The endpoint that deceives

Suppose two systems produce the same response to every item in a test set. That observation supports a narrow claim: under those test conditions, their outputs match. It does not identify the state each system retained, the experience each system selected, the support each system depended upon or the objective that made one outcome preferable.

This hidden assumption appears whenever a learning curve is treated as a biography. A curve plots performance against examples, time or compute. It usually compresses away who chose the examples, what happened after failure, whether assistance adapted, whether the learner changed the task and whether the learned change survived interruption. Those omissions are not decorative context. They may be the mechanism.

Three terms help, provided they are used as instruments rather than labels. Training changes a learner under a largely externally specified objective and experience distribution. Adaptation changes behaviour or state in response to local conditions; the change may be transient or durable. Development, as used in this paper, is durable and path-dependent change in which the learner’s activity helps shape the experiences, support or goals that will later shape the learner.

This last sentence is a practitioner synthesis, not a claim that developmental science has one agreed definition. Its value is operational. It directs attention to a reciprocal loop: learner state affects action; action changes the sampled world and elicits responses from others; those consequences alter later learner state. Biological maturation, sleep, embodiment and culture also matter. The proposed definition does not replace them. It isolates a causal property that can be looked for across very different systems.

A biography is a causal object

Calling the path a biography does not mean adding an anecdotal story after the metrics. It means representing the sequence as part of the causal model. Let L denote the learner state at a moment, N the surrounding niche, S the support currently available and G the operative goal. The learner takes an action A and receives an outcome. The smallest useful description has two coupled updates:

The first update is familiar. A learner changes after an experience. The second is the neglected half. The available environment changes because of what the learner did and because another agent, institution or tool responded. When the second update is independent of the learner’s action, the path is open-loop. When it depends on action, the learner helps construct the distribution from which later learning occurs.

Consider two lessons: first name the parts of a pump, then dismantle it; or first handle the pump, then receive the names. The inventory of information can be identical. The order can still change which distinctions become salient, which errors occur and which questions become askable. In a machine system, tool use before a textual explanation may generate traces that make the explanation actionable. Reversing the order may produce fluent vocabulary with no diagnostic competence. Sequence therefore need not commute: experience A followed by B can produce a different state from B followed by A.

The equations assume that the recorded states contain the relevant past. If a relationship history, bodily condition or unlogged tool effect matters, the state description is incomplete and must expand. That limitation is useful. It turns a vague claim about “developmental context” into a search for missing variables. The equations are not a psychological theory. They are a ledger for asking whether the learner merely receives a path or becomes one of its causes.

The thesis can now be stated more precisely. Development is not training plus elapsed time. It is a history in which the learner helps produce the conditions of its next change. That history can be weak or strong, and humans do not exercise it equally in every setting. A child completing a rigid worksheet may inhabit a less reciprocal loop than a robot selecting experiments through intrinsic motivation. The comparison therefore concerns causal organisation, not species prestige.

flowchart TD
  A["The identical report card"] --> B["The endpoint that deceives"]
  B --> C["The learner and the niche<br/>change together"]
  C --> D["Machine adaptation has several<br/>update surfaces"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#D95545,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

The learner and the niche change together

Play, attachment, curriculum and scaffolding are often listed as ingredients of development. That list hides their distinct jobs. Each changes a different control point in the loop that generates experience.

Each mechanism changes a different control point in the experience-generating loop. Play changes who proposes the intervention. A learner who turns, combines, drops or repurposes an object is more than consuming data. The learner chooses actions that reveal new data. Attachment changes the tolerable risk and the route back after disruption. Curriculum changes the order and density of encounters. Scaffolding changes support in response to the learner, then withdraws it so that control can transfer.

Play is an experiment policy

Experiments with preschool children show that exploratory action is sensitive to evidence and social framing. In Bonawitz and colleagues’ multifunction-toy study, children directly taught one function, or watching it taught to another child, performed fewer unique actions and discovered fewer non-demonstrated functions than children who watched an adult-directed or merely intentional demonstration.[4] This does not prove that instruction is generally harmful. It shows that a teaching context can rationally change what the learner expects remains worth searching.

Play and pedagogy are not opposites

The multifunction-toy result is sometimes flattened into a rule that instruction suppresses creativity. That rule would be another endpoint error. A teacher can reduce a search space by signalling that one action is relevant, but can also ask a question, expose a contradiction or arrange materials so that the learner performs the revealing intervention. The causal issue is how communication changes the learner’s search policy.

The 2025 study cited in Figure 3 sharpens the boundary. When pedagogical signals were present but the adult did not demonstrate the target function, children’s exploration was not necessarily constrained in the same way.[5] A plausible interpretation is that learners infer not just “someone is teaching” but “this demonstrated information is intended to be sufficient or especially relevant for me”. That inference can be efficient when the teacher is knowledgeable and the task is narrow. It can hide undisclosed possibilities when the object has several useful functions.

Developmental design therefore does not maximise unguided variation. Random action can waste time or create danger. Nor does it maximise direct instruction. The design problem is to set constraints that make informative action possible while leaving genuine control with the learner. In machine terms, an exploration policy needs safe action bounds, but it also needs room to choose tests whose outcomes can change its model. A script that samples every permitted action is broad, yet not self-directed. A learner that selects one high-value intervention can be narrow, yet causally active.

Attachment is more than a safety buffer

In attachment research, the secure-base idea joins two functions: a child can explore from a relationship and can return to that relationship as a haven when distressed.[8] It is tempting to map this directly onto a machine’s rollback point, sandbox or recovery controller. That comparison preserves one useful relation: a reliable return path can change how far a system ventures. It fails at the most important human properties: affection, interpersonal history, embodied regulation, grief, trust and felt vulnerability.

Curriculum arranges difficulty; scaffolding transfers control

A curriculum changes the sequence or distribution of tasks. In machine learning, curriculum learning often begins with easier or more revealing examples before harder ones, with a connection to continuation methods that smooth an optimisation problem.[11] In human settings, a curriculum may also encode cultural priorities, norms, explanations and institutional power. The shared structure is ordered exposure. The human meaning of that order is not preserved automatically.

Scaffolding is narrower and more dynamic. A review by van de Pol, Volman and Beishuizen identifies three central characteristics: support is contingent on the learner, it fades, and responsibility transfers to the learner.[7] Static hints are support, but not necessarily scaffolding. Permanent automation may improve output while preventing transfer.

Here, competence is the learner’s current independent capacity, difficulty is the task demand and support lowers effective demand. The equation assumes that learning is strongest when challenge is slightly above competence. It is an illustrative engineering model, not a law of child development. Remove that assumption and the numerical values change. The comparison still exposes the key issue: observed success mixes what the learner can do with what the environment is doing for the learner.

Three ways a scaffold can counterfeit development

First, support may be called contingent when no diagnosis occurs. A tutor supplies a hint after every third error, regardless of why the learner failed. The timing depends on behaviour, but not on an adequate model of learner state. Second, support may fade according to a timetable rather than competence. Withdrawal then measures endurance, not transfer. Third, responsibility may appear to transfer while a hidden service continues performing the hardest inference.

A scaffold audit should therefore record the evidence used to diagnose need, the rule that raises or lowers help and the task component that becomes independently controlled. The clean test is a no-support transfer condition on a changed task. If performance collapses, the earlier score described the coupled learner-plus-scaffold system. That system may still be valuable. The mistake is attributing its full competence to the learner alone.

This distinction matters in both classrooms and software. Autocomplete can increase writing speed while reducing recall of commands. A diagnostic copilot can improve case resolution while leaving the operator unable to recognise a novel failure. Conversely, well-timed examples can reduce avoidable search and accelerate independent skill. The question is not whether help was present. It is whether the control structure was designed to disappear where independence is required.

Machine adaptation has several update surfaces

Asking whether a model “learns during use” is too coarse. A machine system can change at several surfaces, with different persistence and authority. It may update weights, alter a recurrent or context state, write an external episodic record, revise a world model, change a tool policy or receive a new objective from a human. Calling all of these memory or learning hides where the causal change occurs.

Curriculum learning changes which examples arrive first. Intrinsically motivated robotics moves part of task selection into the agent: systems can seek situations where learning progress is high, producing open-ended developmental trajectories.[12] Meta-learning trains an update rule so that a few new examples can produce rapid task adaptation.[13] In-context learning changes expressed behaviour through the current input state, often without a lasting weight change.[15] Test-time training adds an update objective during deployment.[17] Continual-learning methods such as elastic weight consolidation try to protect earlier competence while learning later tasks.[14]

These mechanisms are substantial. None is disqualified because a human engineered the initial objective. Human development also occurs inside inherited biology, designed institutions and culturally supplied goals. The relevant question is more exact: which parts of the experience-generating loop can the machine alter, and under what authority?

Where an update occurs is not why it occurred

A durable weight update can be less developmental than a transient context update. Offline gradient descent may change millions of parameters while the learner has no influence over examples, timing or objective. A short-lived recurrent state may instead encode the result of experiments the agent selected and tools it chose to invoke. Persistence and reciprocity are separate axes.

The same separation applies to external state. Writing a note to a case ledger may preserve experience across sessions, but only if later policy is authorised to read it for the same subject and purpose. A retrieval system can simulate continuity by replaying old text while the underlying learner remains unchanged. That can be an excellent architecture. It should be described as external state-supported adaptation rather than silently credited to model development.

Objective change is even more delicate. A child’s goals can be redirected by hunger, play, approval, norms, reflection and conflict with other people. A deployed machine usually operates under an objective and permission structure that it must not rewrite unilaterally. Bounded negotiation can still exist: the system may propose a subgoal, ask which trade-off matters or revise a plan after human confirmation. Developmental reciprocity does not require unrestricted self-modification. It requires that the allowed change points be explicit.

This yields a two-question audit. First ask where did the update land? Then ask what caused that update opportunity to exist? The first question separates weights, runtime state and external records. The second separates fixed training from learner-shaped experience. Conflating them is why a model with a long context is sometimes called a lifelong learner, and why a continuously fine-tuned service is sometimes assumed to have a biography when its data pipeline is entirely external.

Now inject failure. The simulator changes the pump casing, so the familiar acoustic clue becomes unreliable. A fixed demonstrator keeps showing the old test. A responsive tutor notices the apprentice’s repeated contradiction and withholds the hint, allowing another test. A curiosity objective sends the agent toward observations with high prediction error, but may chase sensor noise. A continual learner preserves old knowledge so aggressively that it cannot update. The mechanisms can fail in opposite directions.

Machine learning research has long argued that human-like learning requires more than pattern recognition, including compositional representations, causal models and learning-to-learn.[18] The present thesis adds a different pressure: even an excellent internal model may not be developmental if it cannot participate in constructing the path by which it changes. Conversely, a modest system can instantiate reciprocal path construction on a narrow task without possessing a human-like world model.

stateDiagram-v2
  [*] --> Claim
  state "A biography is a causal object" as Claim
  state "Play is an experiment policy" as Test
  state "Play and pedagogy are not<br/>opposites" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Compare biographies, not only scores

The practical consequence is a new evaluation design: the Matched-Endpoint Developmental Comparison, or MEDC. It begins by making two learners look equivalent on purpose. Then it applies interventions that expose how the equivalence was produced.

The method does not yield a developmental similarity score. A single number would trade clarity for false precision. Instead, it produces a causal profile whose dimensions can disagree. A system may show strong learner-selected exploration, partial support fading, no objective revision and weak continuity after reset. That profile is more useful than a verdict.

Endpoint matching is a design problem

Exact equality is neither necessary nor usually credible. Human participants vary, stochastic models vary across runs and assistance itself can be difficult to quantify. MEDC uses a declared equivalence band: the candidates must reach outcomes close enough that the planned intervention, rather than an obvious baseline gap, becomes the main explanation for later divergence.

Matching may require different amounts of support. That is not a nuisance to hide. Support is one of the measurements. A novice with a checklist and an expert without one may both complete the task, but the amount, timing and type of assistance reveal different competence. For machine systems, token budget, retrieval access, tool calls and human corrections belong in the same ledger as the final score.

The endpoint must also be matched on more than the benchmark used to tune the system. At minimum, use a held-out task family and several random seeds or repeated human cases. Otherwise, the intervention can merely expose overfitting. The goal is not to make the candidates identical. It is to remove easy explanations so that the learning history becomes experimentally visible.

The MEDC protocol

1. Match the endpoint. Calibrate task difficulty or assistance until candidate systems reach a comparable observable result. Record the support required to achieve it.
2. Record the biography. Version the learner state, objective, experience generator, support policy, social or tool partners and all persistent external state.
3. Perturb one relation at a time. Remove support, swap the partner, change the goal, offer a new tool, interrupt the trajectory, replay the same interaction open-loop or allow the learner to choose the next task.
4. Measure consequences separately. Report independent transfer, exploration breadth, recovery, retention, goal sensitivity and the amount of support that remains necessary.
5. Build rivals. Give an open-loop system the same observations, or give a non-learning system a powerful retrieval store. A developmental mechanism earns credit only when a matched rival cannot generate the same evidence.

Pre-register the revealing intervention

Each intervention should correspond to one claimed mechanism. Support removal tests transfer and dependence. Partner swap tests whether learning depends on a particular interaction history or merely on information content. Open-loop replay tests contingency. Task choice tests whether learner-selected experiments improve adaptation beyond a matched external curriculum. Reset tests persistence and the location of state. Goal perturbation tests whether the system can reinterpret action under changed value or authority.

Specify the expected direction before observing results. A reciprocal learner need not win every metric. Greater exploration can reduce immediate reward. Faster updating can increase forgetting. Strong attachment-related regulation in a child is not expected to look like mechanical fault recovery. A useful profile contains trade-offs, not a heroic green line.

Negative results are informative when the implementation was adequate and the measurement bridge was explicit. If contingent support and open-loop replay produce the same transfer after information and timing are matched, contingency loses explanatory weight for that task. If a model retains adaptation after every relevant state is reset, the state inventory is incomplete. If task choice helps only because it consumes more samples, the result supports resource allocation rather than learner-shaped development.

How to read the lab

The lab tests whether path construction leaves evidence after the visible endpoint has been matched. “Learner initiative” governs how strongly the reciprocal system seeks informative challenge. “Support contingency” governs whether help responds to current competence. “Fading” transfers control by withdrawing help. “Persistence” determines how much competence survives the shock. The baseline keeps a fixed curriculum and fixed support policy.

A positive result is not simply a higher green line. It is a stable pattern in which matched endpoint performance separates after theoretically chosen interventions, and matched open-loop rivals fail to reproduce that separation. A negative result means the proposed reciprocal mechanism did not matter under the tested conditions. It may have been implemented weakly, measured badly or genuinely irrelevant.

The lab cannot establish human-equivalent development. Its variables are engineering abstractions. It cannot establish attachment, felt safety, intrinsic value, consciousness or moral status. It does expose assumptions that prose often hides: support can inflate visible success; exploration can help or distract; persistent updating can accelerate recovery while risking interference.

type DevelopmentalProfile = {
  learnerInitiative: number;      // 0..1, who selects informative actions
  supportContingency: number;     // 0..1, help responds to learner state
  supportFading: number;          // 0..1, help withdraws with competence
  updatePersistence: number;      // 0..1, state retained after interruption
  objectiveRevision: "none" | "bounded" | "negotiated";
};

type MEDCResult = {
  supportedEndpoint: [number, number];
  independentTransfer: [number, number];
  recoveryEpisodes: [number, number];
  explorationBreadth: [number, number];
  breakConditions: string[];
  aggregateScore: never;          // prohibited: dimensions must stay visible
};

function learningGain(competence, challenge, support, initiative) {
  const effectiveGap = challenge - support - competence;
  const productive = Math.exp(-Math.pow((effectiveGap - 0.12) / 0.22, 2));
  return 0.032 * productive * (0.55 + 0.45 * initiative) * (1 - competence);
}

// Run two biographies, calibrate their final support to the same observed
// endpoint, then compare no-support transfer, shock recovery and exploration.
// The complete executable implementation is the page script below.

What each comparison preserves, and where it breaks

Play and intrinsic-motivation systems can preserve learner-selected intervention and uncertainty reduction. The comparison breaks at enjoyment, pretence, social meaning and felt curiosity. Attachment and recovery controllers can preserve the relation between a return path and risk-taking. The comparison breaks at affective bond and embodied distress. Human and machine curricula can preserve ordered exposure. The comparison breaks where curriculum carries normative, cultural or communicative meaning. Social scaffolding and adaptive tutoring can preserve contingency, fading and transfer. The comparison breaks when the tutor has no model of the learner, support never fades, or the system cannot assume independent authority.

This discipline avoids two symmetrical errors. The first declares that machine learning is merely metaphorical because machines are not children. The second treats any scheduled optimisation, reward signal or context update as a machine version of childhood. The right unit of comparison is a named causal relation with a named failure point.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The identical report card
  M->>E: A biography is a causal object
  E-->>M: compare rival explanations
  E->>D: Let history predict what the endpoint cannot
  D-->>Q: revise scope or act