The Boundary of Computation

Most of the work in a political maneuver is clerical. Someone has to remember what the actor wanted, who was thought to control it, which levers were declared and what each one needed, which of them stalled for want of a resource, what was spent and what came back. An executive running maneuver after maneuver carries all of it in their head, and carries it badly, and a machine could carry it well. Interleaved with that work at every step is a much smaller set of decisions that nobody can carry for the actor: whether the chief operating officer was persuaded, whether to give up a lever, whether the thing the maneuver was opened for is still worth having.

The Lam System for Political Interaction is built around the line between the two. Everything a maneuver writes into its record, and everything that follows from what has been written, the system specifies exactly enough for a machine to administer. Every step that moves the maneuver forward waits on a judgment, and the system leaves each of those to the actor. This paper argues that the line sits in the right place, that moving it in either direction breaks something, and that a system drawn this way can still be rigorous, in the one sense of rigor that political work admits.

What the machine can hold

Take the maneuver the hub introduces, in which Lena Ortiz needs the chief operating officer, Daniel Okafor, to approve moving the night dispatch desks onto her budget. She has two levers. One is the overtime case, which shows Daniel that a single desk cuts overtime he has taken criticism for; the other is the savings figure, carried to him with the signature of the controller, Priya Natarajan. She holds the overtime data and her standing with Daniel, and she does not hold Priya's sign-off.

A machine keeping Lena's record can do a great deal with that. It can file each entry at its address, so that the second lever's required resource has a place anyone can point to. It can refuse an entry that does not fit: an execution with no lever above it, a selection made by Lena when only Daniel selects, a second adverse alternative filed under the same name as the first. When the second lever's gate fails, it can open the sub-cycle aimed at exactly the missing sign-off, hold the parent where it stopped, and hand the sign-off back once it flows in. At any moment it can say which cycle has control, what Lena holds, counted, and which exits remain open to her. Asked what the maneuver looked like before the sub-cycle opened, it can replay the record to that point.

Each of these is a consequence. Given the entries already written, the next state of the record follows, and a machine that knows the system's grammar can work it out without being told anything about Daniel. That is what administering a maneuver means: keeping the record, checking its form, and telling the actor what follows from what. A later essay in the series, How an AI Administers a Maneuver, takes up how a machine does that well. This paper is about where it has to stop.

The questions that move a maneuver

A maneuver advances when certain questions are answered, and three of them cannot be answered from anything in the record.

The first comes after a lever is played. Lena puts the overtime case to Daniel and then has to say whether his reading of it matched what she projected: whether approving the move now looks better to him than leaving dispatch with the depot managers. If it did, a perception is written and the cycle moves toward his selection. If it did not, a failed perception is written, and Lena has to decide what to do next. Which of the two entries the record gains depends entirely on her answer. The record holds what she projected and the fact that the lever was played. What Daniel thinks is not in it, and nothing could put it there.

The second is an election that cannot be taken back. After a failed perception Lena can declare a new lever, moot the cycle, or abandon it. Earlier, with two levers declared, she can resource the second now or press on with the first. When one lever has become unnecessary she can moot it while the other continues. She can revise what the cycle is after, which sends detection and the objective around again, and she can walk away at any point. Each of these closes something off. A mooted lever does not come back; an abandoned cycle stays abandoned, and trying again means opening a new one. The system permits each move where it is available and prefers none of them.

The third is an attestation: the actor's statement that something the cycle rested on has changed. Suppose that while Lena is getting the sign-off, Priya tells Daniel herself that the overtime saving is real. Lena's first lever was built to make that point to Daniel, and it may now have been made. Whether it has, and so whether to moot the lever as already served, is Lena's to say. The record shows what Priya was asked for and what she signed. It cannot show what Priya said in a corridor, or whether Daniel heard it as the point Lena's lever was meant to make.

Even the gate, which looks like the most mechanical moment in the cycle, waits on a judgment before it runs. Its test is a count: every resource the lever was declared to need must be among what the actor holds. Once both lists are written, the comparison is exact; deciding what goes on them is a judgment. Lena's standing with Daniel is a resource, and whether she holds enough of it for this particular ask is a reading of Daniel, made from what she knows of him. The machine compares the lists. It cannot say whether a relationship is in good enough repair to spend.

What these questions share is that each is answered by looking at the world, or at the actor's own aims, and none by looking at the record. The system's own position on utility is that it is appraised qualitatively, and two careful people appraising the same target will not always agree, however much structure surrounds the appraisal. A machine asked to supply the answer would be guessing about Daniel, and the guess would go into the record as a fact, with every later entry following correctly from something nobody observed.

In the illustrative sorting below, five of the eleven steps taken from Lena's maneuver are her judgment and the other six follow from the record. The count belongs to this example. In any maneuver, the steps that land on the actor's side are the ones where it could have gone more than one way.

Two surfaces of judgment

The questions are of two kinds, which differ in what they answer and where they arise.

Interpretive judgment reads the world. Did Daniel's reading validate the projection? Does Lena hold what the lever needs? Has the lever's purpose already been met? Each question has an answer that is true or false about the situation, even though only the actor is placed to give it. Once the actor gives it, the system has nothing left to choose, because the answer selects the next entry. A perception that validated is written as a perception; one that did not is written as a failure.

Strategic judgment shapes the sequence: which lever to declare first, whether to lay every lever out at the start or add each one after seeing how the last landed, whether to resource a pending lever now or carry on, whether to moot a lever that another lever's outcome has made unnecessary, when to revise and when to walk away. These questions have no answer in the situation. Two competent actors who read Daniel identically can sequence differently for good reasons, because they take different stances toward the uncertainty of whether their levers will be enough. One commits to the next lever only after seeing what the last one did. Another plans every lever at the start and secures each one's resources before moving to the next. A third lays out the whole plan, finds everything missing, and goes after all of it. The system receives whichever stance the actor takes and ranks none of them.

The two kinds arise at different points in the cycle. Interpretive judgment sits at the gate, where the actor's account of what they hold meets what the lever needs, and above all at the reading after an execution, where the actor says whether the target moved as projected. Strategic judgment sits wherever the next entry could be one of several: just after a lever is declared, when the actor can declare another or start resourcing this one; at every point before a pending lever reaches its gate, since resourcing it is a standing option until then; and just after a failed perception, when a new lever, mooting and abandoning are all open.

Even so, they are one boundary. On both sides the system reaches a point where its own rules admit more than one continuation and nothing in the record picks among them. On the interpretive side the actor's reading of the world picks; on the strategic side the actor's choice does. The system is permissive at both points for the same reason: they are where the actor's intervention enters it. Interpretation fills in the tests the cycle runs, and sequencing sets the order in which the tests come up. A machine that settled either kind would be supplying the actor's part of the maneuver, whichever kind it was.

Treating them as one also matters in practice. An actor reviewing a maneuver that went wrong wants to know which of their judgments was at fault, and the record can tell them only if every judgment, of either kind, was written as theirs. If a machine had settled the sequencing and left the readings to the actor, or the reverse, the record would be a mixture, and the actor could no longer tell a bad reading of Daniel from a sequence they never chose.

When an open fork is a defect

A system that leaves judgment to the actor has an excuse ready whenever its rules fall short. Any point where the rules fail to say what happens next can be called the actor's discretion, and any behavior the system cannot write can be called a choice the actor declined to make. Left unchecked, "the actor decides" absorbs every gap in the design, and the boundary of computation becomes a place to hide mistakes.

LSPI holds itself to a strict test, which can be stated as a principle. A fork belongs to the actor only if every legitimate way of proceeding at that point can be written through it. Different actors may take different branches; the fork is the actor's because each branch can be reached by some choice the actor makes. If a legitimate way of proceeding cannot be reached whatever the actor chooses, the fork is a defect in the system, and the system has to change.

An invented illustration shows the shape. Imagine a version of the system that let an actor resource a lever only once every lever in the cycle had been declared. The actor who plans everything at the start could be written without trouble. The actor who declares one lever, plays it, reads the result and only then declares the next could not be written at all, because they resource each lever before the next one exists. That version might describe its rule as leaving the order to the actor. The test disagrees. An adaptive actor is a legitimate actor, the version gives them no branch to take, and so the fault lies with the version. The system is meant to receive the actor's stance, and a rule that makes one stance impossible has chosen one.

The test is asymmetric on purpose. Calling a fork the actor's is an active finding: it requires checking each legitimate way of proceeding against what the system admits, and the claim holds only when every one of them gets through. Calling it a defect needs only one legitimate behavior that the system cannot write. The burden falls on the claim of discretion, because that is the claim a designer is tempted to make.

The test has a third outcome. Some points that look like forks offer no strategic choice. After a lever is executed, the record will gain a perception or a failed perception, and the actor's reading of the world decides which. Such a point is neither discretion nor a defect. It is interpretive judgment, and the test keeps it apart from sequencing so that neither kind is mistaken for the other.

There is a second way to get the boundary wrong. A system can be defective by forbidding what an actor legitimately does, and it can overreach by deciding what only the actor should decide. A rule that mooted a lever automatically whenever another lever's outcome seemed to cover its purpose would close a genuine fork by mechanism. It is the same kind of fault as the adaptive actor's, met from the other direction: there a move the actor makes could not be written at all, and here a move is written by the machine, on an attestation Lena never gave.

Determined, once the judgments are fixed

With every step that moves a maneuver left to the actor, the system still guarantees a great deal, and the guarantee can be stated precisely. Once the actor's judgments are fixed, each reading, each election and each attestation, the rest of the maneuver is determined: which entries the record gains, what the actor holds, which cycle has control, and how the maneuver ends.

The machine keeping the record does have a freedom of its own. When two things in a cycle are independent of each other, it may write them in either order. If both of Lena's levers have passed their gates and each spends a different resource, the two executions can go into the record either way around, and she ends up holding the same resources, with the same entries and the same exits open. Choices of that kind are scheduling, and the system's guarantee is that scheduling never changes where a maneuver ends. Only the actor's choices at the forks that cannot be undone can do that, so the system is determinate up to the order of independent moves. The essay Order Only Matters at the Forks works through which orders are which.

The record follows the same line. It is exact for whatever was sequenced: everything within one lever, everything at the boundaries of a cycle, every escalation and handback, every election the actor made. For moves that ran independently it is exact up to their order, and it records that they were independent, because insisting on one true order for two levers pursued side by side would invent a sequence that never existed. The paper on the record takes this further.

The phrase "once the judgments are fixed" can sound like a hedge, a promise of determinism followed by an excuse. It is the only form of the claim that can be true. A system for political work that promised determinism outright would be claiming to know how Daniel will read the overtime case, and nobody knows that. A system that promised nothing would be a notebook. Between the two, this guarantee says exactly which part of a maneuver's course came from the actor and which part came from the system, and it says the system's part holds without exception.

The precision pays in two ways. If two people record the same maneuver from the same judgments and produce different records, one of them has made an error of form, and the system can find it. And when a maneuver ends badly, every difference between what happened and what the actor wanted traces back to a judgment, because nothing else was free to vary. Lena can go to the record and find the reading of Daniel that was wrong, or the lever she mooted too early, without wondering whether the bookkeeping played a part.

Why it is not a calculus

A calculus, in the sense mathematicians and computer scientists use, computes. Its steps are defined mechanically, so a program can take any expression and run it to the end with nobody supplying anything along the way. That property is what makes a calculus worth having, and LSPI cannot have it. The steps that decide a maneuver wait on the questions above (whether the reading validated, which irreversible election to make, whether what the cycle rested on still holds), and none of them can be decided from the record. Calling LSPI a calculus would contradict the claim this paper has been making, which is why the series gives a separate essay, Not a Calculus, Not an Algebra, to the system's name.

So the system describes the process it records, and it cannot be executed. The actor advances the maneuver according to what they observe, and whoever keeps the record writes each advance down and checks it against the system's rules. Its rigor comes from guarantees that hold whatever the actor decides: every constituent has one address, nothing is overwritten, capital is conserved, every maneuver ends, and order matters only at the forks. Because they hold whatever the actor decides, they can be shown to hold, and the paper on what the system guarantees gives the reason for each.

What the line protects

Moving the boundary toward the machine breaks the record's main use. The record exists so the actor can learn: what they projected stays beside what happened, cycle after cycle, and their habits as a reader of targets become visible over many maneuvers. That works only if the projections and readings in the record are the actor's own. A machine that filled in Daniel's perception would leave Lena with its guesses beside Daniel's actual choices, and the comparison would teach her about the machine.

Moving it toward the actor fails differently. An actor who has to work out by hand which cycle has control, what they hold after three levels of escalation, or whether a sub-cycle has delivered what its parent needed will get some of those wrong under pressure, and the errors will look like failures of judgment when they were failures of arithmetic. The system takes that work off the actor's hands. What is left in them is the reading of Daniel, the choice of which lever to try next and the decision to stop, and the record keeps each of those as the actor made it, beside everything that followed.