Adapting Supply-Chain Decision Support to the Analyst's Stage of Skill

When an export control or sanction suddenly cuts off a supplier, a human analyst usually decides whether to switch sources, accept lateness, or override the recommendation of an optimization tool. Skill research describes how people move from slow, effortful, conscious problem solving to fast, low-effort automatic execution. Most of that work comes from brain-computer interface training, where users learn to control a device with mental imagery. This proposal argues that a disruption-response tool should change how it presents recommendations and threat alerts depending on where the analyst is on that path. The signal for "where" should be the analyst's spare attention, not their accuracy so far. A novice and an expert can make equally good decisions on a routine shock. The difference is that only the expert has attention left over to notice when the tool's recommendation is wrong on an unfamiliar shock.

The opportunity is real because the two literatures fail in complementary ways. Supply-chain disruption work either removes the human or treats the firm as one rational actor. The second direction's own open problems name this neglect of operator stress and decision latency. The skill literature has a clear stage model but, in its own words, has never turned that model into a working computational tracker tested in large groups, and it has never been tested outside benign tasks. The research file contains no paper joining the two, so the closest work is on one side or the other, and the contrast is drawn below.

Core Claim

We claim that a three-state skill tracker, fit to response times and to performance on a secondary probe task, will produce better outcomes than adapting on accuracy alone. The tracker would mark each analyst as cognitive, associative, or autonomous, and would control both how much the tool explains its sourcing recommendations and how threat-laden its alerts are. The comparison is on held-out geopolitical shocks, measured by lateness and cost, and on how often analysts catch deliberately wrong recommendations.

A second, sharper prediction concerns alert framing. Threat-framed alerts should help analysts in the autonomous state and hurt analysts in the cognitive state. The basis is two findings. Advances in user-training for mental-imagery-based BCI control reports that anxiety diverts attention into off-task worry. Communication strategies driving online health community patient awareness and engagement reports that fear framing raises clicks but suppresses follow-through. Together these imply that the same alert style cannot be best for everyone.

Grounding in Prior Work

The stage model comes from Towards a Cognitive Model of MI-BCI User Training and Advances in user-training for mental-imagery-based BCI control. Both papers map brain-computer interface learning onto Ackerman's skill-acquisition phases. In that model, early performance depends on general reasoning and attention, and later performance depends on task-specific skill. They also find two things this proposal reuses as design rules:

  • Mildly encouraging, biased feedback helps novices but frustrates experts.
  • Feedback that maps transparently onto the task lowers workload.

These papers motivate and supply the model, but their model stays conceptual. Building and validating a computational tracker is part of what this proposal contributes.

Brain computer interface learning for systems based on electrocorticography and intracortical microelectrode arrays supports the claim that the effortful-to-automatic progression is general skill learning and not an artifact of one recording method. Modelling Familiarity for Intelligent Personalized Social Mobilization formalizes the same three stages for task allocation in a game setting. It was tested only where adoption succeeded, which is exactly the survival-bias limitation that a high-pressure crisis setting addresses.

On the supply-chain side, three papers supply the task and the automated competitors.

The geopolitical structure of the shocks is constrained by three findings.

Research Program

Simulator. We will reimplement the cockpit instance as described by Bi et al. and add a geopolitical shock layer with four features:

  • export controls by country;
  • volatility amplification for buyers whose dependence exceeds the 9.2% threshold;
  • partially hidden ownership chains;
  • prospective, ambiguous restriction announcements.

An InvManagement-v0 variant with the same shock layer serves as a second, simpler environment. In both, an automated advisor proposes sourcing actions. The advisor is the sample-average-approximation solver in one environment and the distributional reinforcement-learning policy in the other. On a scheduled minority of trials, the advisor is given a wrong model of the shock. This tests whether analysts catch the error.

Skill tracker. The tracker is a three-state hidden Markov model, meaning a model that infers an unobserved stage from a sequence of observations. Its observations per decision are:

  • decision latency and its variability;
  • accuracy on a secondary auditory probe task run alongside the main task, which is the standard behavioral measure of spare attention;
  • how many information panels the analyst opens.

Accuracy on the main task is deliberately excluded from the tracker's inputs. This keeps it separable from the accuracy-based baseline.

Validating the tracker. The test is whether the inferred states reproduce the Ackerman signature. A general reasoning test should predict sourcing quality in the inferred cognitive state. In the autonomous state, that link should fade and a task-specific practice measure should take over. If the states do not show this shift, they are not skill stages, and we report that.

Adaptive policy. The tool's behavior changes with the inferred state:

  • In the cognitive state, it gives step-by-step explanations of each recommendation, uses calm, affiliative alert framing, and keeps a transparent link between the shock and the proposed action.
  • In the autonomous state, it shows compact raw indicators and threat-framed alerts, and offers no encouraging feedback, because Jeunet et al. found such feedback backfires for experts.

Baselines

Unaided analyst. Analysts use the plain dashboard with no advisor. This shows what human judgment alone achieves in our simulator.

Advisor alone. The advisor executes its recommendations without a human. In the cockpit environment, the reference figure is Bi et al.'s report that under a 100% lead-time shock the risk-averse framework cut total lateness from 8 units to 3, while raising cost from 28,792 to 29,208. We first reproduce these figures on the undisturbed instance to confirm our reimplementation. In the inventory environment, Sabal Bermúdez et al. report a 76.6% decrease in the standard deviation of returns relative to constrained policy optimization, which is the bar our advisor must reproduce before use. No reported result was found on Lacuna for human-in-the-loop performance on either environment, nor for either environment under geopolitical shocks. Comparing the advisor alone against the human conditions on identical shock sequences answers whether the human adds anything at all.

Static advisor with a fixed training schedule. Every analyst gets the same advisor interface and the same number of practice sessions. This is the realistic industrial default.

Accuracy-adaptive advisor. This is the strongest competitor. It changes its display whenever rolling accuracy on the main task crosses fixed thresholds, as a standard intelligent tutor would. It receives the same three display modes, the same total sessions, and thresholds tuned on a pilot cohort with the same tuning budget we give the hidden-Markov tracker.

Ablation of framing adaptation. One condition adapts explanation depth but holds alert framing fixed. This isolates the framing prediction.

What counts as winning. The skill-adaptive condition beats the accuracy-adaptive one if, on held-out shocks in the final session, it shows both of the following, preregistered and with confidence intervals excluding zero:

  • lower lateness at no more than equal cost;
  • a higher rate of correctly overriding wrong advisor recommendations.

Why not just…?

Why not just automate and remove the analyst? This is a serious alternative on routine shocks, and our advisor-alone baseline may well win there. The case against it rests on unfamiliar shocks. Bi et al. concede their solver needs known shock distributions. Liming et al. identify missing error handling as a main adoption barrier. Our injected-error trials measure directly whether analysts recover losses the advisor causes. If analysts do not beat the advisor alone even on those trials, the proposal's premise fails, and we will say so.

Why not just adapt on accuracy? This is the real threat, because accuracy and skill stage are correlated. Our answer is that equal accuracy can hide unequal spare attention, and spare attention is what error-catching requires. The head-to-head comparison settles it. If the accuracy-adaptive tool matches ours on error catching, the skill-stage construct adds nothing operationally, and we will report that as the main result.

Why not just use calm framing for everyone? The cited evidence is all from novices or consumers. It says fear framing hurts follow-through, not that it hurts trained operators. Jeunet et al.'s finding that novice-friendly feedback frustrates experts suggests a single framing is wrong for someone. The framing ablation, crossed with the inferred skill state, tests whether the crossover interaction exists. If calm framing is uniformly best, we drop the framing adaptation.

Why not just train longer? The fixed-schedule baseline holds total practice equal to the other conditions. If extra practice under a static interface catches up by the final session, adaptation only speeds learning. That is still useful, but it is a weaker claim, and we will state it that way.

Decisive Checks

We will recruit procurement professionals and supply-chain graduate students. Each completes six sessions, matching the six-session span of the brain-computer interface study, and is randomized across the five conditions, with sample size set by a pilot power analysis.

The study succeeds only if three things hold together:

  • The inferred states reproduce the Ackerman shift in which abilities predict performance.
  • The skill-adaptive tool beats the accuracy-adaptive tool on both lateness and override quality on held-out ambiguous export-control shocks.
  • Threat framing shows the predicted reversal between the cognitive and autonomous states.

Failing the first check means the tracker is not measuring skill. Failing the second means skill stage does not add operational value beyond accuracy. Failing the third removes the framing claim but leaves the explanation-depth result standing.