C.18.1 - Scaling‑Law Lens Binding (SLL)
Status: Stable Type: Pattern
Use this pattern when. Use C.18.1 when a generator, selector, method family, benchmark, or comparison makes an empirical claim that behavior changes with scale—for example, with resource budget, data, model capacity, iteration budget or freedom of action. First name what is scaled, the window and the outcome being compared. Reuse an applicable probe account; new probing is needed only when the required scale claim remains unsupported. If no scale claim is needed, use the ordinary comparison directly. A conditional scale derivation within a stated mathematical model can use its derivation and assumptions directly, without a scale probe. That result supports the model-conditional claim; a claim about observed behavior or the model’s adequacy needs its own applicable evidence.
What goes wrong if missed. Teams compare unequal budgets, call coverage telemetry an objective, claim a knee without probe evidence, or assume more scale means linear improvement across a window where the behavior has already changed.
What this buys. A scale-law lens: declare the scale variables, ScaleWindow, probe points, elasticity class, parity notes, and policy thresholds before treating an empirical scale claim as usable in selection, parity, refresh, shipping, or mathematical-lens work.
Purpose. Declare which scale variables change, the ScaleWindow and the supported behavior class. Numeric fits remain domain work. G.9 governs parity, and a selector policy must admit any decision use of illumination.
Keywords. scaling law; Scale Variables (S); ScaleWindow; knee; diminishing returns; iso‑scale parity; UNM/NormalizationMethod‑based mapping; scale‑probe; DoE (design‑of‑experiments); segmented regression; knee detection.
C.18.1:1 - Problem frame
Teams often say a method “scales” without disclosing which resources, across what window, and how outcomes respond (convex rise → knee → plateau). Without that, parity is skewed (unequal budgets, unmatched windows), coverage/illumination report-metrics leak into dominance, and “knees” are found late. SLL supplies a notation‑independent lens to make scale behaviour explicit and comparable.
C.18.1:2 - Problem
Omitting Scale Variables and the comparison window causes: (i) unfair parity (compute/data/FoA mismatched), (ii) illumination/coverage report-metric creep into dominance by default, (iii) late detection of knees and budget waste. G.9 already forbids scalarising mixed scales and mandates equal FreshnessWindows/pinned editions; SLL complements this with ScaleWindow & elasticity.
C.18.1:3 - Forces
Notation independence vs useful scaling heuristics; local context vs cross‑context generality; telemetry vs objectives (illumination stays report‑only telemetry unless policy promotes it); early exploration vs reproducible policy.
C.18.1:4 - Solution — binding lens for generator/selector profiles (normative)
C.18.1:4.1 - Profile annotation
SLL.Profile annotates a MethodFamily/Generator or Selector profile with the scale claim and its basis.
C.18.1:4.2 - Fields (conceptual descriptors).
- S — Scale Variables. Minimal set of adjustable variables with a declared order for the Context:
compute(steps/tokens/FLOPs/time/energy),data(size/quality),model capacity(params/branches),iteration budget,freedom‑of‑action (FoA)/environment richness, etc. Declare each variable’s Characteristic and Scale, with a unit when applicable under C.16, and bindSto a ScaleWindow. Ordering the input does not assume a monotone improvement in the outcome; do not treat qualitative data quality or freedom of action as a ratio quantity without that measurement basis. Keep planned budget values with A.15.2; bind dated resource-use accounts to A.15.1, B.1.6, and A.10. Where training/inference trade, name the phase the claim concerns. - ScaleWindow. The declared range of resource settings under investigation. Identify the settings actually tested and the response, statistic, task family and conditions to which the result applies. To extend a conclusion to untested settings, state the model, bound or other grounds for that extension.
FreshnessWindowgoverns the age of the evidence. - Scale‑Probe. At least two (preferably ≥ 3) parity-respecting points in
Swithin the ScaleWindow, with uncertainty appropriate to the empirical claim. Use replicates, seeds and confidence intervals when the trial design and variability require them; deterministic results retain their conditions and any applicable measurement or numerical error. Pick points via a small factorial or Latin-hypercube when that design answers the multi-knob question. - ElasticityClass
χ ∈ {rising, knee, flat, declining}— a qualitative class supported by the probe at the settings supported within the declared ScaleWindow. Leaveχunassigned when the observations and their uncertainty do not support a class, and state what remains unresolved. Numeric exponents/fits live in domain annexes, not Core. - ParityNotes.
iso‑scale parity?flag and loss notes if not achieved, plus applicable Bridge, Φ, and Ψ IDs for actual crossings under G.9; any numerical assurance penalty follows its applicable B.3 model.
C.18.1:4.3 - Norms (SLL).
- SLL‑1 (Declaration). Any profile claiming scale behaviour SHALL declare
Sand a ScaleWindow for the Context. - SLL‑2 (Probe). An empirical conclusion about a change with scale SHALL use an applicable scale‑probe (≥ 2 points in
S, with the applicable uncertainty treatment in §4.2) and record the supported χ, or the unresolved classification. An existing probe may supply this basis while its window, definitions and conditions remain applicable. Multi‑knob probes SHALL hold unspecified knobs fixed or pinned, and disclose invariants. A result at one setting can establish what happened there. Repeat trials at that setting can improve the estimate for that setting. A change with scale remains open until the evidence covers distinct settings. - SLL‑3 (Parity). A comparison claiming matched scale SHALL establish iso‑scale parity. When heterogeneous knobs such as FLOPs and tokens require a mapping, apply a justified UNM/NormalizationMethod mapping before comparing outcomes under G.9; their names alone supply no conversion. Keep FreshnessWindows/editions equal/pinned as required by G.9. If iso-scale parity cannot be established, state the loss and restrict the comparison to what its applicable rule admits. Loss notes alone do not make an otherwise unsupported comparison admissible. Record any seeds/replicates used, the uncertainty basis, ComparatorSet and applicable policy ids in telemetry/SCR.
- SLL‑4 (Selection lens). Within the same Context and ScaleWindow, if novelty, use value and constraint fit are tied, selectors MAY use illumination as a tie-breaker when an applicable pinned policy admits that bounded use. This SHALL NOT change default dominance. Promotion into dominance needs its own explicit CAL policy; otherwise illumination remains report-only telemetry.
- SLL‑5 (Knee test). A knee is claimed only when a monotone rise is followed by a slope drop across adjacent intervals formed by at least three distinct probe points within the ScaleWindow. Compare slopes along one ordered resource coordinate with other influential conditions held comparable. If several resources change together, define the path and its unit of advance, and attribute the knee to that path. The slope comparison must be admissible for the declared Scales and establish a drop exceeding the policy-defined threshold under the uncertainty or error treatment appropriate to the probe. For stochastic probes, establish statistical significance under the cited policy using a method whose assumptions fit the trial design; report the significance level and any confidence-interval level used. For deterministic probes, establish the threshold exceedance throughout the applicable error bounds. Cite the threshold policy (E/E-LOG), comparison method and uncertainty basis. Failure to establish a knee leaves that claim unsupported. Report rising, flat or declining only when the probe supports that class; otherwise leave χ unassigned. Absence of a confirmed knee is not evidence of rising performance.
- SLL‑6 (Telemetry invariants). Probes SHALL export the uncertainty basis, any seeds/replicates used, edition pins, applicable policy ids and resource-account units governed by C.16 and B.1.6, with dated-work and provenance links under A.15.1 and A.10, to G.11.
C.18.1:4.4 - Method — minimal SoTA probe recipe (notation‑agnostic; informative).
- Choose ordered scale variables
Sfor the Context, such as compute, data, capacity or FoA, under their declared measurement basis. - Choose attainable settings that can distinguish the effects in question. Three to five settings can reveal curvature along a single ordered resource. A two-state variable permits a two-state comparison. For several resources, choose a design that can distinguish the relevant main effects and interactions.
- For stochastic trials, choose replication and an uncertainty method adequate to the claimed trend; bootstrap intervals are one option when their assumptions fit the data. Retain the seeds used. For deterministic trials, retain the conditions and applicable error bounds.
- Estimate the contrast needed for the claim. For a knee, use the adjacent slopes in SLL‑5. Logarithmic slopes require positive values and operations admitted by the Scales under
A.18. A segmented fit or a knee detector can assist that comparison. - Record the response and statistic, task family, settings tested and conditions held fixed. Make the evidence available through the profile or
SLL.Card@Context. An observed run demonstrates that its outcome occurred under those conditions. A feasible solution with a verified objective value supplies a lower bound when maximizing that objective. Meeting a target with a verified amount of an ordered resource supplies an upper bound on the minimum needed under the same conditions. If several resources are tracked separately, report the feasible combination. A claim about expected performance, population coverage or a required success probability needs evidence for that statistic. - If χ changes across the window, split the ScaleWindow and re‑classify per segment.
C.18.1:5 - Consumer relation fields - minimal inputs and outputs (conceptual)
G.9 parity planning uses S and ScaleWindow to align budgets, pin editions, and declare any required UNM or NormalizationMethod mapping; run evidence records the applied mapping and pins. G.11 carries policy-id, PathSliceId, edition pins, the uncertainty or error basis, and any seeds, replicates or CI level used, under the applicable parity conditions.
C.18.1:9 - Archetypal Grounding (post-2015; informative)
- LLM scaling. Historical Kaplan-style and Chinchilla-optimal regimes illustrate different compute/data allocations. Mixture-of-Experts and retrieval-augmented families require their own capacity and inference-budget account.
- RL/Planning. Model-based optimization & general agents vs hand-tuned controllers; slopes reported wrt budget/FoA under safety envelopes.
- QD/OEE. MAP-Elites, CMA-ME, DQD, QDax; POET/Enhanced-POET families: coverage/illumination as telemetry metrics; parity uses fixed grids/spaces and edition pins.
Constructed classification case. A matched-probe account supports unchanged performance across the declared resource window, with uncertainty small enough for the comparison’s stated tolerance. Report flat; the absence of a knee does not turn it into rising. In a second account, the uncertainty still permits both an increase and a decrease that would matter to the choice. Leave χ unassigned and retain that limitation. A supported monotone increase without a confirmed slope drop can instead support rising. These supplied accounts illustrate classification and its limit; they are not empirical scaling results.
Constructed bounded-error knee. At resource counts 1, 2, 3, deterministic throughputs are 10, 18, 19, each with absolute error at most 0.1, under the same tested conditions and admissible linear scales. The first adjacent slope is in [7.8, 8.2]; the second is in [0.8, 1.2]. Both rises are positive and the slope drop is in [6.6, 7.4]. A declared drop threshold of 5 is therefore exceeded throughout these bounds, supporting a knee over these three probe counts. No interpolation beyond the admitted counts is claimed. If wider error bounds admit a drop of 5 or less, this knee conclusion is no longer established; use only the behavior class still supported by those observations and bounds.
C.18.1:7 - Bias-Annotation
| Bias | Symptom | Correction |
|---|---|---|
| Bigger-is-better bias | More compute, data, capacity, or freedom of action is treated as automatic improvement. | Declare S, ScaleWindow, and elasticity class before using the scale claim. |
| Telemetry-as-objective bias | Coverage or illumination is promoted into dominance by default. | Keep telemetry report-only unless the selector policy explicitly admits it. |
| Knee-by-story bias | A plateau or knee is asserted from one anecdote or one late observation. | Require scale-probe points, replicates or uncertainty, and a cited threshold policy. |
C.18.1:6 - Conformance Checklist (CC-SLL)
Sdeclared orS = N/Awith rationale.- For an empirical scale comparison, give the probe settings, response, task family, invariants and applicable uncertainty. Name the supported class or the remaining uncertainty. With observations at only one setting, report what is established there and leave the scale comparison open.
- A matched-scale claim has established iso-scale parity, with the editions, seeds when used, and comparator recoverable. If the resource regimes cannot be matched, report the individual results or a narrower comparison whose grounds hold. Any numerical assurance penalty has a basis in the receiving B.3 model.
- An illumination tie-break cites the policy admitting it and its lens id in E/E-LOG provenance. When that policy depends on χ, include the supported class or leave the dependent preference open.
- Knee claims cite the policy threshold, slope-comparison method and applicable uncertainty or error basis; include the statistical significance level and any CI level when those are used.
C.18.1:8 - Common Anti-Patterns and How to Avoid Them
Hidden budget mismatches; averaging ordinals across families; illumination in dominance by default; unpinned editions; slope claims without an applicable uncertainty basis; training/inference phase mixing → cure with G.9 parity (equal windows/editions; normalize‑then‑compare; return sets), phase‑label the claim, and record slope uncertainty per Scale‑Audit discipline.
C.18.1:10 - Payload — exports
SLL.Card@Context (UTS row; editioned):
⟨S{knobs, applicable units, phase when relevant}, ScaleWindow, Scale‑Probe{points≥2, design, uncertainty basis, seeds?, CI?}, ElasticityClass χ?, ParityNotes{iso‑scale?|loss, invariants}, BridgeIds?/Φ/Ψ, PolicyIds? (E/E‑LOG), PathSliceId?⟩.
UTS row template (conceptual).
SLL.Card@Context := S=(ordered variables; applicable units=…; phase when relevant=…), ScaleWindow=[LOW…HIGH], Probe=(points=…, design=…, uncertainty basis=…, seeds?=…, CI?=…), χ?=rising|knee|flat|declining, ParityNotes=(iso=true|false; invariants=…), Bridge/Φ/Ψ=(…), PolicyIds=(…), PathSliceId=(…).
Here χ? means that the class is omitted when unresolved; uncertainty is not a fifth behavior class. Keep the probe account and its limitation available to the receiving comparison. An unassigned class cannot support a class-dependent preference. Use the linked profile or the probe account to recover the response statistic, task family, tested settings and conditions, and the grounds for any conclusion between or beyond those settings.
C.18.1:11 - Consequences
Benefits. A comparison can show which knobs were scaled, what window is covered, how much probe evidence supports the slope class, and what parity or normalization losses are declared, with any numerical assurance penalties justified by the receiving B.3 model. Dominance changes require their own admitting policy.
Trade-offs. An empirical comparison across resource settings needs observations at at least two scale points and recoverable invariants, applicable phase and uncertainty information, any seeds used, and applicable policy thresholds. Reuse of an applicable account can avoid new probing. The gain is that selectors, parity harnesses, refresh telemetry, and mathematical-lens uses can cite one bounded scale claim instead of guessing whether the observed behavior transfers.
Stop condition. Stop at C.18.1 when the scale variable, ScaleWindow, probe basis, supported class or explicit classification limit, and parity notes suffice for the receiving question. An unresolved class can finish this inquiry while leaving a scale-dependent preference unsupported. Select further probing only when its attainable contribution to the decision warrants its burden; use C.11.DUA if that choice needs resolution. Move to G.9, C.19, G.11, C.29, or a domain annex when the next question concerns parity, selector policy, telemetry refresh, a mathematical lens, or a numeric fit.
C.18.1:12 - Rationale
C.18.1 exists because scale claims are easy to overread as universal improvement claims. The pattern keeps scale behavior bounded by a declared scale variable, scale window, probe basis, uncertainty, elasticity class, and parity notes before the claim is reused.
C.18.1:12.1 - SoTA-Echoing
Which evidence is sufficient for a scale claim? Retain the distinction in C.29 between a consequence inside a declared mathematical model and a claim about the modelled phenomenon. The conditional cost equations in C.19.2:5 establish their break-even horizon from the supplied assumptions; empirical probes would answer whether those assumptions describe actual handoffs. Requiring a probe for the derivation adds work without strengthening that conditional result, while using the derivation as observed performance omits the needed evidence. This pattern therefore limits its probe requirements to empirical scale claims and adapts C.19.1’s claim-matched evidence rule: keep parity and uncertainty conditions, and reuse an adequate account. Reopen when the assertion extends beyond the model conditions, the empirical window or the supported uncertainty.
Scaling-law fits, knee detectors, segmented regressions and experimental-design methods support the scale claim under their own conditions. Keep the measurement, parity and selector-policy grounds required by the use. Apply C.29 only when a mathematical representation choice, transfer or reliance question remains unresolved.
C.18.1:13 - Relations
C.27 temporal-claim relation.
- C.27 may flag: a claim that more review capacity, tool calls, tokens, data, model capacity, parallelism, freedom of action, sprints, or another declared scale variable changes rate, learning, recovery, throughput, stabilization, or improvement.
- This pattern keeps: scale variable, scale window, scale probes, and elasticity value.
- Non-admissible use: more scale does not by itself establish linear improvement, and a scale word does not create a C.27 rate-change claim by itself.
- Neighboring-pattern use: for a comparison or benchmark, cite G.9 for parity; if linear improvement is assumed, name the scale variable and scale window or state the improvement claim as an assumption.
Builds on: C.16/17/18.
Coordinates with: C.19 (lenses/policies), G.5 (set‑returning selector), G.9 (parity; ParetoOnly default; UNM/NormalizationMethod‑based mapping), G.10 (shipping).
Use C.24 only when a fixed action or option needs tool or service call planning; a scale claim about budgets or freedom of action alone does not require that plan.
Pedagogical cue. Say what is scaled, use the probe basis and steer only by the behavior it supports.
C.18.1:13a - C.29 mathematical-lens use relation
C.18.1supplies empirical scale-window and scaling-law evidence forC.29when a mathematical-lens use claims observed scale behavior, universality, knees, exponents, coarse-graining validity or diminishing returns. Apply that evidence requirement to the empirical claim; an adequate conditional derivation within a declared model retains its mathematical grounds.C.19.1governs a separately claimed method scale preference. If no such claim is live, keep the local result and its stop condition.