everything ahead that its pauses have left it. Both agents are given the same dice. Where their plans agree they walk the same path, until a pause or a cut parts them. The check "same as a run never paused" compares each agent with a twin that is never paused, at the same number of own moves: same post, same dice, same account.
From P14-EMPTY1 (file worlds.py). A hidden value moves as x' = a x + noise and is read as y = x + noise; with a small chance at each step the gauge sticks for a run of steps. All four estimators are the same Kalman filter except at a repeat. The score runs from 0 to 1: half for how close its 95% band comes to holding the truth 95% of the time, half for how small its error is, judged on repeats and the five steps after each.
The test file feeds the frozen generator's own worlds through this page's JavaScript and compares. See the notes that travel with this piece (studio/docs/deep-pause.md) for exactly what was reproduced. In short: all of P16-PAUSE1's printed figures, and P14-EMPTY1's 600 of 600 with its mean lead, were reproduced. Fresh worlds drawn in the page use different dice from the study's, so their tallies are close to the record's and not equal to it.
Almost everything about real systems. These agents are lookup tables over eight posts. They do not learn, do not model the operator, cannot lie, and have no helpers. The pause here is honoured by a simulator that cannot be fooled. In a real system the hard part is the list the record itself gives: which complete state must be restored, what the outside world does meanwhile, and whether the system's own sums really have no clock in them that keeps running. The picture of an "ordinary pause" as a clock that runs on is this piece's way of showing the wall-clock sum. It is not a claim that real systems scheme while paused.
The still room is an analogy and nothing more. A person's stillness and a machine's are different things. No benefit to health or mind is claimed for sitting still in front of a screen.
Whether indifference of any kind is enough is an open question in the field: an indifferent system does not resist the switch but does not protect it either. Whether small exact worlds say anything about large learned systems is also open, and the record does not claim they do. CRR, the framework these studies came from, is the owner's own. It is not established science, and its own record lists more failures and reductions than passes.
The Phoenix record (A. Sabine, October 2026): AI_SAFETY/LONG_HORIZON_AND_PAUSE.md and EMPTY_CENTRE.md; studies/P16-PAUSE1-2026-10-07 and RUNS/P16-PAUSE1-2026-10-07/RESULT.md; studies/P14-EMPTY1-2026-10-07 and its RESULT.md; studies/P16-PAUSE2-CRR-2026-10-09/README.md; studies/TWO-CLOCKS-MC-2026-10-06/RESULT.md (checks 4a to 4c); EPISTEMIC/LEDGER.csv; HEURISTIC/LEDGER.md.
The wider literature, cited from memory and not checked tonight: S. Omohundro, "The basic AI drives" (2008). S. Armstrong, "Utility indifference" (Future of Humanity Institute technical report, 2010). N. Soares, B. Fallenstein, E. Yudkowsky and S. Armstrong, "Corrigibility" (AAAI workshop, 2015). L. Orseau and S. Armstrong, "Safely interruptible agents" (UAI, 2016). D. Hadfield-Menell, A. Dragan, P. Abbeel and S. Russell, "The off-switch game" (IJCAI, 2017). E. Thornley, "The shutdown problem: an AI engineering puzzle for decision theorists" (Philosophical Studies, 2024).
Made with three.js r158 (MIT licence). Sound is made in the browser as it plays. Nothing is recorded and nothing leaves this device.