diff --git a/README.md b/README.md
index 0cb47d3..480efc9 100644
--- a/README.md
+++ b/README.md
@@ -874,4 +874,177 @@ where none exists.**
 
 ---
 
+## Appendix A. When the Metric Is the Product
+
+The preceding twelve sections rest on an implicit assumption: that client
+satisfaction is a function of *experienced service quality* — how long
+*their* task took, relative to its size and urgency. If this assumption
+holds, the proof is valid and the unweighted mean is a destructive metric.
+
+But there exists a scenario in which the assumption fails and the entire
+argument collapses.
+
+### A.1 The Self-Referential Metric
+
+Suppose the service provider reports the unweighted mean completion time
+directly to the client — on a dashboard, in an SLA report, on a marketing
+page — and the client's satisfaction is derived primarily from *that number*
+rather than from their individual experience.
+
+Define client satisfaction as:
+
+$$U_{\text{client}} = f\!\left(\bar{C}(\sigma)\right), \quad f' < 0$$
+
+That is: the client sees "Average resolution time: 6.56 hours" and is
+satisfied, without checking whether *their* ticket — the critical email
+outage — took 6.56 hours or 18.75 hours.
+
+Under this model, SPT genuinely maximizes client satisfaction (Theorem 1).
+The service provider's throughput is unchanged (Theorem 6). The business
+outcome improves: same work done, happier client.
+
+**Every theorem in this paper remains mathematically correct. But the
+conclusion inverts.** The metric is no longer a proxy for service quality
+that can be gamed — it *is* the service quality, because the client has
+agreed to evaluate quality by the aggregate number rather than by their
+individual experience.
+
+### A.2 The Economics
+
+This creates a coherent, stable business equilibrium:
+
+| Actor | Behavior | Outcome |
+|-------|----------|---------|
+| Provider | Optimizes unweighted mean (SPT) | Metric improves, no extra work |
+| Client | Reads dashboard, sees low average | Reports satisfaction |
+| Management | Sees satisfied client + good metric | Rewards team |
+
+Throughput is unchanged (Theorem 6), so the same revenue-generating work
+is completed. The only thing that changed is the *order* — and therefore
+the reported number. Real resources were rearranged, no additional value
+was created, but the business metrics all moved in the right direction.
+
+This is *profitable*. The provider extracts satisfaction from the client
+at zero marginal cost, by optimizing a number that the client has accepted
+as a proxy for quality. The client is no worse off *in their own estimation*,
+because they evaluate the aggregate, not their individual experience.
+
+### A.3 The Fragility
+
+This equilibrium is stable only as long as the client never inspects
+their own experience. It breaks the moment any of the following occur:
+
+**1. The client checks their own ticket.**
+
+A CTO whose email server was down for 18.75 hours will not be reassured
+by a dashboard reading "Average resolution: 6.56 hours." The aggregate
+metric and the individual experience diverge maximally for high-priority
+tasks (Theorem 4). The clients most likely to inspect their own experience
+are exactly the ones receiving the worst service.
+
+**2. A competitor offers per-ticket SLAs.**
+
+If an alternative provider guarantees "P1 incidents resolved within 4 hours"
+instead of "average resolution under 7 hours," the aggregate-metric provider
+cannot compete for clients with critical needs — which are typically the
+highest-value clients.
+
+**3. The provider's team internalizes the metric.**
+
+If the team believes the metric reflects real performance (rather than
+consciously gaming it), they lose the ability to recognize when critical
+work is being neglected. The metric becomes an epistemic hazard: it
+tells the team they are performing well, preventing them from seeing that
+they are not.
+
+### A.4 The General Pattern
+
+This is not unique to task scheduling. The structure is:
+
+1. A measurable proxy is established for an unmeasured quality.
+2. The proxy is reported as if it were the quality itself.
+3. The proxy is optimized, improving the reported number.
+4. The underlying quality diverges from the proxy, but no one measures
+   the underlying quality because the proxy exists.
+5. The system is stable until an exogenous shock forces inspection of
+   the underlying quality.
+
+This pattern appears across domains:
+
+| Domain | Proxy metric | Underlying quality | Divergence |
+|--------|-------------|-------------------|------------|
+| IT support | Avg. resolution time | Critical system uptime | Server down for 19 hrs, avg says 6.5 |
+| Education | Standardized test scores | Actual learning | Teaching to the test, understanding declines |
+| Healthcare | Patient throughput | Patient outcomes | Faster discharges, higher readmission rates |
+| Finance | Quarterly earnings | Long-term value creation | Cost-cutting inflates EPS, erodes capability |
+| Software | Velocity (story points) | Deliverable product quality | Point inflation, features half-finished |
+
+In each case, the proxy is optimized, the number improves, and the system
+*functions* — profitably, even — until the moment the underlying quality
+is tested by reality.
+
+### A.5 A Mathematical Note on Equilibrium Stability
+
+Model the system as a game between provider (P) and client (C).
+
+**Information structure:**
+- P observes individual completion times $\{C_i\}$ and chooses schedule $\sigma$
+- C observes only the reported aggregate $\bar{C}(\sigma)$
+
+**Payoffs:**
+- P's payoff increases with C's satisfaction and is independent of schedule
+  (throughput is invariant)
+- C's *reported* satisfaction $U_C = f(\bar{C})$ is maximized by SPT
+- C's *actual* welfare (if they could observe it) depends on individual
+  $C_i$ values, especially for high-priority tasks
+
+This is a **moral hazard** problem. P has private information (the
+distribution of $C_i$) that C cannot observe. P's optimal strategy is to
+minimize the observable signal ($\bar{C}$) regardless of the unobservable
+distribution — which is exactly SPT.
+
+The equilibrium is a **pooling equilibrium**: P's schedule looks identical
+to the client regardless of the underlying priority-weighted performance.
+A provider with PWCT = 10.2 and a provider with PWCT = 10.167 both report
+$\bar{C} = 6.56$ under SPT. The client cannot distinguish between them.
+
+This equilibrium is stable under the standard game-theoretic condition:
+**C has no incentive to deviate** (they have no better information source)
+and **P has no incentive to deviate** (any other schedule worsens $\bar{C}$
+with zero throughput benefit).
+
+It is *unstable* under **information revelation**: if C obtains access to
+individual $C_i$ values (via a customer portal, a competing vendor's
+transparency, or a sufficiently painful incident), the pooling equilibrium
+collapses and C's evaluation shifts to the underlying quality.
+
+### A.6 The Uncomfortable Conclusion
+
+The honest answer to "does optimizing the unweighted mean hurt the
+business?" is: **not necessarily, as long as the client never looks
+behind the number**.
+
+The honest answer to "does it hurt the client?" is: **only when they
+have a problem large enough to notice** — which is precisely when the
+metric's distortion is largest (Theorem 4).
+
+The honest answer to "is this sustainable?" is: it is exactly as
+sustainable as any system in which the seller knows more than the buyer.
+Such systems are historically stable for extended periods and then
+collapse rapidly when the information asymmetry is punctured — by a
+crisis, a competitor, or a regulator.
+
+The mathematical structure is clear: the unweighted mean creates an
+information asymmetry between the metric and the reality. Optimizing
+the metric under this asymmetry is *locally rational* for the provider,
+*locally satisfying* for the uninspecting client, and *globally fragile*
+for the relationship.
+
+Whether one calls this "efficient market behavior" or "a dystopian
+consequence of optimizing legible numbers over illegible reality" is not
+a mathematical question. The math says only this: **the incentive exists,
+the equilibrium is real, and it holds until it doesn't.**
+
+---
+
 *This proof was developed conversationally and formalized on 2026-03-28.*