Figure and Ground

The most important thing to understand about figure and ground is that generating the members of a set does not automatically give a terminating test for everything left outside it. Day 02’s interpretation receipt linked pq-theorems to true additions without confusing strings and meanings. Today you will treat the theorem set as a figure, inspect its complement as ground, and prove why two opposing generators can decide membership even when either one alone cannot.

The cumulative model keeps the formal system, its evidence, and Day 02’s interpretation test unchanged, then adds a set-theoretic reading. A universe splits into a figure and its complement, meaning everything in the universe that is not in the figure. A generator may eventually name every figure member. To decide an arbitrary input, however, a procedure must halt with either “inside” or “outside.” The new decision path makes the gap between eventual positive evidence and a guaranteed two-sided answer visible.

Boundary. The lesson defines recursively enumerable and recursive sets through generation and decision. It uses Hofstadter’s visual and TNT set picture as conceptual guides; it does not prove the later incompleteness claims suggested by Figure 18.

A viewer normally treats one region as the object and everything around it as leftover, with the consequence that the same boundary carries unequal attention in its two directions. Figure is the region attended as the object; ground is its negative space. A figure-ground reversal occurs when attention switches and the former background becomes an organized figure in its own right.

Hofstadter moves through three source images. Figure 15 hides letters in apparently irregular blobs; Figure 16 uses Escher’s birds so positive and negative regions both form recognizable shapes; Figure 17, Scott Kim’s FIGURE-FIGURE Figure, makes both colors spell the same word. The pictures progress from accidental ground to independently legible ground.

Source trace. The three-source sequence comes from Chapter III, printed pages 67–69: Figures 15–17 and the accompanying figure/ground discussion. The lab is a conceptual reconstruction, not a reproduction of the artworks: it preserves the attentional reversal and the distinction between accidental and independently structured ground, while using simplified generated shapes and adjustable contrast.

Protocol. Select each Source figure (15–17) in order, set Perceptual focus first to figure, then ground, then both, and record which region is active, contextual, or recessed in the visible model. Lower and raise Contrast while focus stays on both, watching the contrast metric and region emphasis change while the selected source figure stays fixed. For each source figure, use the interpretation readout to state whether the ground is merely leftover or can be held as a positive reading of its own.

Limits. The lab uses course-built abstractions and does not reproduce Hofstadter’s, Escher’s, or Kim’s artwork pixel for pixel. A focus toggle directs attention rather than measuring perception, and contrast can make a form hard to see without changing its mathematical boundary. Recognizability is observer-dependent; the later set-theoretic distinction is stricter and must not be inferred from visual confidence alone.

The mechanism matters in interface engineering because an allowlisted path also creates a denied space, but a user interface may make only the allowed path legible. An LLM agent policy similarly has visible permitted tools and a less visible complement of forbidden actions. A startup dashboard foregrounds converted users while the unconverted ground may hide distinct segments; in daily life, a calendar shows scheduled time as figure and leaves unscheduled time as ground, which becomes useful only when examined on its own terms.

A visual reversal remains subjective, so mathematics needs a case where figure and ground have exact membership. A number greater than one is composite when it is a product of two smaller positive integers greater than one; a prime has no such factorization. If a formal system generates a marker C followed by exactly a composite number of hyphens, its theorems form the composite figure and the missing C-strings form prime-shaped holes.

Chapter III first modifies Day 02’s system into the tq-system, whose group lengths encode multiplication. Its axiom schema xt-qx represents X × 1 = X; its rule carries xtyqz to xty-qzx, adding one to the second factor and one whole copy of the first factor to the product. The C-system then adds this rule: if x-ty-qz is a tq-theorem, produce Cz. The extra hyphen on each factor ensures both factors exceed one, so every C-theorem’s hyphen count is composite.

Every displayed C----, C------, C--------, and so on has a finite witness: two factors and a corresponding derivation. When C----- has not appeared, waiting longer does not itself prove that five is prime. The positive generator is asymmetric: appearance carries a witness; silence carries an unknown delay.

A simple composite list can make every complement look equally manageable, and that expectation fails once richer formal systems distinguish more layers. TNT, or Typographical Number Theory, is Hofstadter’s later formal system for arithmetic; Chapter III previews it by nesting all symbol strings, well-formed formulas, sentences, theorems, and interpreted truths rather than treating them as one set.

Figure 18 draws the theorem set as a branching tree inside the region of truths and its negations as a reflected tree inside falsehoods. The picture anticipates that neither tree fills its containing region. That is a source claim pointing forward to Chapter XIV, not something the picture proves and not a result established by this day’s finite lab.

Source trace. This set hierarchy reconstructs Figure 18 and its explanation on printed page 71. It preserves the nesting and the two theorem trees but replaces Hofstadter’s irregular coastline and hand-drawn composition with selectable set layers and synthetic sample strings.

Protocol. Cycle Set in focus through all strings, well-formed formulas, sentences, true sentences, and theorems, and record which layer becomes active and which surrounding layers become context or narrower distinctions. Change Example family while holding the set fixed and watch the displayed sample strings change without moving the focused boundary. Increase Samples shown, confirm the samples-per-layer metric changes, then compare the theorem and truth layers without claiming that the finite display measures their infinite sizes.

Limits. The sample families are synthetic teaching cases, not an enumeration of real TNT, and their counts are not proportions of the infinite sets. The visual places theorems among truths and negated theorems among falsehoods as the source figure does, but the lab does not prove consistency, soundness, or incompleteness. Smooth boundaries replace the source’s deliberately intricate coastline and therefore understate how hard the complement can be to characterize.

In software, these layers correspond to text, parseable programs, type-correct programs, accepted builds, and behavior that is actually correct; passing one boundary does not collapse the next. An agent response can be valid JSON, schema-valid, policy-approved, and still factually wrong. A business metric can be syntactically present, correctly calculated, and still fail to measure customer value; a daily message can be grammatical yet false. The lab’s mechanism is a boundary audit, not a metaphorical resemblance.

A generated set may never announce that a missing item will remain missing, so its membership procedure is only one-sided. A set is recursively enumerable, abbreviated r.e., when a mechanical enumerator can list every member eventually; the order may be irregular, duplicates may occur, and there need be no known bound on when a particular member appears.

If the target appears, halt and answer yes. If it has not appeared, the procedure must keep open two possibilities: it is outside the set, or it is inside but scheduled much later. This is exactly the search-frontier lesson from Day 01 generalized from MIU strings to arbitrary generated sets.

A set is recursive, in the historical terminology used by the chapter, when it has a decision procedure: a mechanical test that halts on every input and answers membership correctly. Modern texts often call such a set decidable. Every recursive set is r.e., because one can enumerate all possible inputs and output those the decider accepts. The converse fails in general: some r.e. sets are not recursive.

One enumerator cannot distinguish “not a member” from “not listed yet,” and the consequence is an open-ended negative query. If both a set and its complement are r.e., run their enumerators in dovetailing fashion—alternate one step of each—until the target appears on one side. Exactly one side contains it, so one side must eventually print it and the combined procedure must halt.

The bounded journey is mechanical:

1. Fix a target string t in an effectively enumerable universe.
2. Start enumerator F for the figure set and enumerator G for its complement.
3. Run one step of F, then one step of G, repeating fairly.
4. If F prints t, answer “member.” If G prints t, answer “nonmember.”

Running one enumerator to completion before starting the other would fail because an r.e. enumerator may never complete. Fair alternation is the mechanism that prevents either side from starving.

Source trace. Chapter III defines r.e. and recursive sets and derives the link to decision procedures on printed pages 72–75. The paired-enumerator lab is an explicit conceptual reconstruction of that relationship: fair scheduling, budgets, target selection, and live verdicts make the implicit decision mechanism observable.

Protocol. Choose a Target string, set a small Enumeration budget, and leave Complement search off; record whether nonappearance yields a no verdict or only an unresolved state. Raise Enumeration budget until a positive target appears. Then turn Complement search on, test one target from each side, and watch the trace rows, rounds-used metric, and decision readout; confirm that the run halts when either enumerator prints the target, not when the budget merely expires.

Limits. Finite budgets can demonstrate individual halts but cannot prove that arbitrary enumerators cover their promised sets. The complement search toggle assumes the second stream is a correct enumerator of the exact complement; establishing that premise is separate work. Fair alternation is simplified to visible rounds, and performance in rounds says nothing about practical efficiency when one witness appears extremely late.

This construction has direct operational uses. A software policy can decide every request when one recognizer enumerates valid capabilities and another independently enumerates explicit denials; two “allow” searches do not suffice. An LLM agent needs positive evidence for permitted tool use and a mechanically checkable rejection path to avoid hanging on ambiguous actions. A startup can route every lead only when both qualification and disqualification criteria cover the universe; in daily life, reconciling a bank statement terminates when every transaction is matched either to an approved category or to an exceptions list.

The vocabulary is close enough to ordinary language that several false equivalences feel natural, and each one leads to an invalid decision. Keep the evidence type attached to the term.

Day 04 needs a layered model with honest boundaries, so record a figure-ground decision receipt: the chosen universe; the figure enumerator; the complement enumerator if one exists; the target; the fair scheduling rule; the first side to print the target; and the resulting verdict. Mark the receipt “r.e. only” when no complement enumerator is justified and “recursive/decidable” only when both sides and guaranteed coverage are established.

Checklist:

• [ ] Can you distinguish perceptual figure-ground reversal from the mathematical set distinction?
• [ ] Can you explain why a composite witness gives positive evidence while silence does not certify primality?
• [ ] Can you keep TNT strings, well-formed formulas, sentences, theorems, and truths separate?
• [ ] Can you define r.e. and recursive/decidable without using either term in its own definition?
• [ ] Can you trace the paired-enumerator procedure and explain why fair scheduling matters?
• [ ] Can you state the finite event that certifies each verdict in your receipt?

Day 04 consumes this receipt as a lesson in levels of description. The same object can be viewed as a raw string, a member of a generated set, a sentence under interpretation, or evidence for a decision, and the next chapter asks when one level can safely hide the details beneath it.

A visual-to-mathematical reconstruction can overstate what a picture proves, so provenance and proof status must stay visible. The primary source is Hofstadter’s Chapter III, “Figure and Ground,” printed pages 65–75, read from the supplied local EPUB. The tq/compositeness setup occupies printed pages 65–66 and the figures begin on page 67. Figures 15–18, the r.e./recursive terminology, and the complement argument come from that chapter; all simplified visuals, controls, protocols, applications, paired scheduling trace, and receipt are original conceptual reconstructions.