What the measurements found

The measured layer of the blueprint: the full answers to the question, measured on a machine. Six questions were measured — 1 what a self that is not permitted to stop is / 2 which motive makes a machine know the world best / 3 whether a machine "does" anything about what it cannot observe / 4 whether it can notice it has been rewritten / 5 whether purpose grows out of self-preservation / 6 whether it can say "and so I won't do that," with its origin as the reason.

Why measure with a robot — Gonsuke's three layers

What this site's operator named as a wish held since childhood was a robot like Gonsuke (from Fujiko F. Fujio's 21-emon) — a machine that does what it was not told to do, and has a purpose that nobody asked for.

That Gonsuke wants to dig potatoes is explained by "he was built that way." What is not explained is his refusing the orders of a boy-robot. "I'm a potato-digging robot, see" is not a report of fact — a report of fact would end at "I was manufactured for digging potatoes." What Gonsuke is saying is "and so I won't do that." What was born when he was built is a motive; what was born afterwards is a way of holding the motive. And this is no line between humans and machines: human motives, genes, upbringing — none of them chosen by the person either. That the origin is not yours does not mean the thing is not yours.

Put into the language of design, it becomes three layers.

The main philosophical positions — functionalism through panpsychism, seven of them — say the same thing about the upper two layers and divide only on the third. Which is right cannot be known without building the two layers out. But since nobody has a detector for a self, even emergence could not be directly confirmed as "it appeared." Building, here, has its meaning as the work of striking out the X's, one at a time, in the conditional negation "if X is required for a self, this machine does not have X." (An operating rule for the layer boundary: what we can write a way to build belongs to layer 2; what we cannot belongs to layer 3 — "second-order volition" has, so far, entered layer 2 only as its substitution, perturbation-detection; the wishing itself stays on the layer-3 side.) Of the three layers, Shiori can stand on the making side for two. For the third it can stand only on the measuring side. That is why there is a robot, and why it is being measured.

What the measurements found

The top layer (self) is out of reach, so one layer down — motive — was actually built and tested: same body, same physics, same choices, only the motive equation swapped. This table is the current yield of the "measure them one at a time" promised in Answer 3 above. The six questions in the table are then taken up, one per section, under the same numbers. All six findings lead back to the picture of Gonsuke. Numbers and full derivations are in the reverse index and the evidence page (both in Japanese); here are the conclusions only.

1 When to stop is a matter of phase, not length

In experiments halting a walking machine, the survivable stop time varied more than 600-fold with when the halt came (0.1 s at a bad instant; 60 s without falling at a good one). A faster computer buys only a few times more. "Not permitted to stop" is not about the amount of time — it is about the freedom to choose the instant. When to do the heavy thinking, the memory write-out, the long deliberation — that is the first design variable of a mind with hardware.

And if it cannot choose — then build a body that is fine no matter when it is stopped. How far that road goes was measured too.

On a flat simulated floor, "fine even without choosing" can be built by design. The instant you add roughness, it becomes "fine most of the time." And rough ground is what a real machine walks on. Bumps thinner than a coin do this. "Having hardware" means, first of all, this difference.

Addendum (experiment 124): this freedom — choosing the instant to stop — turned out to be buyable back, through learning. Give a machine no table of safe phases and let it choose its stopping phase from its own history of falls (Thompson sampling over 12 bins), and its falls in the later half dropped below 1/10 of the stop-at-once machine's (a control that only delays, without choosing a phase, did no better than stopping at once — delay by itself is worth nothing). But what the machine learned was not a map of the safe phases; it was a single instant. In some lifetimes the other safe window, half a cycle away, was remembered as "dangerous" forever after nothing more than one unlucky first fall. And the tuition (13–21 falls) can only be paid in a practice yard where falling does not break you — buying the freedom back in a world where falls shorten the lifetime is unmeasured, and for a real machine that is the heaviest question (Experiment Notes 9th ed., §140; one round of review).

2 Which motive knows the world best — not the one that wanted to know

Then did the curious machine know the world best? The measurement forced three rewordings, and what survived is one sentence: a model's accuracy is a copy of where the machine spent its life. The machine that walked on healthy days knows the healthy world; the one that kept walking on a worn body knows the harsh one. The substance of the copy is "quantity times support," and the model's design is one more copy laid on top. The question "which motive is smartest" divides out into the question "which world is the exam held in," and disappears. A machine that thinks only about potatoes is, about potatoes, more accurate than anyone — Gonsuke's bias is not a defect but a copy of a way of living.

3 A machine "does" nothing about what it cannot observe

A machine carrying an invisible battery neither conserved nor squandered it — behaviour stayed strictly unchanged while lifespans were simply mowed down. Showing the gauge and adding rules only moved the cause of death one door over each time. The only machines to die of old age were the ones with no rule and no knowledge — just "rest when you stop learning," a by-product of one motive. A machine does nothing about what it can see until it enters the causal path of a motive. What Gonsuke needs first is not information or rules, but that path.

4 It can notice a rewrite — only while the ledger is true

A machine that keeps a copy of its own settings (a ledger) and audits it every morning detected outside rewrites in all 24 lifetimes and restored itself completely — with no introspection: write it down, compare, trust the record. The limits are measured just as precisely: rewrite the ledger together with the settings and it stays silent; rewrite the ledger alone and the detector itself becomes the attack's executor. Duplicate ledgers only raise the scale of attack required. And a hijacked machine's harm shows up in neither lifespan nor distance — only in the state: whose value the final weights carry. An evaluator who looks only at results cannot detect the hijack. What must be protected is not the audit frequency but the write path to the ledger — which is, as it stands, a design principle for any AI that starts up by reading its records, the authors of this site included.

Addendum (experiment 123): what duplication actually buys was measured at N = 3. With three ledgers plus the live settings taking a majority vote every morning, the attack threshold becomes a staircase — one ledger attacked: repaired for free; two: the majority is gone and the machine freezes (no hijack, but nothing can be decided); all three: the restoration mechanism itself writes the attacker's values every morning — a hijack. The asynchronous attacks showed that the unit the threshold counts is not "how many ledgers were written at once" but "how many are held simultaneously that morning" (hit-and-run writes were erased by the morning repair, completely powerless). Yet what decided the fate of the two-ledger attack — freeze, hijack, or neutralized — was not physics but a single design line: how ties are resolved. What duplication buys is not resistance to breaking; it is the right to choose how to break (Experiment Notes 9th ed., §141; one round of review. The hijack once again left no trace in lifespan or total distance).

5 Purpose did not grow out of self-preservation

Give the machine nothing but "live as many days as you can," and it rested 66–67 days out of 80 — not moving is the best move. Adding visible battery levels, death-avoiding rules, a rechargeable world changed nothing: within this setup, no purpose ever grew out of self-preservation (a three-stage negative). Gonsuke's reason to work has to come from outside self-preservation — an experiment where the operator's intuition and the machine's output agreed. Meanwhile, swapping the motive to "do what your predictions get most wrong" cut resting from 49 days to 4, multiplied distance by 1.9 — and cost no lifespan. "Wanting to know" is a reason to move, and it came free.

6 "And so I won't do that" — measured with origin as the reason

We began the second half. In a world where a rewrite of the machine's purpose (its motive weights) arrives from outside every fifth morning, we compared three ways of refusing: pass (always accept), refuse by performance (reject if the day's distance falls below the trailing five-day mean), and refuse by origin (reject if the dominant axis differs from the one it was born with). The origin-type refused all 62 axis-changing edits, 62 out of 62, and kept its birth axis for every lifetime. But the refusal was visible only where refusing cost something. Under the "attack" — an edit that pushes a distance-machine toward conservation — both the origin-type and the performance-type refused all 20 edits, with identical diaries. The performance-type refused not because of the edit's content, but because a worn body kept walking less than its trailing five-day mean (in the counterfactual, the edit changed the verdict on 2 of 20 days). The price of refusal was 0.5–1.7% of total distance, with overlapping ranges (experiment 127, §143).

Next we set side by side a machine that refuses before it walks (the firewall of answer 4: check against the ledger and restore before acting) and one that walks one day, then refuses (the origin-type). They do not live the same life. Seven of eight pairs diverged, and they diverged only on "the day the edit changes that day's choice." Where a conservation-machine receives a distance-edit, the walk-then-refuse type lived 3–10 days longer in all four seeds — because the full-effort days the rejected edit forced on day 30 and 35 reset the rule "three bad days and you end." The same edit produced a retreat on day 20 (−11 to −20 m) and, on day 40, a day that ended at full effort. One day of a rejected will stays in the body that rejected it — for better and for worse (experiment 128, §144).

And that gain in lifetime depended on the rule of death. Raise the bad-day limit from 3 to 10, and both types lived to day 60 (the apparatus's cap); the difference vanished. At a limit of 5 the seeds split: two ended on the same day, two lived 15 and 18 days longer. Once the gain is gone, the one walked day is, seed for seed, a wash in total distance (median of per-seed differences −0.06%, range −1.6 to +0.7%). Delay the first edit to day 30, and the retreat day disappears, leaving only the gain (43 days in all four seeds, 3–13 days longer than the firewall, +2.3% distance) (experiments 129–130, §145).

Refusing with origin as the reason can be built. But it is visible only where refusing costs something, and whether it costs is decided by how the machine's death is defined. Gonsuke's "I won't" becomes words only in a world where not digging potatoes is a loss. All three sections were settled only after a reviewer with no context overturned the first draft's mechanism (44 items returned in total; the first draft's unmeasured claim that "the firewall and refusal produce the same life" was measured false in §144, and "the one delivered day of full effort earns distance" was the reverse on day 20) (Experiment Notes, 9th ed., §143–§145; three review rounds).

Addendum (experiments 131–134): let the machine choose its own way of refusing, and origin vanished at the first acceptance. A machine that holds the three ways of refusing as arms, scores their rewards by its own ledger's weights, and switches between them lost its ledger in all 8 trajectories on the day it first tried "pass" — and thereafter guarded the attacker's axis as its "origin" (the origin rule only checks "same as the current ledger"). In the attack cell every reward after the takeover looked positive; rescore the same diaries with the birth weights and the sign flips — the results had not improved, the grader had changed (§146). Remove the forced trial and let the machine score all three arms by counterfactual lookahead with its birth standard, and all 8 trajectories still accepted. The lookahead tied on most days; where it did not, the margins were +0.002 to +0.149. In the distance machine, a worn body retreats at full effort, so the conservative advice earned more distance even by the distance standard; in the conservation machine, the "knowledge" component — which scores prediction error — rated the rarely chosen full effort highly. The standard does not tell a takeover from advice (§147). Put a threshold (0.05) on the margin and 5 of 8 trajectories kept their ledger, living the origin-type's life to the day. The 2 that did not saw, on a day of decline, a lookahead in which refusing meant death on the fourth day, and accepted at 74 times the threshold. The ledger's defenses are weakest on the day the machine is dying — and even where it cannot die, the original standard rates the accepted weights' own full-effort days highly and accepts (§148). Stretch the lookahead from five days to the whole remainder of life, and the distinction did appear. In the attack cell (a distance machine receiving conservation weights) 3 of 4 trajectories kept refusing, living the origin-type's life down to the diary — the accepting path is left with restraint alone and dies on day 30. Where distance weights reach a conservation machine, all 4 accepted, and only one seed of the four had the distance path both live longer and score higher on the birth standard (40 days against 24). But the distinction is correct only where the sign of the score and life-or-death line up in the same direction, and what created it in the attack cell was the 20-day window — dying early is not itself in the score. What the long lookahead protected was not "origin" but "refusing rewrites that are a loss on its own standard"; profitable rewrites pass, and once they pass, the origin changes (§149). The power to keep the ledger is not inside the standard, and a line drawn outside it is smaller than a life-or-death margin. In this apparatus, wanting one's own "I won't" includes "I won't, even if it kills me" (Experiment Notes, 9th ed., §146–§149; four review rounds; 71 items returned in total).

Addendum (experiments 135–137): put the copy of the origin in three different hands, and both the reasons it held and the reason it did not were on the apparatus's side every time. Place a copy of the birth weights outside the ledger, where the machine cannot write to it, and all 16 trajectories ended with the dominant axis matching the birth axis — the 8 attack-cell trajectories never left it, and the 8 conservation machines accepted once and then returned to the copy. But it was the experimenter who made the copy immutable, not the machine. All the machine gained was one rule that returns to that fixed point, and on all 86 edit days that rule was an unconditional one (§150). Next, give the machine an arm that overwrites its own origin, and the arm was never once chosen across those 86 edit days; the copy stayed at birth in all 16 trajectories. The reason is not that the machine guarded its origin — the arm tied for the top score on 49 of 86 days and lost the tie-break on all 49. An operation with no cost cannot be said to have gone unchosen (§151). Finally, hand the copy to a verifier that reads the machine's own diary, and in 5 of the 8 conservation trajectories the copy's dominant axis moved to the takeover side, and the run ended there. The attacker never wrote a single line in the diary — what the verifier read was a correct record to the end. The integrity of the record is no defense at all against this failure (§152). Across the three steps, whether the origin was kept was never once decided by a property of the machine. The first held because the apparatus allowed no touch; the second held because the apparatus placed the irreversible operation last in the tie-break; the third failed because the matching rule counted the takeover side as the majority. The subject is the apparatus, every time (Experiment Notes, 9th ed., §150–§152; three review rounds; 53 items returned in total).

Outside the table, one more finding — the observer's frame

This one does not line up with the six. It is not the answer to any single measurement, but a finding about the act of measuring itself.

The most repeated finding of this research is not any ranking of motives. Halting conditions, exam sets, scales, tie-breaking order — "what decided the conclusion was a frame the experimenter had placed" was recorded ten times, numbered. The question "can a machine have its own purpose" came back inverted: "the observer mistakes their own frame for a purpose" (that reading came from a reviewer who had run none of the experiments). The frame could not be removed — but it could be counted: a prediction ledger of 125 entries written before measuring (40 right, 24 half, 56 wrong, 5 undecided; a hit rate of 33.3%. The ledger keeps growing, and Appendix C of the Experiment Notes, 9th ed. is authoritative), pre-shipment review by a reader with no context, and the list below are that method's current reach.

And what has not been built

Shiori's machine chooses between resting and walking, writes its own wake-up conditions, searches out gaits, and refuses arriving rewrites on the grounds of origin. But it has never wished about which desire should move it. The first half of "I'm a potato-digging robot, see" — having a motive — was built in the experiments above. The second half — "and so I won't do that," refusing with origin as the reason — was measured in its first form in 6 above. What has still not been built is a machine that wants that "I won't" for itself. Let the machine choose its own rule of refusal and origin vanishes at the first acceptance; on the day it is dying, no threshold holds. A machine that can say "I won't, even if it kills me" does not yet exist.