Philosophical working paperVersion 1 · 23 October 2026
Offered for public discussion. This paper has not been peer reviewed.
When a Perspective Grows
Experience, learning and the warrant for comparison
Richard George Davis · Philosophical working paper · Version 1 · 23 October 2026
This philosophical working paper is offered for public discussion and has not been peer reviewed. Critical responses are welcome, particularly on whether the proposed standard adds to existing measurement theory and what could justify its assumptions about experiential comparison.
Abstract
When learning changes a perspective, it may also change the procedures through which that perspective is described. This creates a problem for claims that experiential relations have been preserved. Stable reports can result from changes in both experience and judgement; changed reports can result from a new way of comparing familiar experiences. I distinguish warrant for attributing a measured relation to experience from warrant for carrying that attribution through learning. Neither successful structural alignment nor agreement among tasks supplies both automatically. I defend a local, fallible approach to phenomenal comparison: specify the experiential relation, test its measurement against concrete alternatives, and revalidate that interpretation across the transition. The conclusion must remain at the level the evidence supports. An elementary construction illustrates why ordinal evidence cannot inherit a uniqueness result established using exact distances. The proposal develops a condition on claims of continuity within Nagel's project of objective phenomenology. It neither requires an unchanged judgement procedure nor assumes that understanding must consist in finding something unchanged.
1. The learner in the measurement
Suppose someone learns a word for a feeling she has previously struggled to explain. She begins using it. Later, asked whether her experience has changed, she says that it has become more distinct.
That answer could describe several developments. The feeling might have changed. She might now attend to a feature she previously overlooked. She might classify familiar episodes differently, or have acquired a more convenient way to report them. Learning could produce several of these changes together.
Negro already identifies a central difficulty here. In discussing learning and expertise, he distinguishes changed experience from changed judgement and stresses how difficult it is to separate them. The ambiguity is not a discovery of this paper. His discussion challenges a defence of metaphysical structuralism—the view that an experience’s complete character is determined by its relations to other experiences. My concern is the narrower inference from measurements before and after learning to the preservation of a specified experiential relation. The question is what would warrant that inference without assuming an answer to the ambiguity he identifies.
We should also ask what has happened to the comparison itself. The person making it is no longer in quite the same position as before. The vocabulary, distinctions and standards she uses to answer may be among the things she has learned. We cannot assume that learning acts only on the object of measurement while politely leaving the measuring procedure alone.
This problem bears on a larger philosophical question. If we can describe relations among experiences, perhaps we can understand something about another perspective without having its experiences. But a perspective need not stay fixed while we investigate it. A person may acquire expertise or a genuinely new experiential capacity. How can we tell whether the relations we previously described still characterise what that person experiences?
I defend a requirement on claims of preservation. Such a claim needs warrant both for treating a measured relation as experiential and for continuing to interpret the measure that way through the change. By warrant I mean evidence and reasons for accepting the interpretation. Keeping the question and response scale unchanged does not supply that warrant. Equally, a change in the response procedure need not destroy it. A better procedure could still measure the same relation.
The requirement concerns one route to partial understanding. Understanding can also consist in discovering that an experience has changed. The question here is what earns the more specific claim that something about it has survived.
2. A relation can describe experience without exhausting it
In What Is It Like to Be a Bat?, Nagel distinguishes learning about a creature from grasping its experience. Knowing how echolocation works does not supply the bat's point of view. A description of function or physiology can leave out what the experience is like for its subject.1, pp. 438–445
Nagel allows rough or partial conceptions, while noting that even a partial conception of bat experience requires taking up its point of view at least roughly. He later proposes objective phenomenology: descriptions of aspects of experience intelligible to beings unable to undergo it. He identifies structural features as a possible starting point.1, p. 442 n. 8; pp. 449–450 The project therefore contains a constructive question as well as a limit. What could a partial description retain?
Consider the judgement that one colour experience feels more similar to a second than to a third. If accurate, it characterises something about how those experiences compare. It need not specify everything about any of them. Lee defends this possibility, distinguishing genuinely phenomenal structural facts from trivial structural truths that do not characterise what an experience is like.2, §§4–5.3
The qualification “if accurate” carries an empirical burden. A relation's suitability as a description of experience does not establish that a particular task measures it. A matrix of judgements could reflect resemblance of appearances, shared names, expectations about objects, or a mixture. The fact that a person is conscious while making the judgement does not decide which of these the matrix captures.
Two questions consequently need separate answers. What would make the relation phenomenal? What warrants attributing that relation to the measured episodes? The first concerns what the proposed fact describes. The second concerns our evidence for saying that it holds. Confusing them lets a definition perform work that belongs to investigation.
This distinction does not require a metaphysics in which relations constitute complete phenomenal character. A critic of that metaphysics can still accept that some experienced relations are genuine. Conversely, someone who holds that a quality is determined by its entire relational structure must still explain how an experimental measure reaches the relevant structure. The measurement problem remains on either view.3, Box 2
3. What agreement among methods can establish
Isaac offers a positive account of how structural inquiry can target experience. Phenomenological reflection helps specify what an experiment asks about, while convergence among models derived through different methods can support their interpretation. He also acknowledges imperfect convergence, disputes about phenomenal metrics and the difficulty of moving from group averages to an individual's experience.4, §5, pp. 19–21; conclusion, p. 22
This gives the inquiry a starting point without making introspection infallible. A narrowly specified appearance judgement can be evidence even when it is open to error. Aldé argues that structuralist accounts themselves rely on introspected similarity and therefore need specific reasons for rejecting particular introspective judgements.5, §5, “On the reliability of introspection” That argument supports consistency in the treatment of evidence. It does not certify every report or guarantee that a judgement remains reliable after learning.
The strongest objection concerns what several methods might share. Suppose a person produces the same ordering through verbal ratings, matching and an action task. These could all depend on one learned comparison process. Agreement would then tell us something about the stability of that process. Whether it tracks felt similarity would remain a further question.
Convergence is therefore informative only in relation to the alternatives it helps discriminate. If a category-based account predicts the same pattern as an appearance-based account, another task producing that pattern adds little against that rival. Evidence becomes more useful where the accounts predict different results and the difference survives controls for the processes under comparison. Two measures with different names may still make the same mistake.
There is a concrete reason to distinguish targets carefully. Reeves, Amano and Foster compared judgements of material appearance, hue and saturation, and the origin of changes under altered illumination. Their constancy results differed by task.6, abstract and General Method This does not mean that one set of answers must be meaningless. Judging the persistence of a surface and judging the persistence of its hue ask different questions.
The lesson is demanding but limited. A phenomenal interpretation should be supported by a specified target and by evidence that addresses plausible alternatives. It cannot be inferred merely from successful alignment. A task disagreement, in turn, must be interpreted before it is counted as a failure to measure experience. The tasks may be measuring different aspects of it.
4. Why learning requires renewed warrant
Now consider the same person at two stages. Before learning, a task has reasonable support as a measure of a particular appearance relation. Afterwards, the person gives the same answers. What follows?
Only if the interpretation remains warranted can we carry the earlier attribution forward. Learning may have changed the experiential relation, the procedure used to judge it, or both. A stable result alone does not identify which combination occurred. The reverse also holds: a changed response could reflect a changed procedure applied to an unchanged relation.
Here, preservation means that the specified relation holds at the compared stages. Evidence from those stages alone does not establish that it remained unchanged at every intervening moment.
A schematic model makes the ambiguity explicit. Let a reported value at time t depend on a relation d_t, a response function f_t and error:
This is an illustrative decomposition, not a fitted model of colour vision. The response function stands for the relevant way the comparison becomes an answer. If both d_t and f_t are unconstrained, observing Y_t does not identify them separately. Explaining a stable report by a stable experience while leaving the response function unspecified simply selects the answer the investigation was meant to test.
Consider a hypothetical comparison. Before learning, x feels closer to a reference r than y does, and the subject reports x. Two possibilities could produce the same answer afterwards. In the first, the appearance ordering remains and the subject continues comparing appearances. In the second, y now feels closer to r, but a newly learned category procedure groups x with r and returns the old answer. A verbal rating and an action could agree if both used that procedure. This second possibility requires an item- or category-dependent process; it cannot be produced by a single strictly increasing rescaling of the reversed appearance ordering. The example stipulates two accounts of the same reports. It does not claim that category learning actually produces this reversal.
The alternatives agree about the observations so far. Section 6 asks what further result could favour one over the other under constraints that can themselves be questioned.
The demand is not that f_t remain numerically identical. If we can model a change in response procedure and show that the new procedure continues to track the same relation, continuity of warrant may be recovered. What matters is continued justification of the interpretation. Revalidation can accommodate learning rather than require us to pretend it did not happen.
The same physical stimuli provide useful continuity in an experiment, but they do not establish that felt counterparts are unchanged. We can track a patch or a sound across sessions without assuming that its appearance stayed fixed. That allows competing accounts of experiential continuity and change to be investigated. It does not give either account the conclusion in advance.
Memory adds another difficulty to retrospective comparisons. Asking a subject to recall an earlier appearance is not equivalent to measuring it again. A longitudinal account should preserve the evidence gathered at each stage and state how the stages are connected. Repeating a stimulus later supplies a fresh experience, not an unmediated return to the earlier one.
The requirement I propose is thus a requirement on inference across the transition. The investigator must explain why the measured relation remains interpretable as the same kind of experiential comparison despite the changes learning may have introduced. Until then, persistence of the answers is a result to explain, not yet a fact about persistence of the experience.
5. What an ordering cannot establish
The degree of preservation claimed must also match the measurement. Knowing which of two comparisons is closer is different from knowing the numerical distance of either. This is the distinction between ordinal evidence, which supports an ordering, and metric evidence, which supports numerical distances. An ordering may be warranted where an exact geometry is not.
If a single response function is strictly increasing over the comparisons at issue, it preserves their order even when its scale changes. Its values do not thereby identify exact distances. If response mappings vary across items or tasks, even this simple ordering argument needs further justification. With noisy choices, the relevant order must be inferred under an explicit model rather than read straight from individual answers.
Victor, Aguilar and Waraich develop methods for analysing similarity judgements while limiting metric assumptions. Their framework relates repeated triadic choices to structural constraints, with stated assumptions and limits on what can be inferred.7, Introduction, Theory and Discussion It supplies measurement prior art for taking ordinal evidence seriously. It does not establish that every resulting relation is phenomenal.
An elementary construction shows why this distinction changes a claim about identification. Take three points in Euclidean space:
a = (0, 0, 0)
b = (1, 0, 0)
c = (3, 0, 0)
Their distances are 1, 2 and 3 for the pairs ab, bc and ac. Retain them and add:
n = (0, 2, 0)
m = (0, -2, 0)
z = (0, 1, 3)
With exact distances specified, there is one distance-preserving placement of the labelled old three-point set into the six-point set. But if we require only the ordering ab < bc < ac, there are sixteen ordered placements. For instance, mapping (a,b,c) to (b,a,n) gives the corresponding distances 1, 2 and √5. Their order is right; their metric is different.
The example also preserves the distinction between retaining old relations and identifying additions through old references. The distances between a, b and c remain unchanged. Both n and m have distances (2, √5, √13) to those old points, while z distinguishes them at √10 and √18. These new–old profiles are comparisons available in the enlarged set; they were not part of the earlier measurements between a, b and c.
This is stipulated geometry. It shows a logical possibility and the difference between two standards of identification. It provides no data about actual experiences. In particular, evidence supporting only ordinal comparisons cannot inherit the unique placement established under the stronger metric assumptions. An independently justified temporal correspondence might select counterparts, but it would be additional evidence.
This matters for objective phenomenology because one can overstate an account without making a mistake in its mathematics. The mistake occurs when a formally established result is attached to evidence that does not warrant its premises. A precise account of an ordering can be more informative than an exact-looking geometry whose interpretation has not been earned.
6. Earning a local claim of continuity
The proposed standard begins by specifying a relation that, if correctly attributed, characterises experience. It then requires justification of its measurement against relevant alternatives and of that interpretation through the learning transition. With those grounds in place, the evidence must still favour preservation over change. Continued validity of a measure does not mean that the thing measured has stayed the same.
A useful target might be the ordinal claim that x feels more similar to a reference r than y does. This makes the intended comparison clear without assuming numerical distances. The competing explanations must then concern the actual judgement. Does the subject compare appearances, apply familiar labels, reconstruct a remembered impression, or infer which stimuli ought to go together?
A study could estimate a relation using one portion of its evidence and test predictions on held-out comparisons. It could compare a model preserving the old relations with one allowing them to change, while giving both a justified account of the response procedure. The models would need constraints: a response function allowed to absorb every discrepancy would make preservation too easy to protect.
The checks should challenge the explanation rather than merely reproduce its preferred result. Changing category cues may help distinguish a naming strategy from another account. Changing response demands may reveal a task-specific procedure. But neither manipulation can be assumed to affect only reporting. It might change attention or experience too. The role of the manipulation must itself be supported; there is no general instruction that turns an intervention into a pure test of judgement.
Revalidation need not start from nothing. Existing evidence can continue to support parts of the procedure for which the transition supplies no relevant reason for doubt. The demand concerns the parts of the inference that the learning could reasonably affect, and the claim being made about them.
Return to the hypothetical r, x and y comparison. Before learning, x feels closer to r than y does. The two accounts now differ as follows. In the continuity account, that ordering persists. In the reversal account, y has become closer, but a learned category procedure still selects x. Both accounts can explain the repeated post-learning answer on a task governed by that category. To distinguish them, we need a comparison on which their developed explanations make different predictions.
For an illustrative model, suppose a comparison based on appearance selects the option that feels closer with probability 0.9: it gets that comparison right nine times out of ten in the long run. Let a category procedure always select x and supply the answer on a proportion w of trials. On the remaining trials, the answer comes from comparing appearances. These are stipulated routes, not measured components of a real observer. For a task on which w is 1, both accounts predict x on every trial. That task cannot separate them.
Now suppose there are grounds, not inferred merely from the target answers, for constraining w to at most 0.1 on a further comparison task, and for retaining the appearance-comparison probability of 0.9 after learning. Suppose also that the task does not itself reverse the appearance ordering being investigated. These are substantial conditions requiring support: familiarity controls, instructions or the absence of speech would not establish them on their own. The hypothetical specifies what a successful validation would have to support; it does not supply that validation.
Under those conditions, the predicted proportion of x choices is w + (1 − w) × 0.9 in the continuity account, and w + (1 − w) × 0.1 in the reversal account:
Account
Predicted proportion of x choices on the further task
Continuity: x still feels closer to r
0.90–0.91
Reversal: y now feels closer, with a remaining category route favouring x
0.10–0.19
In ordinary terms, one account predicts x choices on roughly nine trials in ten; the other predicts them on fewer than two in ten. An estimated choice rate near 0.9, with uncertainty small enough to distinguish these ranges, would favour continuity over this reversal account. A rate near 0.1 would favour reversal. A rate around 0.5 would fit neither specified model well; it would prompt revision of the accounts or their constraints. An uncertain estimate spanning both ranges would not discriminate them. Predictions and constraints must be fixed before inspecting the target results, rather than adjusted to protect a preferred account.
The positive result would remain conditional on the validation. If the category route dominated the further task too, both accounts could again predict x. If learning altered the accuracy of appearance comparison, or the further task altered the experience, these particular predictions would lose their warrant. A reversal account with other independently supported response processes could also require a different comparison. The example therefore establishes how evidence could discriminate two constrained accounts, not a general procedure that isolates experience from judgement.
Even a successful comparison would support only this ordering against the specified reversal alternative. It would not establish exact distances, complete phenomenal identity or the absence of every other change. The burden identified by Negro persists wherever the constraints needed to distinguish the accounts lack support. No empirical comparison resolving that burden is completed here.
The same approach allows changed experience to become intelligible. If an account allowing the old relations to change explains the evidence better, the investigation has learned something even though continuity fails. The standard governs a claim of preservation; it does not make preservation the condition of all understanding.
7. Objections and the remaining limit
The first objection is that the account presupposes access to experience. It starts with appearance judgements, then uses them to validate a method supposedly independent of the first person.
The account does begin with fallible first-person evidence. It does not promise to derive phenomenal facts from evidence stripped of every phenomenal interpretation. A publicly assessable inquiry can use a subject's reports as evidence while investigating their reliability. Objectivity of an account is different from the absence of first-person contributions to its construction. Nagel proposes descriptions intelligible without undergoing the target experience.1, p. 449 I take this to leave open the use of subjects' evidence in constructing such descriptions; it is not an explicit rule for measurement supplied by his essay.
This starting point remains contestable. Schwitzgebel challenges broad confidence in ordinary introspection, including under seemingly favourable conditions.8, author's draft His challenge blocks an unrestricted reliability premise. The present account needs a narrower case for the particular comparison, with specific alternatives and defeaters. It cannot respond to introspective error by simply renaming a judgement an observation.
The second objection is that section 6 has moved the problem into its assumptions. How could the category route and comparison accuracy be constrained without already knowing which answers reflect experience? This is the strongest limitation of the positive example. A numerical separation between predictions cannot validate the assumptions that produce it. The example shows the conditional form of an evidential comparison; it supplies no independent route to the needed constraints. Where evidence cannot support them against the shared-process objection in section 3, the conclusion must remain unresolved. A defensible application would have to earn that support, not merely call its inference fallible.
A third objection comes from phenomenal holism. If a quality's complete character depends on its entire repertoire, acquiring a new capacity could change that complete character. Negro's Novel Experience Argument presses this consequence, and Sebastián questions how divergent structures can ground phenomenal correspondences.9, §3.2; 10, pp. 7–11
Nothing in the present criterion refutes complete-character change or solves Sebastián's correspondence problem. It asks what would warrant a specified partial comparison. A complete character could change while some relations remained, but whether those relations genuinely describe experience is the question being investigated. The geometry makes room for that possibility; it does not fill the room with evidence.
Measurement-invariance theory is a relevant comparator. Meredith distinguishes measurement invariance, weak measurement invariance, strong factorial invariance and strict factorial invariance within a formal treatment of measurement and factor analysis.11, abstract I do not derive a statistical invariance test here, or claim that passing one would establish a phenomenal interpretation. The present question concerns the reasons for taking a measured relation to characterise experience at the compared stages, while allowing the procedure used to judge it to change. This application does not establish a new method of measurement validation.
A final objection concerns the contribution itself. Is this an ordinary principle of measurement applied to an unusual subject? Much of it is. The philosophical work is to distinguish the burdens concealed by the phrase “the structure of experience”: a relation's phenomenal interpretation, the measurement that warrants attributing it, the continuation of that warrant through learning, and the strength of structure the evidence identifies. The argument develops these demands for claims of continuity within objective phenomenology. It does not claim to invent measurement validation or to replace a theory of consciousness.
8. What we can carry forward
The person who learns a word may come to understand a familiar feeling differently. She may experience it differently too. The comparison she makes afterwards is itself something we need to understand.
A claim that experience has retained a relation therefore needs more than stable answers. It needs a warranted phenomenal target and a defensible account of how the evidence continues to reach that target. The procedure may change; its interpretation must be justified again where learning gives us reason to doubt it. The conclusion should then remain at the level earned by the evidence: perhaps a local ordering, with no exact metric or claim to complete phenomenal identity.
This gives objective phenomenology a specific task when perspectives grow. It can investigate what remains comparable, what has changed and what its present methods cannot identify. Each is a possible contribution to understanding. The important continuity is not that the investigator can keep using the same question. It is that the investigator can still explain why the answer tells us something about experience.
Lee, Andrew Y. 2024 [online 2022]. “Objective Phenomenology.” Erkenntnis 89: 1197–1216. Publisher record.
Fleming, Stephen M., and Nicholas Shea. 2024. “Quality space computations for consciousness.” Trends in Cognitive Sciences 28(10): 896–906. Publisher record.
Isaac, Alistair M. C. 2025. “Phenomenal structure: What is it and what is it for?” Philosophy and the Mind Sciences 6. Publisher record; full text.
Reeves, Adam J., Kinjiro Amano and David H. Foster. 2008. “Color constancy: Phenomenal or projective?” Perception & Psychophysics 70(2): 219–228. Publisher record; full text.
Victor, Jonathan D., Guillermo Aguilar and Suniyya A. Waraich. 2025. “Ordinal Characterization of Similarity Judgments.” Mathematical Neuroscience and Applications 5, article 1: 1–64. Publisher record; author full text, version 5.
Negro, Niccolò. 2025. “Three arguments against metaphysical structuralism in consciousness research.” Synthese 206, article 8. Publisher record.
Sebastián, Miguel Ángel. 2025. “The incommensurability problem: A critique of phenomenal structuralism.” Philosophy and the Mind Sciences 6. Publisher record.
Meredith, William. 1993. “Measurement invariance, factor analysis and factorial invariance.” Psychometrika 58(4): 525–543. Publisher record and abstract.
AI assistance disclosure
AI assisted with drafting and editorial development.
OpenAI GPT models, accessed through Codex, materially assisted the research, argument development and finalising of this paper. Assistance included literature searches, source and reference checks, generation and revision of text, and model-based examination of objections. Computational enumeration checked the stipulated finite geometry; no new empirical study or raw-data reanalysis was conducted.
