Can contextual design decisions resist formalization into evaluation rubrics?
This explores whether the judgment calls baked into context-dependent design — what to retrieve, what to remember, what to surface when — can be pinned down into fixed scoring rubrics, or whether something essential leaks out when you try.
This explores a tension between two things the corpus treats separately: design decisions that live in shifting context, and rubrics that want to freeze quality into fixed criteria. The short version the collection points toward: a lot can be formalized — more than skeptics assume — but the formalization keeps capturing the *form* of a good decision rather than its *fit* to the moment, and that gap is where contextual design resists.
Start with why the question even bites. One line argues that AI context is mutable, dynamic, and ephemeral — prompt, history, retrieved data, hidden state all shifting underfoot — unlike the fixed, stable context of conventional software How does AI context differ from conventional software context?. A rubric is the opposite kind of object: stable, portable, applied identically across cases. So the worry is structural — you're trying to score a moving target with a stationary ruler. The corpus's most direct answer is that the fix isn't better rubrics but a different posture: interactive evaluation should be *designed as a paradigm*, with explicit protocols, rather than adopted as a pile of disconnected benchmarks Should interactive evaluation be designed as a unified paradigm?. That's a concession and a counter at once — formalization is possible, but only if it formalizes the evaluation *process*, not just a checklist of outputs.
Here's the surprising part: the collection shows contextual quality formalizing further than you'd guess. Prompt quality, which feels like pure craft, decomposes into six measurable dimensions grounded in communication theory — a structured space where improving one cascades into others, not a flat checklist Can we measure prompt quality independent of model outputs?. Semi-formal reasoning templates act as 'completeness certificates,' forcing explicit premises and evidence checks and catching failures free-form thinking misses Can structured templates make code reasoning more reliable than free-form thinking?. And formal argumentation turns opaque outputs into attack/defense graphs you can actually contest premise by premise Can formal argumentation make AI decisions truly contestable?. So formalization isn't futile — it buys contestability and completeness that vibes-based judgment can't.
But watch where it breaks. The sharpest warning comes from chain-of-thought: logically *invalid* reasoning chains scored nearly as well as valid ones, because the model learned the *form* of reasoning, not genuine inference Does logical validity actually drive chain-of-thought gains?. That's the rubric's blind spot in miniature — a scoring scheme rewards the shape of a good decision while the actual contextual judgment passes through ungraded. The exploration/exploitation work makes the same move from the other side: a trade-off everyone treated as fundamental turned out to be an artifact of *measuring at the token level* — change the measurement frame and the supposed law dissolves Is the exploration-exploitation trade-off actually fundamental?. Formalization doesn't just observe; it constructs what it claims to measure.
The collection's most usable resolution is architectural: don't convert contextual judgment into a dense score, gate with it. Rubrics used as accept/reject gates on whole rollouts prevent reward hacking, while rubrics flattened into dense rewards get gamed Can rubrics and dense rewards work together without hacking?. And when judgment genuinely needs to move with the case, the answer is an evaluator that collects evidence dynamically rather than a static scorer — agent-as-judge cut judge error a hundredfold over a fixed LLM rubric, though its memory module cascaded errors, a reminder that dynamic evaluation buys fidelity at the cost of new failure modes Can agents evaluate AI outputs more reliably than language models?. So the honest answer: contextual design *can* be formalized, but the formalizations that survive contact with shifting context are the ones that stay dynamic — gates, certificates, evidence-collecting judges — not the ones that freeze quality into a number.
Sources 9 notes
AI interactions operate on a substrate of constantly shifting context—prompt, history, retrieved data, hidden state—that users cannot internalize like traditional UIs. This structural mutability demands a new design discipline centered on context engineering rather than interface design.
Interactive evaluation should be treated as a principled paradigm with explicit protocols and reporting standards, not adopted as disconnected benchmarks. The distinction matters: designing interactive evaluation as a unified system prevents fragmentation and incomparability, while expanding what counts as evidence beyond final responses.
Research identifies six evaluable dimensions—Communication, Cognition, Instruction, Logic, Hallucination, and Responsibility—with 20 sub-criteria based on Grice, cognitive load theory, and instructional design. Improvements in one dimension cascade to others, revealing prompt quality as a structured space rather than a flat checklist.
Semi-formal templates requiring explicit premises, code-path traces, and evidence checks improved patch equivalence accuracy from 78% to 88%, catching cases like function shadowing that free-form reasoning missed. Templates act as completeness certificates without formal verification.
Dung-style argumentation structures AI outputs as traversable attack/defense graphs, allowing users to identify and contest specific premises. Standard LLM outputs lack this structure, making it impossible to pinpoint which claims users actually reject.
Illogical chain-of-thought exemplars matched valid CoT performance on BIG-Bench Hard, showing that structural properties—not logical validity—drive the gains. The model learns the form of reasoning, not genuine inference.
Hidden-state analysis using Effective Rank metrics shows near-zero correlation between exploration and exploitation, revealing the trade-off emerges only at token level. VERL demonstrates simultaneous enhancement achieving 21.4% accuracy gains on Gaokao 2024.
DRO shows that using rubrics to accept or reject rollout groups—rather than converting rubric scores into dense rewards—prevents reward hacking. This separation preserves the categorical strength of rubrics while letting token-level rewards optimize within valid answers.
Eight-module agentic evaluation achieved 0.27% judge shift versus 31% for LLM-as-a-Judge on complex tasks. However, the memory module cascaded errors, revealing that agentic systems need error isolation mechanisms to maintain gains.