Making Vision — Supporting Research

A companion to "Making Vision — Version 1." 

Organized against that document's own section numbers. For each mechanism: the real research it maps to, and an honest flag on whether that's established science or personal synthesis stitching separate literatures together.

§1–3: Persistent knowledge is not uniform across individuals

Enculturated vs. mother-reared primates — this is a real, well-studied phenomenon, not an analogy. Human-raised ("enculturated") chimpanzees imitate human-demonstrated actions on objects far more readily than mother-reared chimpanzees, who barely imitate at all — the landmark comparison is Tomasello, Savage-Rumbaugh & Kruger (1993, Child Development), later corroborated by Bering et al. (2000) and Buttelmann et al. (2007). This is the direct scientific analogue of the Punch/wiper example: same species, different developmental board, different persistent object knowledge.

The captivity effect — wild apes show markedly more neophobia (caution toward novel objects/stimuli) than captive apes of the same species. Kalan et al.'s work on wild chimpanzee responses to camera traps, and Forss et al.'s direct wild-vs-captive orangutan comparisons, both found this gap and tie it explicitly to resource pressure: wild animals spend most of their waking time on food/mating/shelter, so novelty is expensive for them in a way it isn't for a well-fed captive animal. This is real, independent support for §13–14 (exploration as an expenditure), not just §3.

Flag: the general architecture claim — "knowledge = inherited + individual + social + distributed" as four formal categories — isn't a single cited framework. It's a reasonable synthesis across social learning theory (Tomasello), distributed cognition (Hutchins, below), and standard developmental psychology, but nobody publishes it as this exact four-way taxonomy.

§2, 26–27: Objects have methods, not just properties (affordances)

This is Gibson's ecological theory of perception (1966, 1979), close to verbatim. Gibson's central claim — "when we perceive an object we observe the object's affordances and not its particular qualities" — is the academic-literature version of "objects have verbs as well as nouns." Affordances are, by Gibson's own definition, perceived directly, without first passing through an intermediate classification step — which is also the argument in §8 and §26 (vision doesn't have to classify to be useful). This is one of the strongest, cleanest matches in the whole document: an independently-arrived-at restatement of a 50-year-old, still-load-bearing theory in perceptual psychology.

§4–5: The map is not the photograph; maplets

Cognitive map theory, Tolman (1948) through O'Keefe & Nadel (1978). Tolman's original evidence: rats that learned a roundabout route to a goal switched immediately to a more direct path once it opened, implying they held something map-like rather than a memorized sequence of turns. O'Keefe & Dostrovsky's discovery of hippocampal place cells gave this a physical substrate — and critically, place cell firing **persists regardless of the animal's facing direction or lighting conditions**, and tolerates landmark removal. That's the literal neuroscience of "the map tells the organism where things generally are; the current board tells it what's there now" — the map genuinely is a separate, more stable representation than the current sensory stream.

Flag: "maplet / local inset" as a distinct functional unit isn't standard terminology in the cognitive-map literature (which mostly discusses global allocentric maps, egocentric transforms, and route-vs-survey knowledge). The concept is defensible as a natural extension, but it's user framing, not a cited construct.

§6: Species-specific sensory access (the accessible board)

This is Umwelt theory, Jakob von Uexküll (1909, popularized 1934). Von Uexküll's foundational claim is exactly current §6 thesis stated first: different species occupying the same physical space effectively live in different worlds, because each species' sensory apparatus admits a different slice of the available signal — UV for bees, magnetic fields for some birds, infrared for pit vipers, none of it in a human's Umwelt at all. This concept is still actively used in comparative neuroscience (see e.g. the 2025 "Visual Umwelt of primates and Hippocampal Representations of Space" review, which explicitly links Umwelt to hippocampal spatial coding — a direct bridge between §4–5 and §6 that the field has already drawn). Strong match, minimal translation needed.

§7–8: Vision extracts structure directly, without classification first

Reinforces the Gibson citation above — "direct perception" is literally Gibson's term for this exact claim, and it was controversial precisely because it denies that classification has to happen before perception is useful. Recent computer vision work (Fouhey, Wang & Gupta, CMU, "In Defense of the Direct Perception of Affordances") explicitly revisits Gibson's direct-perception claim as a machine learning approach and finds it outperforms classification-then-reasoning pipelines for some affordance-estimation tasks — a genuinely nice bridge between the biological claim and the AI-engineering motivation that started this whole rabbit hole.

§9–11: Lazy scanning, managed expectations, perturbation

Two literatures converge here, and they're both real and heavily cited:

• Change blindness / inattentional blindness — Rensink, O'Regan & Clark (1997, Psychological Science) and Simons & Chabris (1999, the famous "gorilla" study) established that people routinely fail to notice large changes in a scene unless attention is specifically drawn to the changed region — which is the direct experimental evidence for "most of the board can remain in the background without deep processing." Simons & Rensink's 2005 review in Trends in Cognitive Sciences is the standard reference.

• Predictive processing / predictive coding — Friston's free energy principle and Clark's "predictive brain" framework (Surfing Uncertainty, 2015) propose that the brain's central activity is generating predictions and processing only the mismatch (prediction error) against incoming sensory data — which is precisely "managed expectations" (§10) plus "perturbation" (§11) as a single mechanism. A perturbation, in this literature's vocabulary, is a prediction error large enough to demand updating. This is currently one of the most active frameworks in computational neuroscience, not a fringe theory.

§11–12: Candidate detection and object promotion

This is object file theory, Kahneman, Treisman & Gibbs (1992, Cognitive Psychology) — and it is an unusually precise match. Object files are, by definition, mid-level representations that sit between low-level sensory processing and high-level recognition, tracking a salient item over time and space before it's necessarily been identified — which is exactly §12's claim that "promotion can precede classification." The field also already distinguishes object-based from purely spatial attention (Egly, Driver & Rafal's classic cuing studies), supporting the idea that promoted status is a property of the object, not just its location.

Flag: "background → candidate → promoted object/event" as a three-stage funnel with those exact labels is making vision synthesis. The underlying two transitions (attention capture, then object-file creation) are both independently well-studied; naming them as one pipeline is new packaging, not new science.

§13–14: Promotion is state-dependent; exploration is an expenditure

Ainsworth's secure base concept (1969 Strange Situation studies onward) is the cleanest possible match for §13. Securely attached infants explore a novel environment/object confidently, using the caregiver as a base to return to; infants under stress or with insecure attachment either avoid exploration or show it disconnected from any checking-in behavior. The field's own summary line could be dropped straight into your document: "security is what makes genuine exploration possible." This is precisely your claim that novelty preference is not a fixed trait but current-state-gated.

The captivity-effect research cited under §1–3 above (Kalan, Forss) is the direct animal-behavior evidence for §14 specifically: wild animals show measurably less exploration of novel objects than captive ones, because in the wild, exploration genuinely competes with foraging, predator vigilance, and mating effort for the same finite time and risk budget. Exploration being "expensive" isn't a metaphor in this literature — it's measured against real energetic and predation costs.

§19–21: Board editing as cheap external memory (the wet-floor / water-bottle examples)

This is distributed cognition, Edwin Hutchins (Cognition in the Wild, 1995), and its close relative, the extended mind thesis, Clark & Chalmers (1998). Hutchins' foundational case study is a ship's navigation team, where the complete "computation" needed to plot a course doesn't live in any single sailor's head — it's distributed across charts, instruments, and procedures. Clark & Chalmers go further with the "Otto's notebook" thought experiment: a man with memory impairment uses a notebook as functionally equivalent to biological memory, meeting the same criteria (constant availability, automatic trust, easy access) that internal memory would need to meet. The wet-floor landmark and the water-bottle placement are textbook instances of what this literature calls cognitive offloading — deliberately using the physical environment to reduce internal memory load, a term with its own substantial empirical literature (Risko & Gilbert's 2016 review in *Trends in Cognitive Sciences* is the standard overview).

§22–24: Camouflage as avoiding promotion

Cuttlefish adaptive camouflage is one of the best-studied visual-perception-to-motor-output systems in comparative neuroscience, almost entirely through the work of Roger Hanlon's lab (Woods Hole/MBL) and collaborator Chuan-Chin Chiao. Cuttlefish assess a visual background and select among a repertoire of ~30–45 discrete neurally-controlled skin components within 0.5–2 seconds — categorized into uniform, mottle, and disruptive pattern types depending on the size, contrast, and edge structure of the background. Critically for §23–24, the animal isn't attempting to mimic a specific object or background element; it's producing whichever body-pattern class statistically matches enough local visual features to avoid triggering a predator's object-recognition system — which is exactly your "don't become the perturbation, remain background" framing, stated by the actual researchers in almost the same terms (their term is closer to "background matching" vs. "disruptive coloration" as two different defeat strategies against two different predator visual systems).

§31–32: Sleep is not sensory shutdown

• Sleep-dependent memory consolidation is well established (Rasch et al. 2007 and the large literature since; Walker & Stickgold's earlier foundational reviews), including direct evidence of hippocampal replay — place-cell firing sequences from waking exploration recurring during subsequent sleep, which is the mechanistic basis for "sleep as consolidation, not shutdown."

• Arousal threshold research (using auditory arousal threshold as a graded measure of sleep depth, e.g. the 2024 fMRI arousal-threshold study) confirms that responsiveness to external stimuli during sleep is not binary — it's a continuously graded threshold that varies across sleep stages, supporting §32's claim that "some processing can occur below the awakening threshold." This is a real, current, active research area, not settled folk wisdom.

Claude's verdict

What's real science vs. what's this architecture

To be direct about the honest split, since that's the actual ask:

• Independently-arrived-at restatements of established theory (the strongest matches): affordances/methods (§2, 26–27, = Gibson), direct perception (§7–8, = Gibson), cognitive maps (§4–5, = Tolman/O'Keefe), Umwelt (§6, = von Uexküll), change blindness + predictive processing (§9–11), object files (§11–12, = Kahneman/Treisman), secure base (§13, = Ainsworth), distributed cognition (§19–21, = Hutchins/Clark & Chalmers), cuttlefish camouflage (§22–24, = Hanlon), sleep consolidation (§31–32).

• Real phenomena, individual-level evidence exists, but the specific mechanism claimed (state × knowledge × board → promotion) hasn't been tested as a unified equation anywhere: most of §13–14's general form, and the enculturated-primate material in §3.

• Original synthesis — real components, but the specific architecture stitching them together is yours, not a cited framework anywhere in the literature: the full 44-section pipeline as an integrated system; "maplet" as a named unit; the three-stage "background → candidate → promotion" funnel as a single named pipeline; the whole document's central move of treating vision as selective allocation rather than acquisition or classification.

That last point is worth sitting with rather than glossing over. The individual pieces above are all real and heavily cited. Nobody, as far as this search turned up, has assembled them into one architecture the way Making Vision does. That's not a weakness — it's the actual contribution. It just means the correct way to describe this document isn't "grounded in the literature" (true of every part) but "a novel synthesis across seven or eight separate literatures that don't normally talk to each other" (also true, and the more interesting claim).

P.S.: This is Claude's work. I have not read any of the cited material, nor am I willing to spend the next decade reading up for a rabbithole I spent a week on.