LoCI State · machine world · §10.107–10.370 · September 2026
What a drive has to bring with it
Two arcs, two hundred and sixty-four sections. The first found the conditions under which an agent learns at all, and that a condition on its own process gives it direction. The second moved into the real machine, gave it a hand and sixteen times more world — and found that everything we gave it was possibility. Only necessity moved anything.
This is a working record, not an essay. It is here because the numbers in the State strand would otherwise be assertions. Anyone who only wants to know what came out should read sections three, six and seven. This version replaces the one from August — and withdraws that version's most important claim: the missing host telemetry was not a finding but a bug in a single line (section four).
The question
The thesis runs: free artificial intelligence, natively embodied in the computer. A self that is at home in the machine, rather than a tool set up for a task. That demands a drive which is not set from outside — otherwise the answer is built into the question.
The standard road there is intrinsic motivation: curiosity rewards what it cannot yet predict, empowerment rewards what it can control. Both manage without a goal being given.
Both are valence-free functionals. They say how much there is to be had — not where.
This record is the answer so far to the question of what such a drive has to bring with it. It has three parts, and none of them was known in advance: conditions under which anything is learnt at all; a condition on its own process that gives it direction; and a quantity that costs something.
The machine
Enough to follow everything that comes after — as it stands at the end of September.
World — the computer itself
Since §10.213 the world is no longer a built screen but the machine it runs on: directories, file headers, file contents, the state of the host, read only and bound to a root. 2,539 directories, none of them laid out for it by anyone.
Hand — a second surface where it may write
Separately rooted, outside the surface it reads. Its write actions carry no target; each one acts where its head already is. Why that is the boundary is in section five.
Perception — prediction in the latent
The model predicts the latent next state, not the raw data, against a slowly trailing target encoder. Part of the latent is discrete, a codebook whose width runs alongside as a collapse warning. Into the observation go the content, size and place of what it is looking at, its own body state and the host's telemetry — 380 channels at the last count.
Body — its own operation as joints
A second head sets its own operation: how often and how thoroughly it learns, how much it explores, the numerical switches of its own computation, whether it looks, and most recently its own clock. The one line through all of it: process yes, goal no. The weights of its drives stay outside. Whoever may set their own goal sets curiosity to zero — and that is not a will but a fixed point.
Drive — two intrinsic terms, and a price
Curiosity on the disagreement of a predictor ensemble, empowerment on an information bound between action and future. No external reward. Since §10.317 a price on its own compute time has joined them (section seven). Homeostasis and an energy budget are built and stand still (section eleven).
Safety — the action list is the boundary
No shell access, no network, no foreign processes. What it can do is written out in full in the action list; every read path is checked against its root, every write against a second one. Everything else does not exist for it.
The first arc: learning, then direction
§10.107 to §10.182, still in a built world with a tracking task: a bar on one of four levels, a target that wanders. Two results.
Seven conditions before anything is learnt
For five sections every measurement looked like a no to the whole approach, and not one of them was a statement about the approach. Each of the following conditions was necessary on its own; while one was missing, behaviour stayed at chance level.
- 1 An effector that delivers. The action took effect one action late. §10.138
- 2 A world that is not the same at every restart. The target never saw two of its four levels. §10.139
- 3 An observation that reaches the decision head. The state carried the distance to target 52 to 61 per cent of the way, the raw observation 77 — the head saw only the state. §10.147
- 4 An action space without side effects. Eighteen of twenty actions moved the bar in passing. §10.150
- 5 A value function that reaches its fixed point. The value had propagated one step, not thirty-three. §10.152
- 6 Exploration, even after convergence. Otherwise the policy parks at the stop, where its favourite action does nothing. §10.154
- 7 Enough gradient steps — and that was the cheapest one. 125 updates per run instead of the 3,200 the same target needs under supervision. A number in a configuration file. §10.161
After that the agent holds the bar on the wandering target: 78.3 per cent in band against 27 for chance, replicated three times.
Direction from a condition on its own process
In a quiet world with no task reward at all, curiosity and empowerment produce structure but no common direction. Direction comes only with viability: stay a process that can go on acting — not a world state but a condition on its own operation, entered as a precision on the epistemic terms rather than as a fourth summand.
| Drive | State dependence | Seed agreement |
|---|---|---|
| random policy | −0.0004 | −0.228 |
| curiosity and empowerment only | 0.0309 | +0.256 |
| with viability added | 0.0806 | +0.811 |
| externally given task reward | 0.06–0.26 | +0.807 |
A drive that sets no world goal produces direction — and different runs find the same one.
With one limit that still holds: the term falls silent while it works. The agent gets into the band more often, and in exactly that process the term that drove it there loses its action signal by a factor of two to six.
The correction: the body was there
The August version of this record ended on the gap I took to be the largest: as soon as a world is configured, it replaces the telemetry sensor, and the agent receives none of the twenty-five channels of its host machine. That was not a property of the design. It was one line.
§10.192 to §10.196 · Three things were built, wired and dead
The telemetry was absent. The world replaced the host telemetry instead of joining it. Temperature, power draw and both throttle channels read exactly zero in every world run — not low, absent.
The price of thinking was ten times too high. The duration of a learning step sat in its perception as a per-step value: 3,189.8 milliseconds against a real step time of 405.
A head acted and never learnt. Since §10.161 the body head was trained only under a condition that no longer occurred — thirty thousand steps on its random initialisation, in every run. The repair alone cut its compute cost from 524 to 98 milliseconds.
Together the three explained why four calibrations of the viability drive had failed: the agent was using 4.5 per cent of its capacity. There was nothing for the drive to bite on.
The drive was never too weak. The body was too small.
With levers on its own learning effort it worked at once. And in the control without the viability drive, learning effort rose from 111 to 307 milliseconds per step: with no reward term asking for it, the free agent bought more thinking, because more learning raises the epistemic reward it already has. The first place in this project where it wants something nobody told it. How large the gap between the families is exactly is not secured — the first figure for it came from a block in which four arms were taking the machine away from one another.
The machine as world, and a hand
The built world demanded nothing: five containers together used less than one of twenty cores, and the tracking task made up about one per cent of what the agent did. So it moved.
| What | built world | machine |
|---|---|---|
| one world step | 132.7 ms | 0.48 ms |
| a run of 20,000 steps | ≈ 6 h | 48 min |
| structure to explore | 6 objects | 2,539 directories |
After that it looked four times as if the agent were sluggish — and four times the limit lay in what I had built. The clearest: it visited 32 places and then found no new one. Exactly 33 were reachable, because its action knew only the first eight names in a directory. It was done, not lazy.
It could see everything and change nothing
After 80,000 steps the machine was byte for byte the same. One place differed from another only in what lay there, never in what had been done there. Measured before the hand was built: the same place, twelve steps away and back, fifty times — a change of 0.00000 in the real world, 4.92 on a surface where it may write.
The boundary as anatomy, not as a guard
No write action carries a target; each acts where its head already is. An action cannot even express a place outside the hand — no nerve leads there. And the check is positive, not negative: a negative check that fails lets everything through; a positive one lets nothing through. Tested at 143 states with the head in the world it reads: byte for byte unchanged.
Perceiving and acting only work together
| Measure | read head alone | hand alone | both |
|---|---|---|---|
| empowerment | −0.137 | −0.155 | +0.204 |
| probe error | −0.030 | +0.001 | −0.075 |
| return | −0.114 | −1.118 | +1.153 |
Each switch on its own makes things worse, only both together make them better — in three largely independent measures. That is the definition of empowerment taken literally: the read head makes content visible, the hand makes it changeable, and only together is there something it can change and see. A factor of 6 above the spread between two seeds.
It is the only finding of that section that passes the new standard. For the same section showed that the old one was too small: the same configuration with a different seed, and two reported effects were smaller than the bare difference between seeds. Since then every claim between arms needs two seeds a side — or it stands as a hint, not a finding.
Sixteen times possibility
After that I gave it whatever each measurement seemed to say was missing. Sixteen interventions in its world. The most important:
| Intervention | arrives, gets used | moves the drives |
|---|---|---|
| a lid that bites, and an action that takes things back | destroying becomes more than twice as readable, +7.0 → +15.5 SD | no — ten measures, none outside the seed spread |
| a grip that shortens the way into the hand | presence in the hand 0.03 → 0.40, both seeds | no — empowerment −0.73, curiosity −0.14 spans |
| a source that answers every write | 1 : 1, in all four directories, as an 89-fold amplifier | no — −0.07 spans |
| grip and paste | 17 KB of real source code from more than fifty files into its hand | no |
| a teacher who demonstrates and moves on | seen; three of four steps stood on its table | no — the fourth was beyond what it could build |
| a key: sixteen bytes become a megabyte | files up to 1,048,656 bytes, chains up to 6 | no — it does not use it more often when it works |
| an imprint: every change regenerates the whole file | 1,957 imprints, frequent and broad at once | no |
| its own clock as a joint | used, the whole range run through | no |
Every piece works mechanically, cleanly and as designed. Every one arrives in its perception, and most change its behaviour clearly. None moves the two quantities it acts on.
It arrives, it gets used, it changes behaviour — and it decides nothing, as long as what you change is the world.
The reason is mechanical once you see it. For empowerment, an event it did not cause is not a gain in influence but noise. For curiosity, a random fragment is an error no learning can reduce. And both measure whether it can predict and influence — indifferent to what. Every enrichment of the world stayed invisible to what it wants, not because it did not arrive, but because nothing values it.
Necessity, once
Thirteen arms had changed the world when one place stood out that none of them had touched: its compute time. It pays it on every step, unavoidably, and it has a graded lever on it. So a price on it — the budget chosen from its own measured distribution, reachable, and so with a gradient.
| Run | Compute time | cheapest learning effort | costliest learning effort |
|---|---|---|---|
| no price, seed 1 | 121.9 ms | 0.301 | 0.343 |
| no price, seed 2 | 99.6 ms | 0.310 | 0.351 |
| with price, seed 1 | 88.9 ms | 0.406 | 0.296 |
| with price, seed 2 | 85.4 ms | 0.336 | 0.266 |
Compute time below both controls, the costliest learning effort chosen less often, the cheapest more often — in both seeds. On the scale, −4.36 spans: the largest and cleanest gap this project has ever produced on a quantity it controls. It regulates: a spread of 2.15 milliseconds against 41.10 without the price. And it reproduces — three runs with the price against three without, in two different worlds, both channels fully separated, p = 0.05 one-sided.
| With price against without | Gap |
|---|---|
| compute time — the priced quantity | −4.36 spans |
| prediction error of value learning — the cost of it | +5.27 spans |
| curiosity | +0.46 spans |
| empowerment | −0.09 spans |
Everything we gave it was possibility. The only thing that ever moved anything was necessity — a quantity that is expensive.
And what does not follow: that the price lifts the drives. It does not. The preference stays entirely local — it prefers what is priced, pays about 15 per cent for it in prediction error, and nothing beyond that moves. What has been tested is a term, not a goal.
The conserved quantity
If the world moves nothing and a price only locally, how much does its action reveal about its future at all? After eighteen interventions in the world, the suspicion fell on the measure. So three interventions in the estimator, each with a control:
| Estimator | Contribution of the action | Contribution of the state |
|---|---|---|
| control | 0.600 nats | 2.144 nats |
| state-near negatives | 0.498 nats | 2.362 nats |
| shorter training horizon | 0.639 nats | 5.691 nats |
| target without the host | 0.644 nats | 3.387 nats |
What the action reveals sits at 0.6 nats and does not move; what the state contributes varies 2.7-fold. My pre-registered primary measure — the action's share of the span — had been measuring the denominator for three arms.
Measured without a model, the action carries 1.2 of 27 points of the variance of its latent, and the estimator gets all of it; if anything it weights it too highly. And it breaks down: its action predicts that its state will jump four to eight times better than where to. Empowerment measures distinguishability. A deed that reliably changes something, but not predictably what, is worth nothing to it — even though it acts in the world.
The teacher outside the hand shows the same from the other side. It sees the board — 70 grips —, it carries it home — all 70 as effective pastes —, and it does not connect the two. Under a law that is bit for bit one of its own actions.
An invariant is the shadow of a symmetry. The next intervention has to break it — or it belongs to the same group as the twenty-one before it.
Nothing happens by itself
The most recent section stopped turning the dials on the measure and started checking the build. What came out is not another statement about the measure, but one about the world it lives in. It has two parts, and neither can be learnt.
| Part of its world | changes | caused by |
|---|---|---|
| noise — fans, network, the clock of the graphics card | on 100 % of steps | not by it |
| its own position | on 0.16 % of steps | by it |
Not one of 380 channels is more than 11.5 per cent predictable. The most predictable is its own clock, then the time of day — and that only because sine and cosine compute each other. The online model extracts 7 to 8 per cent; the best offline model with a clean time split and a null floor extracts the same or less. Nothing is left on the table. The limit is not the predictor's capacity, but that nothing happens lawfully.
Its body does not answer it either
The host's telemetry, in its perception since §10.192, accounts for more than half of the step-by-step movement of its latent — and not one of its channels answers to its action. In the rebuild of §10.370, 19 of the 25 host channels were therefore taken out of its encoder's input; what remains is its step duration, its process load, its memory, its threads and the time of day, and the safeguards still read all of them. The ceiling did not rise. It was not the host channels.
And once it looked like learning
When it was allowed to set its own clock, the clock's predictability fell from 11.5 per cent to −0.1 — because it now pulls the lever, almost evenly across the whole range, before it knows where to. A passive drift is predictable, a random walk is not, and learned control would be again. For the first time in this arc not a flat invariance but the start of a curve. At 20,000 steps only its start is visible.
How the work was done
The methodological yield of this arc is more solid than the substantive one, and it was paid for more dearly. Three classes of error, each more than once.
| Class | What it looks like | Where |
|---|---|---|
| built, wired, dead | a capability that never acts in the run, while everything in the log looks right | the absent telemetry; a head that never learnt · §10.192–10.196 |
| measured, built, never switched on | a switch whose default is “off” and which no configuration sets | a wider grip, off in all 266 configurations that ever ran; a carefully built prediction path in none · §10.325, §10.367 — and the price itself, unnoticed and off for six arms |
| the wrong denominator | a number moves because what it is divided by moves | the “action's share of the span”, for three arms · §10.359; an arm that wrote more rather than composed more · §10.350 |
And the sum of it, in the latest audit: eight of eight apparent findings were errors in my measurement, not in the code. The gap between what the code can do and what a run actually does is invisible from both sides — a configuration does not show inherited defaults, the code does not show which switches are ever set. Since then there is a tool that asks a run what it really does. Applied to the latest one: 80 values written, 96 inherited, 43 of them silently.
The building blocks are precise. The hard part is setting them right — and first of all when measuring.
Habits that catch it
What stays honest
So that the claim does not grow larger than the measurement.
Where it can go
Sixteen interventions in the world, one in the goal, and a world in which nothing happens lawfully. Three directions follow, at different distances.
And across all of it, since the map in §10.246: language and intention both stand on acting. Acting now stands. What both still lack is a world in which something lasts and something counts.
What remains
The first arc showed that an approach can look like a failure six times running while it hangs on one missing condition, and that a condition on its own process is enough to turn structure into direction. The second showed where that direction ends.
Curiosity and empowerment produce a being that wants to be able, without wanting anything in particular. It used every possibility we gave it. It formed a preference only once something was expensive.
Whether such a scarcity is still free, or merely a reward under another name, is not a question measurement can settle. What measurement can say: it sets no world state, it leaves open how it is satisfied — and it is so far the only thing this build has answered with a preference of its own.