Making a food healthier usually makes it worse to eat. Bran hardens bread, mushroom fiber softens meat patties, and polyphenols turn astringent. The binding constraint in reformulation is not generating candidate recipes, which is cheap. It is finding out whether people will eat them. Human sensory panels taste only a handful of samples per session before palate fatigue sets in, and instruments miss masking, synergy and in-mouth breakdown. This note proposes to borrow the decision structure of a journal editorial board, not its conversational style. Cheap specialized reviewers screen many candidates. A middle editor compares their disagreements. Only contested candidates get "sent out" to the expensive human panel. The question is whether this tiered triage finds better healthy-and-palatable formulations per panel session than flat ensembles, single models, or classical statistical design.
In two sentences: journals handle thousands of submissions with a small pool of expert reviewers by triaging, desk-rejecting obvious failures, and escalating only the uncertain cases. We build the same tiered routing for food reformulation and measure, against a fully tasted ground-truth pool, whether it reaches the sensory-nutritional tradeoff curve with fewer human tasting sessions.
Core Claim
The research file's obvious intersection is a multi-agent "editorial board" of language-model personas that critiques and rewrites recipes. No paper in the file does this. We think it is the wrong target anyway, because it has no ground truth. A board of models agreeing that a high-fiber cookie is pleasant tells us nothing about whether it is. The useful part of editorial structure is the triage logic under a scarce expert budget. We claim that when evaluations come at three fidelities, a tiered board with explicit accept, reject and escalate decisions dominates flat allocation in panel efficiency. The three fidelities are model-based screening (nearly free), instrumental texture analysis (cheap), and trained human panels (scarce). The decisive comparison holds the panel budget fixed.
The specific mechanistic bet is about the escalation signal. We expect disagreement between reviewers grounded in different evidence to predict panel surprises better than any single model's stated confidence. Here, different evidence means a texture reviewer fit to instrument data, an aroma and astringency reviewer, and a nutrition reviewer. Single-model confidence is known to be miscalibrated in peer-review agent systems. The file flags this as an open problem.
Relation to Prior Work
Three editorial-structure papers supply the method, and each was evaluated only on text. MARG splits manuscript review into section-holding workers, criterion experts who write their expectations before reading, a leader, and a fact-checking refinement group. It reports 3.7 "good" comments per paper versus 1.7 for single-agent GPT-4, and cuts generic comments from 60% to 29%. Multi-agent peer review collaboration shows single-model self-correction can collapse: Colored Objects accuracy fell from 66% to 28%. Cross-agent review with 1–10 confidence scores reached 79.45% on Penguins versus 47.95%. FIRSTPASS shows that whether a manuscript needs extended review is predictable from multi-round review text. A fine-tuned Qwen2.5-7B reached 80.5% accuracy and 78.2% macro-F1.
We reuse MARG's pre-committed criteria and FIRSTPASS's framing of "standard versus extended review" as a predictable routing decision. Our departure is that the reviewers' verdicts are checked against physical measurements and human tasting, not against reviewer text. Hierarchy is valuable here only if it saves panel sessions. Organized LLM teams motivates the hierarchy: a designated leader improved task efficiency by 9.76% in VirtualHome-Social while keeping communication cost in check.
On the food side, RecipeCrit edits recipes by gradient steps on a latent text representation. It reports 30–50% relative gains in ingredient-addition success over BART, RecipeGPT and PPLM, but gives no guarantee of physical realizability. NGQA reasons over conflicting nutrient-health rules with Think-on-Graph but ignores palatability. FEAST on WineSensed aligns text and image embeddings with human flavor distances, reaching 0.91 triplet agreement versus 0.82 for single-modality baselines, but never intervenes on a formulation. These papers supply candidate generators and reviewer components. None allocates human evaluation.
The food chemistry papers constrain the design. Fiber dilutes gluten disulfide networks and raises hardness (soybean-residue cookies; reconstituted whole wheat). Mushroom stalks soften and decohere pork patties at 25–100% substitution (Agaricus patties). The instrumental-sensory gap is documented: most GC-MS volatiles fall below perception thresholds while single trace odorants dominate (Zhi et al.), and synergy and masking escape sensor arrays (Wang et al.). Together these say a texture or aroma reviewer reading only text or concentrations will err systematically. That is why our reviewers are fit to measurements, and why escalation to humans cannot be eliminated, only rationed.
Research Program
We first build the ground truth that does not exist in the file. The search turned up no reported result on Lacuna for panel-evaluated food reformulation benchmarks. We take two matrices whose penalties are well documented: fiber-fortified wheat cookies and bread rolls, and pork patties with mushroom-stalk substitution. For each, we assemble a pool of roughly 120 candidate formulations drawn from several sources:
- RecipeCrit edits seeded from Recipe1M.
- A prompted large language model.
- Food-scientist proposals.
- A grid of substitution levels.
Each candidate's nutrition is computed from USDA composition, as in pic2kcal. Every candidate is then fully measured. Texture profile analysis records hardness, springiness and cohesiveness. A trained panel rates descriptive attributes and overall acceptance on a 9-point hedonic scale, using a balanced incomplete block design to respect fatigue. The result is a replay benchmark: any allocation policy can be simulated by revealing a candidate's panel outcome only when the policy pays for it.
The board has three tiers.
- Reviewers. Three specialists, each of which writes its acceptance criteria before seeing a candidate, in MARG's style:
- A texture reviewer, a small regressor from formulation features to instrument texture, retrained as instrument data arrive.
- An aroma and astringency reviewer, a language model given threshold and polyphenol knowledge, with an optional embedding check against WineSensed-style human distances.
- A nutrition reviewer, deterministic from composition tables, with NGQA-style health rules.
- Associate editor. Combines the three reports into an estimate $p(x)$, the probability that candidate $x$ is both acceptable (panel mean at or above a preset threshold) and meets the nutrition target. It also computes $d(x)$, the spread of the reviewers' normalized verdicts.
- Editor. Holds the budget and decides, for each candidate, whether to accept, reject, send to the instrument, or send to the panel.
Its rule is to escalate the candidates with the highest
$$s(x) = d(x)\, p(x)\,(1 - p(x))$$
Here the product $p(1-p)$ is largest near the decision boundary, and $d(x)$ upweights cases where grounded reviewers conflict. Every panel result recalibrates $p$ by isotonic regression. A secondary loop lets an author agent revise rejected-but-close candidates once ("major revision"). The revisions are drawn from the pool's neighbors, so outcomes stay measurable.
Baselines
The naive reference is uniform random allocation: spend the panel budget on randomly chosen candidates and report the best one found. Every policy draws from the same pool, gets the same budget of panel slots, and gets the same instrument budget, so the comparison is purely about allocation.
The strongest text-only competitor is a single frontier language model with the same nutrition tool and the same written criteria, run with self-critique. It ranks candidates and spends the budget on its top choices. This is what a practitioner would deploy today. Given Xu et al.'s self-correction collapse, we give it the same number of model calls as the board, so it does not lose for lack of compute.
The strongest structural competitor is a flat ensemble: the same three reviewers, scores averaged, with escalation by mean uncertainty. It shares every component except the editor tiers and the disagreement term. This makes it the ablation that isolates our key ingredient.
The strongest classical competitor is sequential design over a low-dimensional parameterization (substitution level, water, one binder) on the instrument and panel data. It is response-surface methodology as in the purple basil study and the rabri study, plus a Gaussian-process Bayesian optimizer with matched evaluations. We found no reported result on Lacuna for either on a panel-budgeted reformulation task.
Two further ablations isolate parts of the board. One replaces the measurement-fit texture reviewer with a language-model texture persona. The other removes the pre-committed criteria.
Beating these baselines means two things at panel budgets of 10, 20 and 30 of 120 slots:
- A higher dominated hypervolume, relative to the full-information Pareto front of nutrition gain against panel acceptance. The improvement over the flat ensemble must exceed its bootstrap confidence interval across matrices.
- A lower rate of "certified acceptable" formulations that the panel actually rejects.
The closest published reference numbers are MARG's 3.7 versus 1.7 good comments and FIRSTPASS's 80.5% routing accuracy. Neither is commensurable with our metrics, which is exactly the gap.
Why not just…?
Why not just use one strong language model with a nutrition calculator? A single model is cheap, and with tools it can enforce nutrient targets well, so it may tie on the nutrition axis. The risk is the sensory axis. Xu et al. show single-model self-correction can drive accuracy down rather than up. MARG shows single-agent critique skews generic. Our matched-call baseline settles the question directly. If the single model matches the board's hypervolume at every budget, we report that the hierarchy adds nothing.
Why not just average an ensemble of reviewers? This is the real threat. Most of any gain could come from having grounded reviewers at all, not from tiers. The flat ensemble differs from the board only in the editor rule and the disagreement weighting. If it ties, the honest conclusion is that grounding matters and editorial structure does not. We would then reframe the contribution as the benchmark plus grounded reviewers.
Why not just run response-surface methodology or Bayesian optimization? In a one-dimensional substitution sweep, such as the patty series at 25–100%, classical design should win or tie. We expect that and will report it. The board's claimed advantage is restricted to pools that mix structurally different edits: swapped flours, added binders, changed processing such as the microwave and steam bran treatments. A fixed parameterization cannot represent those edits. We test both regimes separately, so the claim is falsifiable in each.
Proposed Experiments
The primary experiment replays all policies on the two fully measured pools across the three budgets, with 50 random seeds for the random policies. It reports hypervolume, false-certification rate, and panel slots needed to find the first formulation that meets both the nutrition target and the acceptance threshold.
The decisive check targets the mechanism. On held-out candidates, we compute how well reviewer disagreement $d(x)$ predicts the absolute gap between the board's predicted and the panel's observed acceptance. We compare that against the single model's self-reported 1–10 confidence. If disagreement is no better a predictor, the escalation rule has no basis, regardless of the headline numbers.
A transfer check trains the escalation calibration on cookies and applies it unchanged to patties. This asks whether the editor's rule generalizes across matrices or is merely fit to one.
A cost accounting reports model tokens alongside panel slots, since MARG needed over 1.2 million input tokens per review. Board overhead must stay small next to the cost of a human session for the efficiency claim to hold.
References
- MARG: Multi-Agent Review Generation for Scientific Papers
- Towards Reasoning in Large Language Models via Multi-Agent Peer Review Collaboration
- FIRSTPASS: A Multi-Domain, Multi-Round Peer Review Dataset Grounded in Real Editorial Outcomes
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- Assistive Recipe Editing through Critiquing
- NGQA: A Nutritional Graph Question Answering Benchmark for Personalized Health-aware Nutritional Reasoning
- Learning to Taste: A Multimodal Wine Dataset
- Digestibility, textural and sensory characteristics of cookies made from residues of enzyme-assisted aqueous extraction of soybeans
- Microwave and Steam Processing: A Novel Approach to Modifying the Characteristics of Reconstituted Whole Wheat Flour and Dough
- Physicochemical and Sensory Properties of Pork Patties with Partial Replacement of Lean Pork by Stalks of Agaricus bisporus
- Predicting Sensory Quality of Longjing Tea on the Basis of Physiochemical Data
- Emission characteristics and health risks of volatile organic compounds and odor from PVC-overlaid particleboard
- Multi-Task Learning for Calorie Prediction on a Novel Large-Scale Recipe Dataset Enriched with Nutritional Information
- Leveraging Response Surface Methodology, Variance Analysis, and Purple Basil for Process Optimization
- Optimization and Storage Study of Rabri Enriched with Date Syrup and Makhana