Origin of Inquiry

Chapter 7 · Dewey 1896, von Uexküll 1934, Simons-Chabris 1999, Gibson 1979, Miller-Galanter-Pribram 1960, Powers 1973, Boyd 1976

The unpaid debt

The diff carries an unpaid debt. It partitions state into figure and ground, but only over the variables you chose to snapshot. The diff requires you to know what to observe. Bi-abduction pays half of it: it infers the frame without a global snapshot, reading the relevant variables off a function's own reads and writes. That works where the system has a legible footprint. Code does. The mechanic does not.

Return to the mechanic. There is no source code to read her variables off. She watched the dash lights, listened to the starter, metered the resting voltage. Six readings, out of the thousands the car can yield. Nothing in the diff told her which six; the diff runs on the basis, it never chose it.

They came from what she was trying to do.

The honest answer is one Peirce gestured at but never built. He named the question and pointed to instinct, to guessing, to musement, the free play of mind from which hypotheses are supposed to emerge. He did not say what fixes the variables. The core inversion that does was written down in 1896, in a paper that never mentions abduction.

Dewey: the unit of behavior

John Dewey's "The Reflex Arc Concept in Psychology" (1896), the paper he would later reprint under the name of its thesis as "The Unit of Behavior," attacks a picture everyone took for granted: that a stimulus arrives, and a response follows. Dewey inverts it. The response is not downstream of the stimulus. It is what selects the stimulus in the first place.

"We begin not with a sensory stimulus, but with a sensori-motor coördination... it is the movement which is primary, and the sensation which is secondary, the movement of body, head and eye muscles determining the quality of what is experienced."

— Dewey, "The Reflex Arc Concept in Psychology" (1896)

His example is a child and a candle. The standard story: the child sees the flame (stimulus), reaches (response), gets burned. Dewey's correction: the seeing is already an act. The eye moves, fixes, holds the flame because the child is doing something, and the holding is what makes the flame a stimulus at all. Change the act and the same flame stops being the stimulus. A child trying to read by candlelight registers brightness; a child trying to grab registers reachable heat. One candle, two stimuli, and the difference is the act.

Stated flatly:

"It is the motor response which assists in discovering and constituting the stimulus... stimulus and response are not distinctions of existence, but teleological distinctions, that is, distinctions of function."

— Dewey (1896)

That is the missing chapter in one sentence. The stimulus is not given and then attended to. It is constituted by the act, and which features count as the stimulus is a teleological distinction, a distinction of function, of what the act is for. The mechanic's six gauges are not the salient features of the car. They are the features salient to fixing it. Her purpose constituted her basis before any diff could run on it. The variable basis is the wake of the act, as the figure was the wake of the change.

Dewey thought enough of this argument to reprint it thirty-five years later, and renamed it for its thesis instead of its target. "The Reflex Arc Concept" became "The Unit of Behavior." The unit is the coordination: the act and its perception held as one thing, with the stimulus a role inside the act rather than a trigger outside it. The unit of behavior is the act, and the act fixes what is seen.

The gorilla

Dewey argued it from an armchair. Psychology put it on film. In "Gorillas in Our Midst" (Simons and Chabris, 1999), viewers watch a video and count how many times the players in white pass a basketball. Midway through, a person in a gorilla suit walks into the scene, faces the camera, thumps their chest, and strolls off. They are on screen for nine seconds. About half the viewers never see them.

The gorilla is the candle, controlled. The counting task is the goal. The goal frames the basis to white shirts and the ball, and a black gorilla is, by that frame, irrelevant. The light from the gorilla lands on every viewer's retina the entire time; the signal is never missing. What is missing is its promotion into task-relevant awareness, the step from present-on-the-retina to noticed. That promotion is what the goal grants or withholds. Without a task to fix relevance there is no figure, only input. The gorilla stays input. It is easy to miss something you are not looking for, which is the slogan a later road-safety ad hung on the same effect, and it is also the entire thesis here.

The demonstration is iconic. It dramatized a line that Neisser and Becklen (1975) had opened with overlapping transparent videos, and that Mack and Rock (1998) had named inattentional blindness. The contribution was the vividness: a chest-thumping gorilla is harder to wave away than a transparent ghost, and it made undeniable what the tradition already held. Perception is not a camera that records and then asks what matters. It is an act that fixes what matters and records that. Selection is not a filter applied after seeing. It is the condition of seeing at all.

Note the double edge, because the chapter ends on it. The same experiment that shows purpose frames perception shows purpose blinds it. The engagement that holds the counting task is the same engagement that hides the gorilla. A sharp frame is a sharp blindness. Hold that; the closing section spends it.

Nested acts, nested bases

An act is never solitary. The mechanic diagnoses because she means to fix; she fixes because someone means to drive; the driving serves some errand past that. Each act sits inside a larger one, and each frames a narrower basis than the act above it.

drive somewhere        basis: roads, fuel, time
  └ fix the car        basis: the fault, the parts, the repair
      └ diagnose       basis: crank, lights, click, voltage
          └ diff(before, after)   ← the primitive starts here

The test at each level sets the basis at the level below. "Is the car drivable?" is itself a contrast, and resolving it spawns "is it fixed?", which spawns "what is broken?", which finally licenses the six gauges. The diff primitive does not just run once. It runs at every rung, and the basis it runs on is handed down from the rung above. Only the bottom rung was formalized; the rest was taken as given. The "given" was the goal stack.

This nesting is not a metaphor borrowed for the occasion. It is the recurring structure under every serious account of goal-directed agents, from HTN planning to Soar.

Five names, one operation

The diff appeared under eight names. The frame has its own list. Five traditions, none citing the others, each arrives at the claim that an agent does not perceive neutrally. Its purpose selects what counts as relevant. They are analogues, not one machine. But the role is identical.

Two of these carry more than corroboration. Powers' Perceptual Control Theory states the inversion as a control law: the organism acts to hold a perception at a reference value, so the variable it controls is precisely the variable it frames. And the reference values are layered, high loops setting the set-points of low ones, which is the nesting above written as a control hierarchy. Miller, Galanter and Pribram had already built the same nesting in 1960 as the TOTE unit (Test-Operate-Test-Exit). The unit imported the cybernetic loop into psychology as a hierarchy of plans, where each Test fixes what its level observes. Two independent statements of the nested basis, thirteen years apart.

Reinforcement learning supplies the one place the claim gets a narrow formal analogue. In the state-abstraction taxonomy of Li, Walsh and Littman (2006), one criterion keeps exactly the distinctions that bear on optimal action and discards the rest: a distinction that never changes what you should do is provably droppable. That is "purpose frames the basis" made precise, in one setting. But read the quantifier carefully. The abstraction is defined relative to a fixed reward and a fixed underlying state space. It compresses a basis you already have. It does not invent one. That gap is the seam.

Boyd: orientation frames observation

One more witness, from the least academic corner. John Boyd's OODA loop is usually drawn as a cycle, Observe, Orient, Decide, Act. Boyd insisted the drawing lies. In "The Essence of Winning and Losing" (1995–1996) he makes Orientation dominate the loop and feed back onto Observation itself: orientation, shaped by "genetic heritage, cultural tradition, previous experiences," determines what the observer can even pick up. You see what your orientation lets you see. That is purpose framing the basis, stated in a fighter pilot's idiom, rarely connected to Dewey or to abduction.

Code: the basis is goal-relative

The same world, queried under two goals, yields two bases. Nothing about the car changed. What changed is what the act needs to know.

# A world is a bag of measurable quantities.
world = {
    "crank", "dash_lights", "starter_sound", "battery_rest", "battery_load",
    "fuel_level", "tire_pressure", "cabin_temp", "radio_preset", "wiper_angle",
    "coolant_temp", "oil_pressure", "odometer", "paint_hue",
}

# A goal supplies the predicate that decides relevance:
# a variable is in the basis iff knowing it could change the act.
GOALS = {
    "diagnose_no_start": lambda v: v in {
        "crank", "dash_lights", "starter_sound", "battery_rest", "battery_load",
    },
    "plan_a_road_trip": lambda v: v in {
        "fuel_level", "tire_pressure", "odometer", "coolant_temp",
    },
}

def frame(world, goal):
    """The variable basis is whatever the goal makes action-relevant."""
    relevant = GOALS[goal]
    return {v for v in world if relevant(v)}

for goal in GOALS:
    basis = frame(world, goal)
    print(f"{goal}:")
    for v in sorted(basis):
        print(f"    {v}")
    print()

# The diff runs AFTER this. It never chose the basis.
# The goal did. Change the goal, change what there is to diff over.

The relevance predicate is hand-written here, on purpose. The diff machinery begins one step to the right of this function: it consumes basis and never produces it. The producer is the goal.

No root, and none needed

This relocates the cold-start. Does it abolish it? "The goal supplies the basis" pushes the question up a rung. Where does the goal come from? From the goal above it. Diagnose because fix, fix because drive, drive because there is somewhere to be, and the why's keep going. The chain does not end in a final purpose the agent was issued at the factory. It is open. There is no root to reach.

That looks like a regress problem and is not, because inquiry never needed the root. To fix the basis for the task in front of you, the live purpose one rung up is enough. The mechanic need not justify driving, or living, to know that fixing the car means watching the starting circuit. You frame relative to the purpose you are actually in, not the ultimate one, and you stop climbing the why's the moment the basis is fixed. So the cold-start is not solved but dissolved. There is no frameless first moment to bootstrap from, because an agent is always already inside some live purpose, and a live purpose is enough. Musement was Peirce reaching for a starting frame in a void. There is no void; there is always a why one rung down from the one you are answering.

The traditions admit as much, in their honest moments. Powers added a process he called reorganization: when intrinsic error stays high, the control hierarchy rewires itself, more or less at random, until the error falls. Boyd, in "Destruction and Creation" (1976), argued from Gödel and the second law that any orientation decays and must be torn down and rebuilt. Both are accounts of where a new basis comes from when the old one fails. And both are undirected: they generate variation and let error select.

This is where the folk wisdom is sharper than the theory. Necessity is the mother of invention, and a mother does more than trigger a birth at random. The need shapes what gets born. Reorganization says that the agent re-frames under pressure; it does not say the pressure aims the re-framing. The claim worth defending, and the one the prior literature does not quite make, is that the basis invention is directed by the need: motivation does not only select among given frames, it gives birth to the frame the act requires. That is a generation problem, not a selection problem, and it is what separates basis invention from the economy of research that follows, which only ever selects.

No tradition gets basis invention for free. Predicate invention in inductive logic programming, Lenat's discovery systems, the structure-learning extensions of active inference: each invents new terms, and each presupposes a prior search space, a grammar, a body, or a purpose to invent within. The defensible statement is not that the cold-start is unsolved but that it is never frameless: invention always runs inside a frame it did not itself produce, and that frame is whatever purpose is live, never the last one, because there is no last one.

Observe, then perturb

Framing has two halves. The observe half: the goal fixes what to look at. The perturb half, the economy of research, is which experiment to spend on, given a basis and a budget. Peirce wrote the second and left the first to musement. You cannot rank experiments over a basis you have not framed, and framing without a budget to test is contemplation, not inquiry. Observe, then perturb. Dewey, then Peirce.

What breaks

Purpose frames perception. It also blinds it. The same act that constitutes the right stimulus makes the agent unable to register the wrong one, and "wrong" is judged only against the current goal. The mechanic framed for a no-start will not see the slow coolant leak. Her basis is correct for her act and blind to everything outside it.

Three ways the frame fails, each naming work still owed:

• The blind spot is structural. A basis is a commitment to ignore. If the real cause lives in a variable the goal made irrelevant, no amount of diffing recovers it, because the diff never sees it. The gorilla showed it: the tighter the frame, the surer the blindness. It is the same failure one level up: before, a variable nobody measured; here, a variable the purpose never framed.
• Salience overrides purpose. Pain, startle, a loud bang: the world sometimes seizes the basis from the bottom up, against the goal. Purpose frames much, not all. An account that makes perception purely top-down cannot explain the interrupt.
• Goals conflict. The agent rarely holds one goal. Diagnose-fast, diagnose-cheaply, and avoid-liability frame three different bases over the same car, and when several drives contest one perception, which one frames? That is the fork, and it is where the fan of branches gets interesting.

This is the bind the method inherits. Purpose is what makes observation finite: without it there is no figure, only the indefinite mass of everything measurable. Purpose is also what makes observation partial: every frame that picks something out is a frame that leaves something out, and the thing left out is chosen by the same goal that picked. The frame is, in Boyd's reading, an attack surface: name what you are looking for and you have named what you are blind to. A method that frames by purpose inherits the vulnerability of purpose. What follows spends the budget the frame defines, then asks what to do when the frame itself is the thing that has to change.

Sources