Where Apologies Stop Working: Distance and the Cost of Repairing Referral Relationships

When a rural doctor sends a patient to an urban specialist who did not need to see one, the relationship between the two practices takes damage. Healthcare systems usually repair that damage cheaply, with an excuse, an apology note, or a "better safe than sorry." Evolutionary game theory, the study of which behavioral strategies spread when agents copy more successful neighbors, has a sharp result about this situation. Forgiveness after an apology sustains cooperation only when the apology costs more than the gain from cheating and less than the penalty for an unrepaired breach. Below that window, opportunists defect, apologize, and defect again. Above it, honest agents who made genuine mistakes give up on the relationship rather than pay. These models assume that every agent faces the same apology cost. In healthcare, the cost of the usual repair act is set by geography: a reinstatement visit, an in-person case review, or a lost clinic day for travel. It is close to free near a hospital and very expensive 50 km away over mountain roads.

This note proposes to show, formally and in simulation calibrated to real spatial and triage data, that a single repair rule cannot keep a geographically spread referral network inside the window where apologies carry information. The prediction is a two-sided failure. Near the hub, cheap apologies let referral dumping persist. In the periphery, honest but error-prone practices are pushed into abandonment. We then design and test repair rules whose cost is indexed to location and to local diagnostic noise, and whose currency cannot be produced by a language model at no cost. If the idea is right, the disparity in specialist access between central and remote practices is partly a mechanism-design failure, and it can be fixed by repricing repair rather than by adding capacity.

Core Claim

Let $x$ index a referring practice's location. Each practice has a private gain from offloading an inappropriate referral, $c(x)$. Each faces a penalty for an unrepaired breach, $\delta(x)$, meaning loss of referral priority or tele-reward. Each pays an effective repair cost $\gamma(x)$. Following the stability condition in When apology is sincere, cooperation evolves, even when mistakes occur frequently, repair-based cooperation is stable at $x$ only if

$$c(x) < \gamma(x) < \delta(x).$$

Under current practice, repair is bundled with travel, so $\gamma(x) = \gamma_0 + C_t(x)$, where $\gamma_0$ is the non-travel part of the repair and $C_t(x)$ is travel friction at $x$. A uniform $\gamma_0$ therefore places central practices below the window and remote practices above it whenever the spread of $C_t$ exceeds the window's width. Two things make the window itself location-dependent. First, the honest error rate $\alpha(x)$ is higher where diagnostics are scarce; Bai et al. document rural under-detection. Second, generative text drives the cost of written apologies toward zero (Undermining Mental Proof).

The new contributions are three. We derive how $\alpha(x)$ shifts the window. We measure what fraction of a real geography falls inside it under realistic policies. We test a repair currency that is non-spatial, cannot be forged with text, and is sized per practice, specifically temporary forfeiture of referral priority scaled to $c(x)$ and $\alpha(x)$.

Relation to Prior Work

The game-theoretic core comes from the commitment-and-apology line. Apology and forgiveness evolve to resolve failures in cooperative agreements shows that fake committers invade when apology is cheap, and that costly apology matters more under high noise. Conditions for the Evolution of Apology and Forgiveness supplies the upper bound, where overly expensive repair drives honest agents to revenge or exit. When agreement-accepting free-riders are a necessary evil shows that asymmetric costs open loopholes. All of these assume homogeneous costs and noise. We supply the method for relaxing that assumption, and we take the asymmetry from geography instead of treating it as a nuisance.

The application substrate is Can telemedicine help integrate the referral-based healthcare system?. That paper models GPs and a central specialist with distance-dependent travel cost, and shows that telemedicine under constrained capacity widens access but overloads the specialist and degrades quality. Its GP errs non-strategically, and the model has no penalty and no repair. We keep its operational structure, meaning specialist service rate, quality cost, and tele-reward. We add strategic dumping and repeated repair.

Wojtowicz and DeDeo constrain the design. If language models make effortful text free, any repair made of words is a cheap apology regardless of location, so our candidate currencies exclude text. Talk, Judge, Cooperate is the main competing mechanism. Its gossip protocol, ALIGN, lets language-model agents sustain cooperation through natural-language reputation, and it treats gossip as free. We test whether free gossip is enough when the behaviors being judged differ systematically by distance.

The spatial literature motivates the heterogeneity. Guo et al. and Park et al. show that travel friction is large, unequal, and tied to income. MamaCare AI reports that 62% of women of reproductive age in Kebbi State live beyond 5 km from a facility. Supply–demand governance of hierarchical healthcare systems documents patients bypassing local clinics, which is a symptom of the trust breakdown we model.

Research Program

The first component is analytical. We extend the finite-population commitment model with Fermi imitation dynamics, so that agents copy better-performing strategies with probability increasing in the payoff gap. The extension places agents on a distance distribution with location-specific $C_t(x)$, $\alpha(x)$, and $c(x)$. Agents imitate within and across distance bands. We derive the per-band window and the conditions under which a single global repair rule can satisfy all bands at once. Our expectation is that no such rule exists once the spread of $C_t$ across bands exceeds $\delta - c$. We also derive how $\alpha(x)$ narrows the window, since higher honest error means more frequent repair payments for honest agents.

The second component is a calibrated agent-based simulation of a referral network built on the operational model of the telemedicine paper. Specialist congestion feeds back into quality cost. Travel friction comes from distance-to-facility distributions reconstructed from the public sources used in MamaCare AI (GRID3 facility locations and WorldPop population grids for Kebbi State). Honest misreferral rates come from published triage accuracy. In Knitza et al., the expert rule reaches 17% specificity at 90% sensitivity on the RHADAR registry, and logistic regression reaches 33%. These figures give realistic bounds on how often an honest referrer sends an unnecessary case. We set $\alpha(x)$ in remote bands toward the expert-rule figure and in central bands toward the machine-learning figure. Sensitivity analysis covers that choice.

The third component replaces scripted GPs with language-model agents in a smaller version of the network. This tests whether, given a free text-apology channel, the models discover the defect-apologize loop, and whether the proposed repair currency removes it. The test reuses the agent-game setup style of ALIGN.

Baselines

The naive reference point is unconditional acceptance. Every referral is accepted, there is no penalty, and repair is a free written note, which is the cheap-apology regime. It sets the floor for appropriate-referral rate and the ceiling for specialist overload.

The status-quo baseline is travel-bundled repair, $\gamma(x) = \gamma_0 + C_t(x)$, with $\gamma_0$ tuned to the best single value. It represents how reinstatement actually works and is the configuration the core claim says must fail at both ends.

The strongest homogeneous competitor is the costly-apology strategy of Han et al. with one $\gamma$ chosen inside the population-average window. Its repair cost is decoupled from travel but uniform across practices. For this class of model, the reported reference is a cooperation level of about 0.6 under noise $\alpha = 0.1$ in the play-play scenario of Apology and forgiveness evolve to resolve failures in cooperative agreements. That result is for a homogeneous population, and we will reproduce it before adding geography.

The grim-trigger baseline permanently revokes referral rights after a breach. It is included because administrators propose it, and the cited work predicts it collapses under noise.

The reputation competitor is ALIGN-style gossip among practices and the specialist, with free natural-language reputation and no costly repair. ALIGN reports 100% cooperation for reasoning models in its infinite-horizon donation game in Talk, Judge, Cooperate, so it is the strongest existing mechanism. No reported result on Lacuna exists for it under heterogeneous noise or in referral settings.

The triage-filter baseline places a referral classifier at the specialist's door, with the operating point of the logistic regression model in Knitza et al.

All mechanisms share the same geography, error rates, specialist capacity, and simulation seeds. Each mechanism's free parameters get the same tuning budget, a fixed grid search over the same number of configurations, optimized for the same objective. Two ablations isolate the key ingredient: indexing repair cost on distance alone, and on local noise alone. The proposal beats the field if it raises the appropriate-referral rate in the most remote quintile without lowering it in the central quintile, relative to the best tuned uniform rule. The gap must exceed seed-to-seed variation, and specialist quality cost must be no worse.

Why not just…?

Why not just subsidize travel or allow free telemedicine reinstatement for remote practices? That makes $\gamma(x)$ small everywhere, which moves remote practices from above the window to below it and reproduces the cheap-apology loop the theory predicts. The telemedicine model already shows that lowering referral barriers without volume discipline overloads the specialist (Can telemedicine help integrate...). The free-reinstatement condition is a named arm of the simulation. We expect dumping rates in remote bands to rise to the unconditional-acceptance level.

Why not just put a better triage classifier at the specialist's door? This is a real threat. If a filter rejects most dumped referrals, dumping stops paying and repair becomes irrelevant. The available evidence suggests the filter is not tight enough. At the clinically required 90% sensitivity, the best model in Knitza et al. reaches only 33% specificity, so most unnecessary referrals still pass. The model also scores self-reported forms, which cost nothing to exaggerate. We run the filter alone and combined with our repair rule. If the filter alone matches the combination across all bands, we will report that pricing repair adds nothing.

Why not just tune one uniform repair cost carefully? The analysis gives a concrete test. We compute the fraction of the Kebbi population whose band lies inside the window at the best uniform cost. If that fraction is near one, geography does not matter, and the claim fails.

Why not just use gossip-based reputation? Gossip is free to produce and can be generated by language models, so it is exactly the kind of signal Wojtowicz and DeDeo argue loses meaning. In a heterogeneous network it also risks reading honest remote errors as misconduct. The ALIGN arm, and its combination with our repair rule, will show whether reputation alone separates noisy honest practices from strategic dumpers by band.

Proposed Experiments

The first check reproduces the homogeneous result of roughly 0.6 cooperation at $\alpha = 0.1$. We then introduce Kebbi-calibrated travel friction under travel-bundled repair. The decisive pattern is high dumping in the central quintile and high abandonment, meaning appropriate referrals withheld, in the remote quintile, with the gap between quintiles growing with the spread of $C_t$.

The second check compares all baselines on four metrics: appropriate-referral rate by distance quintile, the remote-to-central ratio of access for genuinely sick patients, specialist quality cost, and the fraction of practices inside their local window. The distance-only and noise-only ablations determine which indexing variable carries the effect.

The third check uses language-model agents. They receive a free text-apology channel and are compared under three repair conditions: text apology, travel-bundled repair, and indexed priority forfeiture. We measure defect-apologize cycles per agent.

We also run a robustness sweep that replaces Kebbi distances with a flatter distance profile. The theory predicts the advantage of indexing shrinks as the spread of $C_t$ shrinks. That prediction can be falsified, and it is the cleanest test that geography, rather than tuning, drives the result.