Desk Rejection for Supply Chains: Exception-Triggered Hierarchical Review of Orders Under Private Information

Supply chains with many tiers suffer from the bullwhip effect, where small swings in customer demand grow into large swings in factory orders as each firm reacts to the orders of the firm below it rather than to real demand. Sharing raw demand and inventory data across firms is the textbook cure. Recent work shows that language-model agents that share information or negotiate cut costs sharply in a simple linear chain (Agentic LLMs in the Supply Chain). Real firms, however, refuse to expose proprietary inventory and cost data, and when suppliers ration scarce stock they have reasons to inflate what they report. Academic journals face the same problem in another setting. Many independent experts hold private judgments, and the system needs a reliable global decision without exposing everyone to everyone, and without the delay of sending every case to the top. Journals solve this with a layered board. Associate editors read confidential referee reports, pass only a summary upward, and handle most cases themselves. The editor-in-chief sets global policy and sees only the cases that are escalated. Most submissions never reach the top tier, and some are rejected at the desk.

This note proposes treating that structure as an explicit, tunable design for multi-tier supply networks. Each firm reports confidentially to an "echelon editor," an intermediary for one tier. The echelon editor checks each report against later realized sales, keeps a credibility score for every reporter, and passes only aggregated forecasts upward. A single top-level editor enforces network-wide constraints such as a carbon cap. The key ingredient is the desk-rejection rule. Each echelon editor decides most orders locally and escalates only orders that look anomalous against its credibility-weighted forecast. The scientific question is quantitative. How much of the benefit of full information sharing can this structure recover while revealing only aggregates? How does the best review depth and escalation threshold shift as demand shocks become more frequent, and as incentives to misreport become stronger?

Core Claim

We claim that an exception-triggered two-tier review structure dominates both flat peer negotiation and a single central planner, under one condition: firms keep raw data private and have an incentive to misreport. Its advantage should be largest when shocks are rare but large, and smallest when shocks are frequent. In the frequent-shock regime, every extra review tier costs a period of delay that outweighs the fidelity it adds. When agents are honest and can share raw data freely, we expect flat information sharing to match the hierarchy. That outcome is a prediction we will test, not a weakness we will argue around.

The novelty is not "add a manager agent." It is three things together. First, we treat review depth, aggregation, and escalation threshold as variables to measure. Second, we impose a realistic privacy constraint. Third, we chart the trade-off between oversight and delay that the governance literature describes but has never quantified for operations (Gap 3 of the file).

Relation to Prior Work

Three works motivate the gap. Agentic LLMs in the Supply Chain has neighbor agents in a linear chain share data and negotiate. It observes that downstream agents become "stubborn" and upstream agents "suggestible." It has no tier that arbitrates disputes or enforces network-wide limits, and it assumes agents are willing to share raw data. We remove that assumption and add the missing tier. Kuan et al. build a chain of custody for carbon data, in which suppliers author records, contractors validate them, consultants endorse them, and the developer finalizes them, with selective disclosure of trade secrets. It is a static ledger for a single project, reviewed by humans. We borrow its idea of "trust without exposure" and apply it to recurring replenishment decisions, where review delay actually costs money. Oh et al. design mechanisms for one supplier to elicit credible forecasts from one buyer, and note that misrepresentation compounds across tiers. Our echelon editor's credibility score is a learned, multi-party replacement for such a mechanism.

Other work constrains the design. Embodied LLM Agents Learn to Cooperate in Organized Teams finds that a designated leader improves completion time, for example by 9.76% for teams of three GPT-3.5 agents. Those agents were fully cooperative and doing household tasks, so the result does not show that hierarchy helps when agents have conflicting interests. Policy-as-Prompt shows that a policy given to a language model in different formats with the same meaning yields different decisions on edge cases, at similar average accuracy of about 0.79. This is why our main experiments use learned policies rather than language-model editors, with language models only in a secondary replication.

Some work supplies methods we reuse. MORL/D provides the Messiah simulator: networks of 7, 13, and 24 nodes, scored on three objectives, namely profit, carbon emissions, and inequality in service level across markets. It also provides the strongest centralized learner. MIRACL splits the network into subproblems to speed up learning. That split is a computational convenience, not a privacy boundary, and we differ by drawing the split along ownership lines. Picard iteration makes the long simulated runs we need affordable to compute. The journal front-matter papers, such as IEEE Transactions on Information Theory, describe board structure but evaluate nothing. ARGO offers a qualitative idea, "interpretive latitude" for local business units, that our escalation threshold turns into a number. Both are inspiration, not evidence.

Research Program

We extend Messiah in three ways. First, demand follows Merton jump diffusion: ordinary random drift plus occasional sudden spikes, as in the agentic study. The rate of spikes, written $\lambda$, is a control variable. Second, when stock is scarce, suppliers allocate it in proportion to orders. This rule creates the classic incentive for buyers to inflate orders during shortages. A parameter $\beta$ scales how much profit a firm gains from inflating, and $\beta = 0$ gives an honest setting. Third, each firm sees only its own state.

Messiah code is not listed as a public resource in the file. If it is unavailable, we will reimplement it from the papers' specifications and confirm that our version reproduces the qualitative ordering of MORL/D above NSGA-II.

The governance structure has three adjustable parts.

  1. Depth $d$, with values 0, 1, or 2. Depth 0 is flat. Depth 1 adds echelon editors. Depth 2 adds the top-level editor. Each tier an order must climb adds one period of decision delay.
  2. Aggregation. Echelon editors pass upward either raw reports or a credibility-weighted aggregate forecast per tier.
  3. Escalation threshold $\tau$. An order is escalated when it deviates from the editor's credibility-weighted forecast by more than $\tau$ standard deviations. Setting $\tau = 0$ escalates everything, and $\tau = \infty$ escalates nothing.

The firms' order policies and the editors' adjustment policies are trained with the MORL/D recipe, so that structure is the only thing that changes across conditions. Credibility is an exponentially weighted record of how accurately each reporter's past forecasts matched realized sales.

We report the bullwhip ratio at each tier $k$,

$$B_k = \mathrm{Var}(q_k) / \mathrm{Var}(D),$$

where $q_k$ is the order stream leaving tier $k$ and $D$ is consumer demand. We also report hypervolume, the share of the profit, emissions, and inequality trade-off space that a method's best solutions cover.

The summary statistic is the fraction of the information-sharing gap that a structure $G$ recovers,

$$R(G) = \frac{H(G) - H_{\text{none}}}{H_{\text{full}} - H_{\text{none}}}.$$

Here $H$ is hypervolume. $H_{\text{none}}$ comes from decentralized agents who share nothing. $H_{\text{full}}$ comes from a centralized learner with full visibility and no delay.

We also log how many raw private fields each structure reveals, so that recovery can be plotted against exposure. The deliverable is a map of $R$ over the grid of $(\lambda, \beta)$ settings, showing which choice of $(d, \tau)$ maximizes it. Our hypothesis is that the best depth falls as $\lambda$ rises and rises with $\beta$.

Baselines

The no-sharing reference. Each firm runs its own learned order policy, or alternatively a classical order-up-to rule tuned on its own history, using only local information. This sets $H_{\text{none}}$. Any structure that does not beat it on bullwhip ratio and hypervolume has failed outright.

The centralized learner with full visibility. This is MORL/D trained on the whole network state. On the Complex Messiah setting, MORL/D reports up to 75% higher hypervolume than NSGA-II, and a Pareto front about eleven times denser than modified PPO. It is the strongest existing method and acts as an oracle, because it sees what real firms would not reveal. We run it on our extended simulator with the same number of environment steps and the same hyperparameter search as every governance variant. We do not expect to beat it. Success means a structure reaches $R$ of 0.8 or more while revealing only aggregates.

Flat information sharing and negotiation. This baseline follows Agentic LLMs in the Supply Chain, which reports a 92.8% cost reduction when Gemini Flash agents share information, against standalone agents in a linear chain. We implement it in two forms. In the first, learned policies receive their neighbors' raw reports. In the second, the language-model protocol is ported to the networks with multiple suppliers and branches. Because it is the most plausible simpler answer, it is the most important competitor. The hierarchy beats it only if it achieves higher $R$ when $\beta > 0$, or equal $R$ with fewer raw fields revealed.

A single central planner that issues all orders. It receives aggregated tier reports but has no intermediate editors, so every decision carries the full delay of a round trip to the top.

Ablations of the key ingredient. The first compares exception-triggered escalation against escalating everything ($\tau = 0$) and escalating nothing ($\tau = \infty$) at fixed depth. The second compares credibility-weighted aggregation against plain averaging. If neither escalation nor credibility weighting matters, the editorial analogy adds nothing beyond aggregation.

Why not just…?

Why not just run a centralized optimizer on all the data? Where the data can be pooled, a centralized optimizer should be used. That is exactly why it serves as our upper bound. Its premise of full visibility is what fragmented, privately owned supply chains violate, as Oh et al. and Kuan et al. both stress. Our contribution is to measure how close a structure that respects privacy can get to it, and our $R$ statistic reports that distance directly.

Why not just let neighbors share information and negotiate, with no hierarchy? This is the real threat. If agents are honest and willing to share, flat sharing may match or beat a hierarchy, because it adds no delay. We test this head-on with the $\beta = 0$ condition, and we will report it plainly if flat sharing wins there. The case for hierarchy rests on two settings. One is $\beta > 0$, where neighbors cannot check one another's reports while an echelon editor with a credibility score can. The other is network-wide carbon caps, which no single bilateral negotiation can enforce. If flat sharing still wins under misreporting incentives, the core claim is wrong.

Why not just use one central planner and skip the middle tier? A single planner pays the full delay on every order and must process every report. The structure we propose escalates only anomalies. The comparison of escalating everything against the triggered rule isolates this. We predict that the central planner's disadvantage grows as the spike rate $\lambda$ rises. If it does not, the middle tier is unnecessary.

Why not just use coordinating contracts such as revenue sharing? These contracts are proven for two firms, but they break down across tiers and under loss aversion (Loss-Averse Supply Chain Coordination; Oh et al.). We still include a revenue-sharing variant of the no-sharing baseline on the 7-node network, where it is most likely to compete.

Decisive Checks

We run 20 seeds per cell on all three Messiah sizes. The grid covers $\lambda$ from rare to frequent spikes and $\beta$ from zero to strong incentives. Picard-style batched simulation keeps runs over long horizons affordable.

The claim survives only if all three of the following hold. First, at $\beta > 0$, triggered two-tier review achieves higher $R$ than flat sharing, with non-overlapping 95% confidence intervals, on the 24-node network. Second, the best depth falls as $\lambda$ rises. Third, triggered escalation beats both escalating everything and escalating nothing.

A secondary replication swaps the learned editors for language-model editors. Each editor's policy is written in plain text, JSON, and YAML, following Policy-as-Prompt. For each format we measure how often the escalation decisions flip. This tells us whether governance by prompt is stable enough to put in the loop at all. No reported result was found on Lacuna for language-model agents on Messiah or on networks with branches, so we will establish those numbers ourselves.