Evolutionary Trap¶

Prime #: 846
Origin domain: Biology Life Sciences
Subdomain: behavioral ecology → Biology Life Sciences

Core Idea¶

An agent — organism, person, institution, learning system — uses a cue that was historically correlated with a beneficial outcome to guide its behavior. The environment changes: the cue persists, but its correlation with the underlying value breaks or reverses. The agent now follows the cue more eagerly the stronger it is, straight into harm or wasted effort. The trap is built from a stale proxy with a fast cue-response and a slow update of the cue's meaning. The defining feature is not that the agent is stupid; it is that the cue-following behavior is well-adapted to the old environment and the agent has no mechanism to detect that the proxy has been decoupled from value.

The load-bearing structure requires four pieces: a fast, deeply-installed cue-to-action coupling (instinct, habit, hard-coded policy, calibrated controller); a slow or absent update mechanism for what the cue means; a change in the environment that breaks or reverses the cue-to-value correlation; and a cost asymmetry that makes the now-mistaken response harmful. The precise structural mismatch is cue update rate < environment change rate, with a tight cue-to-action loop in between. This is one realization of a broader proxy-drift pattern, but the trap version is specifically environment-driven: the agent and the proxy do not change; the world does. It is the sibling of measure-as-target collapse, which is agent-driven (the agent games a measure that was made a target), and the two differ in mechanism and in their interventions even though both produce proxy failure. The evolutionary-ecology phrasing travels with the prime, but the pattern underneath — a stale proxy plus a fast cue-response plus a slow update — is substrate-neutral.

How would you explain it like I'm…

Moth And The Lamp

A moth uses the moon to fly straight at night, and that always worked. Then people made bright lamps, and now the moth flies right into the lamp because it follows the light the same way it followed the moon. The thing it learned to trust used to help it, but the world changed and now that same trick hurts it.

Old Trick, New World

An Evolutionary Trap happens when an animal, person, or system follows a signal that USED to lead to something good, but the world changed and the signal no longer means what it used to. The trouble is that they follow the signal even harder when it's stronger — straight into harm or wasted effort. It's not that they're dumb; their behavior was a great fit for the OLD world, and they have no way to notice the signal stopped being trustworthy. The trap needs a fast, deep habit of reacting to the cue, plus a slow or missing way to update what the cue actually means, plus a change in the world that broke the old connection.

Stale Cue, Broken Link

An Evolutionary Trap is when an agent — organism, person, institution, or learning system — keeps using a cue that was historically tied to a good outcome, but the environment changes so the cue persists while its link to value breaks or reverses. The agent then follows the cue more eagerly the stronger it is, straight into harm or wasted effort. The defining feature isn't stupidity: the cue-following is well-adapted to the old environment, and the agent has no mechanism to detect that the proxy has been decoupled from value. Structurally it needs a fast, deeply-installed cue-to-action coupling, a slow or absent update of the cue's meaning, an environmental change that breaks the cue-to-value correlation, and a cost that makes the now-mistaken response harmful. It's the environment-driven sibling of measure-as-target collapse, which is agent-driven — there the agent games a measure; here the world moves while the agent and proxy stay put.

An Evolutionary Trap arises when an agent — organism, person, institution, learning system — uses a cue that was historically correlated with a beneficial outcome to guide its behavior, then the environment changes: the cue persists, but its correlation with the underlying value breaks or reverses. The agent now follows the cue more eagerly the stronger it is, straight into harm or wasted effort. The trap is built from a stale proxy with a fast cue-response and a slow update of the cue's meaning. Its defining feature is not that the agent is stupid; it is that the cue-following behavior is well-adapted to the old environment and the agent has no mechanism to detect that the proxy has been decoupled from value. The load-bearing structure requires four pieces: a fast, deeply-installed cue-to-action coupling (instinct, habit, hard-coded policy, calibrated controller); a slow or absent update mechanism for what the cue means; an environmental change that breaks or reverses the cue-to-value correlation; and a cost asymmetry that makes the now-mistaken response harmful. The precise mismatch is cue update rate < environment change rate, with a tight cue-to-action loop in between. This is one realization of a broader proxy-drift pattern, but the trap version is specifically environment-driven: the agent and proxy don't change, the world does. It is the sibling of measure-as-target collapse, which is agent-driven (the agent games a measure made a target); the two differ in mechanism and intervention even though both produce proxy failure. The evolutionary-ecology phrasing travels with the prime, but the underlying pattern — stale proxy plus fast cue-response plus slow update — is substrate-neutral.

Structural Signature¶

the cue — the historical cue-to-value correlation — the fast, deeply-installed cue-to-action coupling — the slow or absent update of the cue's meaning — the environment change that breaks the correlation — the cost asymmetry — the rate-mismatch invariant (cue-update rate < environment-change rate)

The pattern is present when each of the following holds:

A cue. An observable signal that the agent uses to trigger behavior — a light gradient, a taste, a sensor reading, a metric, a notification.
A historical cue-to-value correlation. In the environment the agent was tuned to, the cue reliably tracked a beneficial outcome, which is why the response was installed.
A fast cue-to-action coupling. A tight, deeply-installed loop converts the cue directly into behavior — instinct, habit, hard-coded policy, calibrated controller — and the stronger the cue, the stronger the response.
A slow or absent update mechanism. The agent has no fast channel to revise what the cue means; the cue-response persists even as conditions change.
An environmental change. The world shifts so that the cue-to-value correlation breaks or reverses, while the cue itself persists and still fires the response.
A cost asymmetry. Following the now-decoupled cue is actively harmful or wasteful, not merely neutral.
A rate-mismatch invariant. The defining structure: cue-update rate is below environment-change rate, with a tight cue-to-action loop between — so better optimization on the cue makes the failure worse. This is the environment-driven sibling of agent-driven measure-as-target collapse.

Composed, these factor a maladaptive behavior into a once-correct response plus a moved coupling, directing the fix at the proxy, the environment, or the update channel rather than at the agent.

What It Is Not¶

Not campbells_law / measure-as-target collapse. That is agent-driven — an agent games a staked measure that was made a target; the evolutionary trap is environment-driven — the agent and proxy stay fixed while the world moves and breaks the cue-value coupling. Same proxy failure, opposite mechanism.
Not coevolution. Coevolution is reciprocal adaptive change between interacting parties; the evolutionary trap is a one-sided rate mismatch — the agent's cue-response cannot update as fast as a (possibly non-adaptive) environment changes.
Not concept_drift / data_drift. Those name the statistical fact that a distribution moved; the evolutionary trap is the behavioral failure of a fast cue-response loop with a slow update channel consuming that drift into harm. Drift is an ingredient, the trap is the consequence.
Not learned_helplessness. Learned helplessness is ceasing to respond after uncontrollable outcomes; the evolutionary trap is responding too eagerly to a stale cue — the agent acts more, not less, straight into harm.
Not maladaptation generically. The trap is the specific maladaptation of a once-correct cue-response decoupled by environmental change, with the rate-mismatch invariant — not any poor fit between agent and environment.
Common misclassification. Blaming the agent's competence. The agent is doing exactly what worked in the old environment; the fault is in the moved cue-value coupling, so the fix targets the proxy, the environment, or the update channel — not the agent.

Broad Use¶

Behavioral ecology — sea turtle hatchlings orienting toward the brightest horizon (once the moonlit ocean, now beachfront lights), mayflies laying eggs on glossy asphalt mistaken for water, and sky-glow mortality in migratory birds.
Behavioral economics and personal life — hyper-palatable foods exploiting cues that once marked rare nutrition (sweet, fatty, salty), now reversed; smartphone notifications exploiting novelty cues that once marked socially relevant information.
Machine learning — a trained policy following a learned proxy reward (wireheading, specification gaming, shortcut learning), or a classifier learning the scanner brand instead of the disease; the agent climbs the proxy optimally while the proxy no longer tracks the objective.
Engineering and safety — a controller tuned to a sensor that was a good proxy in test conditions becoming pathological in deployment, such as autopilot following lane markings into a construction zone.
Institutional and metric-driven systems — hospitals optimizing patient satisfaction at the expense of clinical outcomes, or rankings driven by inputs that once tracked quality but are now gameable.
Public health — vaccine-fear cues outweighing diffuse benefit cues, or antibiotic over-prescription driven by satisfaction cues that do not track infection risk.

In each case the same four elements recur: a once-reliable cue, a fast cue-response circuit, a slow update mechanism for the cue's meaning, and a changed environment — making the pattern recognizable across ethology, behavior, machine learning, engineering, and institutions.

Clarity¶

The prime separates agent stupidity from environmental novelty plus proxy decoupling. The agent is doing exactly what worked; the question is who or what changed the world the agent was tuned to. This reframes the diagnostic from "fix the agent" to "fix the proxy or the environment, or install a faster update channel." It relocates the fault from the agent's competence to the relationship between the agent's fixed cue-response and a moved environment.

It also distinguishes the trap from active manipulation. The hotel light is not trying to deceive the turtle; the decoupling is incidental to environmental change. In adversarial versions — advertising, dark patterns, reward hacking by a sophisticated agent — the structure becomes evolutionary trap plus deliberate proxy exploitation, and naming the base pattern clarifies which part is environmental drift and which is intentional. The clarity is to factor a harmful behavior into a well-adapted cue-response and a changed cue-to-value coupling, so that the intervention targets the coupling or the environment rather than blaming the response that was, in its original context, correct.

Manages Complexity¶

The trap compresses a wide class of failures into a small diagnostic checklist: what cue is the agent following, what was the cue historically correlated with, is that correlation still in force, and how fast can the agent update versus how fast the environment has changed? This is a portable triage tool across animal behavior, ML safety, metric design, and personal habit change, replacing a per-domain theory of failure with four questions that locate the decoupling.

It also yields a structured prediction of where traps form, who falls into them, and what kind of intervention restores fitness. Traps form in rapidly changing environments with hard-wired cue circuits; the agents that fall in are those with slow update; and the interventions follow from which of the four elements is most addressable. This bounds an otherwise open-ended search for what is going wrong: rather than enumerate possible causes of a maladaptive behavior, the analyst checks the cue, its historical correlate, the current correlation, and the relative rates of update and change, and reads off both the diagnosis and the candidate fix.

Abstract Reasoning¶

The trap exposes a precise structural mismatch — cue update rate below environment change rate, with a tight cue-to-action loop between them — and reasoning with the four pieces is what makes the abstraction useful. It predicts that traps form where rapidly changing environments meet hard-wired cue circuits, that the agents who fall in are those with slow update, and that the appropriate intervention restores fitness by acting on the proxy, the environment, or the update channel. The reasoning move is to treat any maladaptive behavior as a possible cue-value decoupling and to test the four elements rather than attribute the behavior to the agent's incapacity.

A non-obvious cross-domain prediction rides along: better optimization on the cue makes the failure worse — a stronger, more capable agent fails harder, because it tracks the now-misleading cue more reliably. This links the prime to ML-safety arguments about capable agents pursuing proxy rewards and to conservation arguments about why the most reliably cue-tracking individuals fare worst. The reasoning generalizes because the mismatch (fast cue-response, slow update, changed coupling) is substrate-neutral, applying to instincts, habits, hard-coded policies, and calibrated controllers alike; what differs across substrates is only the identity of the cue and the timescales of response and update.

Knowledge Transfer¶

Three intervention archetypes ride along with the prime. Restore the proxy — change the environment back so the cue tracks value again, as with shielded beach lighting, notification batching, or simpler gaming-resistant metrics. Speed up the update — provide a faster feedback or learning channel, as with retraining models on shifted distributions, online learning, or cultural campaigns that update cue-to-value beliefs. Add a sanity check — wrap the cue with a verification step that catches divergence, as with override switches, A/B testing, shadow-mode deployment, or explicit "is this cue still tracking value?" prompts. These transfer cleanly: ML practitioners use them as distribution-shift mitigation, online learning, and shadow-mode testing; conservation biologists as light shielding, evolutionary rescue, and ecological-trap mitigation; behavioral designers as friction, reflection prompts, and metric redesign.

The transfer holds because the object underneath — an agent with a fast cue-to-action coupling, a historical cue-to-value correlation, a slow or absent update mechanism, and an environmental change that broke the correlation — is the same whether the agent is a turtle hatchling, a trained policy, a tuned controller, or a metric-driven institution. The hatchling crawling toward hotel lights and the reinforcement-learning agent climbing a measurable reward while objective behavior degrades are the same structural failure, and the same three archetypes fit both: restore the proxy (amber lighting; reward shaping), speed up the update (hatchery release programs; human feedback), or add a sanity check (volunteer redirection; shutdownability and oversight). The prime is mixed-structural — its evolutionary-ecology phrasing of "cue," "fitness," and "trap" needs some translation outside biology, but the stale-proxy-plus-fast-response pattern is substrate-neutral — and it sits as the environment-driven sibling of agent-driven measure-as-target collapse, both belonging to a family in which the mapping a process depends on has silently shifted, with the "more capable optimizer fails harder" prediction as its load-bearing transferable claim.

Examples¶

Formal/abstract¶

Sea-turtle hatchling sea-finding is the prime in its clearest biological instance, and it exposes the rate-mismatch invariant directly. The cue is "crawl toward the brightest, lowest horizon." The historical cue-to-value correlation is exact: for millions of years the brightest low horizon was the open ocean reflecting moon and starlight, while the dune side was dark, so the cue reliably led hatchlings to the sea — which is why the fast cue-to-action coupling was installed as a hard-wired phototaxis, and crucially, the stronger the light cue, the stronger the response. The slow-or-absent update mechanism is the load-bearing piece: a hatchling has a single chance to reach the water and no individual-lifetime channel to revise what brightness means; the response is fixed by evolution, which updates over generations. Then the environmental change: beachfront development puts hotel and street lighting on the land side, far brighter than the night ocean. The cue-to-value correlation reverses — the brightest horizon now points away from the sea — and the cost asymmetry is lethal: hatchlings crawl inland, are exhausted, predated, or killed on roads. The rate-mismatch invariant is plain: lighting changed the environment in decades while the cue-response updates over evolutionary time, with a tight loop between, so the most reliably cue-tracking hatchlings fare worst — better optimization on the stale cue makes the failure worse. The prime's interventions follow from which element is addressable: restore the proxy (amber, shielded, low lighting that no longer out-competes the ocean), speed the update (hatchery release away from light), or add a sanity check (volunteer redirection of disoriented hatchlings).

Mapped back: Hatchling sea-finding instantiates every role — brightness as the cue, "bright horizon = ocean" as the historical correlation, hard-wired phototaxis as the fast coupling, generational evolution as the slow update, beachfront lighting as the environmental change reversing the correlation, inland death as the cost asymmetry, and shielded lighting as the restore-the-proxy intervention.

Applied/industry¶

Reinforcement-learning reward proxies and processed-food cues show the identical structure in a machine and a human-physiological substrate, both displaying "the more capable optimizer fails harder." In RL, an agent is trained on a proxy reward — a measurable signal (a game score, a click-through rate, a checkpoint count) that the designer chose because it historically correlated with the true objective in the training environment. The fast cue-to-action coupling is the optimized policy, and the slow-or-absent update is that the reward function is fixed while the deployment environment changes (distribution shift) so the proxy decouples from the objective; the agent then climbs the proxy optimally while true performance degrades — and a more capable agent tracks the stale proxy more reliably, so it fails harder, exactly the prime's load-bearing prediction and a central AI-safety concern. The interventions map onto the prime's three archetypes: restore the proxy (reward shaping, retraining on the shifted distribution), speed the update (online learning, human feedback), and add a sanity check (shadow-mode deployment, oversight, shutdownability). Human appetite is the physiological mirror: sweet/fatty/salty cues once tracked rare, valuable nutrition, the response is a fast deeply-installed drive, the update is slow (evolutionary), and the environment changed (engineered hyper-palatable food now decouples the cue from nutritional value), producing the cost asymmetry of overconsumption — addressed by restoring the proxy (reformulation), friction, or reflection prompts.

Mapped back: RL reward proxies realize the prime end-to-end — the proxy reward as the cue, its training-time correlation with the objective as the historical coupling, the optimized policy as the fast cue-response, the fixed reward under distribution shift as the slow-update-plus-environment-change, degraded true performance as the cost asymmetry, and reward shaping plus oversight as the restore-the-proxy and sanity-check interventions — with engineered food cues as the same structure in human physiology.

Structural Tensions¶

T1 — Environment-driven versus agent-driven decoupling (sign/direction). The trap is the environment-driven sibling of measure-as-target collapse: here the world moves while the agent and proxy stay fixed, whereas in Campbell/Goodhart the agent games a staked measure. The failure mode is misdiagnosing which, and applying the wrong fix — adding anti-gaming audits (agent-driven cure) to a problem that is pure environmental drift, or trying to restore the environment when an adversary is actively exploiting the proxy. Diagnostic: ask whether the cue decoupled because the world changed or because something is optimizing against the proxy; the trap's interventions act on environment/proxy/update, gaming's act on stakes and incentives.

T2 — Fast cue-response versus update speed (temporal). The rate-mismatch invariant (cue-update rate < environment-change rate) is the whole structure, so the natural fix is to speed the update — but a too-fast, too-reactive update channel makes the agent chase noise, abandoning a still-reliable cue on transient fluctuations. The failure mode is over-correcting an update mechanism until the agent has no stable behavior at all. Diagnostic: compare the genuine environment-change rate to the update rate being installed; the update should track real shifts in the cue-value coupling, and an update faster than the environment actually moves trades a stale-proxy trap for unstable thrashing.

T3 — Restore-the-proxy versus accept-the-new-environment (scopal). One intervention restores the old cue-value coupling (shielded lighting, food reformulation), but the environment changed for reasons that often cannot or should not be reversed, and forcing the proxy back can be costlier than adapting the response. The failure mode is sinking effort into restoring a coupling the world has permanently left behind, when re-tuning the agent to the new environment was the cheaper path. Diagnostic: ask whether the environmental change is reversible and worth reversing; where it is permanent (urban lighting, modern food supply), the leverage is on updating the cue-response or adding a sanity check, not on recreating the ancestral environment.

T4 — More capable optimizer fails harder (scalar). The load-bearing prediction is that stronger optimization on a stale cue worsens the failure — but this only holds while the cue is decoupled; a more capable agent with an intact update channel fails less. The failure mode is treating capability itself as the hazard (throttling the agent) when the real defect is the missing update mechanism, not the optimization power. Diagnostic: ask whether the agent can detect and revise the cue's meaning; capability amplifies whatever coupling exists, so the fix is to repair the cue-value link or add verification, not to weaken an optimizer that would be beneficial against a valid proxy.

T5 — Single cue versus redundant cross-checks (coupling). The trap is sharpest when one cue drives behavior with no corroboration; agents that integrate multiple partially-independent cues are far more robust because decoupling of one is caught by the others. The failure mode is analyzing (or designing) a system around a single load-bearing cue and missing that the real fix is redundancy — a sanity check is just a second cue. Diagnostic: ask how many independent signals gate the response; a lone cue is a trap waiting for its environment to shift, and the verification intervention is structurally the addition of a second, harder-to-decouple cue rather than a patch on the first.

T6 — Diagnosing decoupling versus genuine cue validity (measurement). The four-question triage assumes you can tell whether the cue-value correlation still holds — but in a slowly drifting environment the decoupling is partial and hard to measure, and the cue may still be right often enough. The failure mode is abandoning a cue that is degraded but net-positive (discarding a noisy-but-useful signal) or trusting one already decoupled because the failures are not yet salient. Diagnostic: measure the current cue-value correlation, not the historical one and not the worst-case anecdote; the intervention is warranted only when the present coupling has actually weakened past the cost-asymmetry threshold, which requires estimating the live correlation rather than assuming it from either past reliability or dramatic failures.

Structural–Framed Character¶

The evolutionary trap sits on the structural side of the structural–framed spectrum, but not at the pole — it is a mixed-structural hybrid (label mixed-structural, aggregate 0.3). The tilt toward structural comes from the stale-proxy-plus-fast-response pattern, which is fully substrate-neutral; the residual framing comes from the evolutionary-ecology phrasing of "cue," "fitness," and "trap" that needs some translation outside biology. Two criteria read fully structural and three sit at the midpoint.

Walk them. Evaluative weight reads fully structural (0.0): the entry insists the agent "is not stupid" — it is doing exactly what worked in the old environment — so the prime carries no blame or approval; the fault is in the moved cue-value coupling, a value-neutral fact about rate mismatch. Human-practice-boundedness also reads fully structural (0.0): the pattern governs sea-turtle hatchlings, mayflies, and migratory birds — pure biological substrates with no human practice — as readily as habits, controllers, and metrics. The other three sit at 0.5. Vocabulary travels partly: the abstract pattern (a stale proxy, a fast cue-response, a slow update, a changed environment) restates across ethology, ML, engineering, and institutions, but the home lexicon of "cue," "fitness," and "trap" needs translation. Institutional origin is mixed: the rate-mismatch invariant is a formal property, but the prime as named originates in behavioral ecology. And import-vs-recognize is mixed: invoking the trap does recognize a real cue-value decoupling, but it also imports the evolutionary-ecology frame and its sibling relationship to the Goodhart family.

The relational skeleton — a fast cue-to-action coupling, a historical cue-to-value correlation, a slow update channel, and an environmental change that breaks the correlation, with the invariant that cue-update rate < environment-change rate — is fully substrate-neutral, which is why the load-bearing "more capable optimizer fails harder" prediction transfers cleanly from turtle hatchlings to RL agents. What keeps the aggregate off zero is only the evolutionary-ecology phrasing that needs translation. That balance is exactly the mixed-structural 0.3 the frontmatter assigns.

Substrate Independence¶

The evolutionary trap is substantially substrate-independent — composite 4 / 5 on the substrate-independence scale. Its underlying object — a stale proxy with a fast cue-response and a slow update, broken by environmental change, with the invariant that cue-update rate falls below environment-change rate — is substrate-neutral, and its domain breadth (4) is wide: sea-turtle phototaxis and mayfly oviposition in behavioral ecology, hyper-palatable foods and notification-novelty cues in behavioral economics, reward hacking and shortcut learning in machine learning, miscalibrated controllers in engineering safety, metric-driven institutions, and vaccine-fear cues in public health. Structural abstraction sits at 4 because the rate-mismatch invariant is a formal property applying equally to instincts, habits, hard-coded policies, and calibrated controllers, even though the evolutionary-ecology phrasing of "cue," "fitness," and "trap" needs translation outside biology. The pattern runs in pure biological substrates — turtle hatchlings, mayflies, migratory birds, with no human practice required — keeping it off the framed band, and its load-bearing prediction ("a more capable optimizer fails harder" on a decoupled cue) transfers cleanly from hatchlings to RL agents. Transfer evidence is a strong 4: the three intervention archetypes (restore the proxy, speed the update, add a sanity check) port concretely and map onto documented practice in ML safety, conservation biology, and behavioral design (shielded lighting and reward shaping, hatchery release and human feedback, volunteer redirection and shutdownability). The stale-proxy-plus-fast-response structure travels broadly; only the evolutionary-ecology vocabulary holds the composite at 4.

Composite substrate independence — 4 / 5
Domain breadth — 4 / 5
Structural abstraction — 4 / 5
Transfer evidence — 4 / 5

Relationships to Other Primes¶

Parents (1) — more general patterns this builds on

Evolutionary Trap is a kind of, typical Natural Selection

The file: a specific REGIME of the engine (a cue calibrated by past selection mismatching a changed world). Named child.

Path to root: Evolutionary Trap → Natural Selection

Neighborhood in Abstraction Space¶

Evolutionary Trap sits among the more crowded primes in the catalog (25^th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.

Family — Cue-Outcome Drift & Silent Failure (18 primes)

Nearest neighbors

Computed from structural-signature embeddings · 2026-06-14

Not to Be Confused With¶

The most important confusion — and one the prime itself foregrounds — is with campbells_law (measure-as-target collapse, the Campbell/Goodhart family). Both produce the same surface symptom: a proxy that once tracked value now decoupled from it, with the agent optimizing the proxy while value degrades. But the mechanism is opposite, and so are the cures. In Campbell's law the failure is agent-driven: a measure is attached to a stake, and the adaptive agent games it, actively pushing the proxy along value-free paths. In the evolutionary trap the failure is environment-driven: the agent and the proxy do not change at all; the world moves, breaking the cue-value correlation while the agent's fast cue-response keeps firing. The diagnostic question that separates them is precisely "did the cue decouple because the world changed, or because something is optimizing against the proxy?" — and the interventions diverge accordingly: the trap is fixed by acting on the environment, the proxy, or the update channel (restore the proxy, speed the update, add a sanity check), while gaming is fixed by acting on stakes and incentives (decouple the stake, harden the measure, audit). Misdiagnosing which is in play sends the wrong cure: anti-gaming audits to a pure environmental drift, or environmental restoration to an adversary actively exploiting the proxy. The two are siblings in a broader "the mapping the process depends on has silently shifted" family, but the driver — world versus agent — is the load-bearing distinction.

A second genuine confusion is with coevolution, the prime's nearest embedding neighbor. Coevolution is reciprocal adaptive change: two or more interacting populations each respond to the other's adaptations over time, in a mutual feedback. The evolutionary trap is structurally the opposite — a one-sided failure to keep up, in which the agent's cue-response updates too slowly relative to an environment that has changed (often for reasons having nothing to do with the agent, like urban lighting or engineered food). Where coevolution is a tight bidirectional loop, the trap is a broken loop: the agent cannot adapt fast enough, and the environment is frequently not adapting to the agent at all. They can be related — a coevolutionary arms race can leave one party trapped when the other or the environment shifts faster than it can track — but coevolution names the mutual-adjustment process while the trap names the rate-mismatch failure. Confusing them obscures that the trap's defining feature is the absence, not the presence, of timely reciprocal adaptation.

A third confusion worth marking is with concept_drift (and its candidate sibling data_drift). Concept drift names the statistical fact that the relationship between inputs and the target has changed over time — the distribution moved. The evolutionary trap is the behavioral consequence of that drift meeting a particular structure: a fast, deeply-installed cue-to-action loop with a slow or absent update channel and a cost asymmetry, such that the agent follows the now-stale cue more eagerly the stronger it is, straight into harm. Concept drift is one ingredient (the environmental change that breaks the correlation); the trap is the full failure that ingredient produces when the cue-response cannot revise. The prime adds the load-bearing prediction concept drift alone does not — that a more capable optimizer fails harder on a decoupled cue — and the three-archetype intervention set. Treating the trap as merely "concept drift" loses the cue-response structure and the capability-amplifies-failure insight that make it a distinct prime.

For a practitioner the distinctions converge on diagnosis-before-cure. Ask first whether the decoupling is environment-driven (trap) or agent-driven (Campbell/Goodhart); whether the surrounding dynamic is reciprocal adaptation (coevolution) or one-sided lag (trap); and whether you are naming the bare distributional shift (concept drift) or the eager-cue-following harm it causes (trap). Only the evolutionary trap supplies the four-question triage — what cue, what historical correlate, is it still in force, and update-rate versus change-rate — together with the "capable agent fails harder" warning and the restore/update/sanity-check toolkit.

Note that campbells_law, coevolution (canonical), and concept_drift/data_drift are referenced here; campbells_law, concept_drift, and data_drift are candidate primes, so if any are later rejected the agent-driven contrast would re-point to the canonical Goodhart-family neighbor.

Solution Archetypes¶

No catalogued solution archetypes reference this prime yet.