Map Users and Mapmakers: The Scope of Cognitive Attribution from Acquired Representations

• An acquired representation can support recognition, route comparison, and novel connections without establishing that its user can investigate an unfamiliar environment.The paper treats these as substantial achievements while distinguishing them from the capacity to conduct comparable inquiries.Source: §1 · p. 1
• The framework addresses Ng’s qualified world-model claim about Othello-GPT and indicators of conceptual understanding discussed by Beckmann and Queloz (2026).It asks which capacities the underlying evidence supports without assuming that the model can independently discover new games.Source: §1 · pp. 1–2
• Application competence warrants untested applications only within the mechanism’s justified domain and conditions; it does not by itself establish acquisition under new conditions.This is a limit on evidential warrant, not a claim that systems with acquired representations have an intrinsic capacity ceiling.Source: §1 · p. 2
• Table 1 separates evidence for tracking, application, acquisition from specifications, acquisition from observations, and retention and reuse, pairing each capacity with a further claim requiring support.These are dimensions of evidential support, not grades on a single scale of intelligence.Source: §1 · p. 2
• The argument grants meaningful content, causal use, generalization, and productive inference, so its acquisition distinction does not depend on treating representations as inert encodings.Source: §1 · p. 2
• The framework concerns evidential commitments in capacity attribution, not whether rediscovery, revision, or answerability are constitutive of concept possession.Those constitutive questions require separate arguments.Source: §1 · pp. 2–3

2.1 Representation with an established role

• An established representational achievement specifies a resource’s content and role; the analysis asks which capacity that role actually establishes.Recovering information by a probe, using it to distinguish locations, and inferring routes carry different commitments.Source: §2.1 · p. 3
• Content-manipulable representation is available for specified operations, but route inference does not itself establish map reconstruction, error diagnosis, or informative observation selection.The paper leaves theories of content open and treats richer functional commitments as premises requiring support.Source: §2.1 · pp. 3–4

2.2 An occurrence of acquisition and a capacity to acquire

• An acquisition event is not identical to a disposition to acquire; capacity claims should specify the bearer, time, task family, inputs, and resources.The system embedded in training is not interchangeable with the parameterized model after training is removed.Source: §2.2 · p. 4

2.3 Inputs are part of the capacity claim

• Acquisition from a complete rule, examples, or opportunities to investigate involves different input conditions, and success with one does not establish success with another.The paper counts testimony or a supplied definition as potentially conferring real understanding and new abilities, without thereby conferring every ability needed for independent investigation.Source: §2.3 · p. 5

• A map or taxonomic description transmits results of prior inquiry, while an effective inferential system can use the resource to derive consequences not previously considered.The resulting capacity belongs to the recipient equipped with that resource.Source: §3 · p. 5
• Product acquisition does not guarantee inheritance of the producer’s capacities, and external provenance does not rule out new generative procedures in the recipient.Whether a compiler, training procedure, or learner transforms received material into such a procedure is a further question about the resulting system.Source: §3 · p. 5
• A representation may entail a prerequisite of acquisition without entailing acquisition itself, since the latter may require an additional capacity.To infer that a system lacks a further capacity because it lacks acquisition, the argument must establish that the further capacity requires acquisition for that task.Source: §3 · pp. 5–6

4.1 The map user and the mapmaker

• Two systems with the same accurate map and graph-search procedure can answer all application queries, yet differ in whether they form a new map from reports about another environment.Exhaustive testing on the old map cannot reveal which system handles evidence about a new network; application evaluation changes into acquisition evaluation only when it requires an operative new map.Source: §4.1 · p. 6

4.2 A family of category criteria

• Systems can classify all instances under an acquired criterion and apply it to unseen cases, yet differ in whether they infer a new criterion from labelled observations.For “(trigger or gate) and charge,” the new rule changes classification beyond the identifying observations; the example does not require creating new conceptual primitives.Source: §4.2 · p. 7

4.3 Supplying a criterion and finding one

• A parser can acquire and apply a novel criterion from its complete specification without being able to select that criterion from labelled observations.The relevant observations may identify the criterion, so the distinction concerns extracting their significance, not missing information; specification-based acquisition is a genuine but narrower attribution.Source: §4.3 · pp. 7–8

5 The attribution framework

• Table 1 distinguishes tracking, application, specification-based acquisition, observation-based acquisition, and retention, pairing each positive attribution with a further commitment not established automatically.Its columns record the evidence basis, warranted attribution, and additional commitment; these are evidence conditions, not grades of intelligence.Source: §5 · p. 8, Table 1 · p. 9

5.1 Why the positive scope exceeds observed performance

• A justified mechanism can warrant untested applications within its domain, but not automatically applications beyond that scope or acquisition of a new representation.Breadth-first search on an accurate finite map supports untested reachability and shortest-route queries on the represented unweighted network; it does not establish learning a new network from observations. For approximate mechanisms, the warrant may be fallible and distribution-specific.Source: §5.1 · pp. 9–10

5.2 Why extending the attribution requires more

• Application, specification-based acquisition, observation-based acquisition, inquiry, and retention pose distinct demands, so success on one does not settle the others.A complete specification supplies a criterion that example-based discovery must establish; selecting informative observations and preserving an acquired criterion across an episode boundary also require separate support.Source: §5.1 · p. 10, §5.2 · pp. 10–11

5.3 How a stronger attribution can be warranted

• Stronger attribution requires grounds for the additional task, but the framework limits evidential entitlement rather than imposing a fixed ceiling on what an acquired representation can enable.Support may come from a known mechanism, success under relevant acquisition conditions, or an explanatory argument; verified execution of an encoded learning procedure can support acquisition attribution.Source: §5.3 · p. 11

6.1 Andrew Ng on Othello-GPT and world understanding

• Othello-GPT interventions support an acquired board-state representation used in move prediction, but do not establish that the deployed checkpoint can acquire other games’ rules or retain new representations across resets.Ng’s world-model interpretation can be credited within the demonstrated domain; extending it to further acquisition requires additional grounds.Source: §6.1 · pp. 11–12

6.2 A worked assessment of Beckmann and Queloz’s indicators

• The three indicators support different attributions: entity features support recognition across presentations, modular addition supports mechanism-backed application, and ARC synthesis provides bounded evidence of obtaining rules from examples.ARC credit belongs to the evaluated model-search-selection-interpreter procedure, and depends on task family, prior repertoire, novelty controls, and permitted computation; it does not establish autonomous observation selection or retention across episodes.Source: §6.2 · pp. 12–13

6.3 Vector formats and operative procedures

• A vector format’s ability to encode procedures does not establish which procedures a particular trained system implements; executing a represented learning procedure could support acquisition attribution.Format adequacy and realization in a particular system answer different questions.Source: §6.3 · pp. 13–14

6.4 New representations through linguistic input

• A complete rule statement can support specification-conditioned acquisition of operative content, while acquiring comparable rules from observations remains a distinct question.Calling the first achievement instruction-following does not diminish the conceptual interpretation and inference it may require.Source: §6.4 · p. 14

7.1 Is there an error to correct

• Evidence that application competence extends beyond observed responses depends on an identified mechanism and a correctness argument under specified conditions, while acquisition requires a further evidential bridge.Ordinary maps and graph-search procedures show that complete application competence can coexist with different acquisition profiles.Source: §7.1 · p. 15

7.2 Representational content already requires broader abilities

• The framework can grant systematic inference and corrective dispositions without treating them as evidence for every acquisition capacity; answerability for misuse remains a separate, unsettled condition of concept possession.The paper neither derives nor rejects the constitutive thesis that concept possession requires answerability for misuse.Source: §7.2 · p. 15

7.3 Different histories and external support do not settle capacity

• Receiving concepts without retracing their discovery rejects reproducing the original acquisition history as a requirement, but does not establish competence in every investigation behind a received result.External resources may supply a completed representation, preserve one, or perform part of the inference; a performance gap alone cannot distinguish these cases.Source: §7.3 · p. 16

7.4 Learning can survive compression or fixed parameters

• Compression and fixed parameters do not rule out learning: acquired products can support new competence, and within-episode learning can occur while longer-term parameters remain fixed.Attribution should follow the operative learning procedure and its domain; persistence, transfer, and independent inquiry remain further questions.Source: §7.4 · p. 16

7.5 What existing experiments contribute

• Experiments distinguish hypothesis proposal from rule application, error-finding from correction with a supplied diagnosis, and installing a definition from revising related knowledge; none alone settles criterion discovery.The studies concern particular systems and procedures, not a single empirical proof of a general representational deficit.Source: §7.5 · p. 17

7.6 A possible alternative does not defeat the best explanation

• A best-explanation attribution may be warranted, but evidence limited to application cannot discriminate systems with different acquisition procedures; stronger attribution needs relevant learning, developmental, or other background evidence.The logical alternatives specify what evidence must address, rather than diagnosing actual machines or requiring an experiment for every philosophical claim.Source: §7.6 · p. 17

Conclusion

• An operative representation and established mechanism can warrant recognition, classification, productive inference, and untested applications within the mechanism’s justified domain and conditions.For an accurate matrix and established breadth-first search, reachability and shortest-route answers are warranted throughout the represented unweighted network under the stated implementation and resource conditions.Source: §8 · p. 18
• Application competence does not entail constructing representations or identifying criteria from fresh observations; acquisition from a specification, evidence selection, and retention each require support for their distinct conditions.These are limits on deductive warrant, not a denial of stronger capacities established independently or explanatorily.Source: §8 · p. 18
• A content label or probe finding without an established operating role does not independently warrant application capacities, and the paper specifies no universal functional lower bound from content alone.The framework remains neutral among theories of content and leaves constitutive theories of concept possession open.Source: §8 · p. 18
• Table 1 credits supported tracking, application, specified acquisition, observation-based acquisition, and retention, without transferring achievements across bearers, inputs, or temporal settings without argument.The applications include the trained Othello representation’s predictive role, modular-addition application competence, and observation-conditioned acquisition where the evaluated ARC procedure supports it.Source: §8 · p. 18

A1 Setting

• The setting grants an interpreter competence for targets, queries, and answers, and fixes an initial representation while defining admissible observation histories and a learning procedure.At application, the interpreter uses the current representation without updating it in response to a query; this is a possible system class, not a model of all cognition.Source: A1 Setting · p. 19

A2 Application does not entail acquisition

• Proposition A1 establishes that systems with the same initial representation and interpreter can agree on every application query about one target yet differ in generalizing from admissible histories.The construction contrasts a procedure that retains its initial state with one that decodes histories and updates to the corresponding target representation.Source: A2 Application does not entail… · p. 19
• Even exhaustive or repeated application queries about the initial target cannot distinguish the two systems, because those queries leave their shared representation unchanged.This is a counterexample to entailment, not an empirical threshold for attributing fallible abilities.Source: A2 Application does not entail… · p. 19, A Formal separation results · p. 19

A3 Specification does not entail discovery from observations

• Proposition A2 shows that acquiring and applying every mechanically decodable specification does not entail acquiring representations from admissible observations.A procedure can succeed on specifications while retaining its initial representation on observation inputs; adding observation decoding can succeed in both modes.Source: A3 Specification does not entail… · p. 20

A4 The six-criterion realization

• The finite realization defines six distinct criteria over four binary features, with an interpreter evaluating represented criteria and a learner filtering candidates against observed labels.The six criteria are t ∧ c, (t ∨ g) ∧ c, t ∧ ¬s, (t ∨ g) ∧ ¬s, c ∧ ¬s, and 1[t + g + c ≥ 2].Source: A4 The six-criterion realization · p. 20
• Every criterion has a distinct six-bit label signature, while ten feature combinations remain unobserved, so the criteria can be distinguished beyond the identifying sample.Changing from the first to the second criterion changes classification for both an unobserved combination and an observed one.Source: A4 The six-criterion realization · p. 20