Why the Curve Has This Shape

A fair objection to the R/K/G equation is that it looks arbitrary. Why a logarithm for reasoning? Why a power law for knowledge? Why not a sigmoid, or an exponential? The honest answer is that the specific curves are not sacred — but they are not pulled from nowhere either. Each one is the simplest shape consistent with a single, defensible statement about how that factor behaves. This note makes those statements explicit and walks from each to its curve, so that a reader can argue with the assumptions rather than the algebra. That is the argument worth having.

The starting picture

Recall the three observations the framework opens with. A brilliant reasoner with no relevant knowledge is often confidently wrong. An ordinary reasoner with a well-indexed library usually wins on real tasks. And either of them, made to work through a structured process — show the work, check each step, consider the counterargument — improves without getting smarter or learning anything new. Three factors, pulling differently. The task is to write down the simplest math that respects all three.

Before the individual curves, one structural choice. We assume the reasoning contribution and the knowledge contribution combine multiplicatively, not additively — that quality looks like (something about R) × (something about K), rather than a sum.

The reason is the genius-with-no-knowledge case. If the terms added, a brilliant reasoner would still score respectably on knowledge they lack, because the reasoning term alone would carry them. Multiplication encodes the harder truth: a severe deficiency in either factor strongly suppresses quality — when knowledge is near zero, quality is near zero however strong the reasoning, because anything times almost-nothing is almost-nothing.

This is a modeling choice about interaction, not a literal claim that quality hits exactly zero. A system with heuristics and latent priors can score above the floor on thin explicit knowledge, so an empirical fit may well need a small baseline or interaction-residual term — Q = ε + f(R)·g(K) — beyond the clean product. We keep the bare product here because it captures the interaction with the fewest moving parts, and flag the residual as the kind of thing calibration adds. That is Assumption 0, worth stating because a reviewer will ask why the factors multiply.

R → a logarithm, from diminishing returns

Assumption. Each additional unit of reasoning capability is worth less than the last. Past a threshold, more raw horsepower buys little; the tenth clever insight moves the answer less than the first.

That sentence is most of the derivation, but it needs one honest addition. Diminishing returns by itself only says the curve bends the right way — its second derivative is negative (f″ < 0) — which a square root and a saturating exponential satisfy just as well as a logarithm. It does not, on its own, pick a function. So we choose the simplest marginal-return schedule consistent with it: an inverse-linear one, where the value of the next unit falls off like one-over-what-you-have.

That choice is what makes the rest follow. Summing the shrinking increments — integrating — gives a logarithm:

So the logarithm is not forced by diminishing returns; it is forced by the inverse-linear schedule we selected one step up — the simplest rate that starts at zero, never turns back down, and flattens without a hard ceiling. A square root or a saturating sigmoid would also honor the assumption. Whether the real curve is exactly a logarithm is a question for data, not for this page.

K → a power law, from compounding

A definition first, because it carries the weight. K is not the size of a corpus or the number of documents in a context window. It is knowledge sufficiency: the accurate, relevant, retrievable knowledge actually available to the task. Doubling the documents while halving their relevance does not raise K — it may lower it. Read as raw volume, the claim below would be plainly false, since retrieval noise, contradiction, and context saturation degrade quality past a point. Read as usable, task-relevant knowledge, it holds.

Assumption. Each unit of that usable knowledge is worth more in the presence of the knowledge it connects to. Facts are not stored in isolation; a new one links to what is already there, and the number of useful connections grows faster than the number of facts.

This is the mirror image of the reasoning story. There the marginal return fell; here it rises — the next unit is worth more than the last. A rising marginal return is the signature of a power with an exponent above one:

The exponent β is where “compounding” lives: β = 1 would be a straight line (each fact worth the same as the last, no compounding), and β > 1 bends the curve upward. We do not claim to know β. We claim only that if knowledge compounds at all, β sits above one, and the curve accelerates. The exact value is the empirical question — and, tellingly, the one skeptics reach for first. That it can be fit rather than assumed is the point.

G → linear amplification, from closing the gap

Assumption. Structure substitutes for what the base system lacks. A reasoning graph does not add a fourth ingredient; it lifts the reasoning and knowledge a system can effectively bring to bear — and it lifts them most when there is the most room to lift.

“Most room to lift” is the operative phrase. If a system is already near the ceiling on reasoning, there is little for structure to add; if it is far below, there is a lot. The cleanest way to write “close a fraction of the remaining gap” is linear in that gap:

This is the term that makes the framework’s central prediction fall out for free: because the lift is proportional to the gap, a weak system (large gap) gains a lot and a strong system (small gap) gains little. The claim that structured reasoning helps smaller models more than frontier ones is not an extra assumption — it is what “close a fraction of the remaining gap” means. Whether the closing is truly linear, or curved, is again for data; linearity is the simplest gap-closing rule, not a measured one.

One caveat the clean form hides. As written, G = 1 drives both effective values to 1 — as if perfect structure could fully compensate for arbitrarily weak reasoning or knowledge. That is almost certainly too strong, and the parent framework is explicit that structure lifts a weak base, it does not replace it. Read G = 1 as the theoretical maximum of structure-mediated effectiveness, not an achievable one; an empirical fit would likely carry a closure ceiling — a factor λ < 1 on the gap-closing term, R_eff = R + λG(1−R) — that stops a perfect checklist from turning a toaster into Einstein. We leave λ off the clean equation on purpose and flag it here.

Putting it together

Substitute the effective values into the multiplicative form and the full model appears. One cleanup first: a bare log term is not bounded by 1, so if Q is to stay genuinely on 0–1 we divide it by its own maximum, log(1 + α). Now both factors run 0–1 as their inputs do:

Dividing by a constant does not change any of the shape arguments — the marginal-return schedule, the diminishing curve, the gap-closing lift are all untouched — it just keeps Q honestly normalized rather than merely called so. Every piece now has a reason behind it: multiplicative because a deficiency in either factor suppresses quality, log because reasoning has diminishing returns, power-law because usable knowledge compounds, gap-linear because structure helps the weak most. None of it is arbitrary. All of it is a choice of the simplest representative.

The simplest, not the only

Here is the part that matters for keeping the argument honest. Every step above chose a family, and every choice has siblings that satisfy the same assumption. Diminishing returns on R is satisfied by a logarithm, a square root, and a saturating sigmoid alike. Compounding on K is satisfied by any exponent above one, and by other accelerating curves besides. Gap-closing by G could be linear or gently curved. The assumptions pin down the shape — rising-and-flattening, accelerating, gap-proportional — not the exact function.

So the equation is best read as one member of a family: the simplest curves that respect three qualitative commitments. What would settle the specific forms is not more argument but measurement — fitting the family to matched tasks where R, K, and G are varied and Q is scored, and seeing which member the data prefers. That is why the parent framework labels the functional forms an assumption (A4) rather than a result, and why we would rather concede the exact curves now than defend them prematurely. The contribution the framework is making is not “β equals some number.” It is the prior claim that R, K, and G are separable factors worth measuring at all.