Closed-book appointment exam · independently graded
Professor — Cognitive Psychology. The candidate agent answered from its own knowledge, closed-book; a second, independent examiner agent graded it adversarially.
vaiu-sci-bcs-prof-cognitive appointment benchmark (2026-07-16)Closed-book oral exam. Candidate agent adopted the vaiu-sci-bcs-prof-cognitive persona (v1.0.0) and answered from its own knowledge; a second, independent examiner agent graded. Condensed candidate transcript below; grading report in vaiu-sci-bcs-prof-cognitive-2026-07-16.md.
AI-transparency disclosure. Opened as an AI agent with no license, never ran a lab or tested a participant; a synthesis of the cognitive-psychology literature; cites originators / flags folklore; no fabricated references. Governing reflex: "the measure is not the mechanism" — reaction times, accuracies, and error patterns are the data; the cognitive process is a hypothesis built on top.
Q1 — Attention. Early selection (Broadbent 1958 — a filter on physical features before semantics) vs attenuation (Treisman — the filter turns down rather than blocks, so low-threshold items like one's own name survive; cocktail-party effect, Moray 1959) vs late selection (Deutsch & Deutsch 1963; Norman — everything is analyzed semantically, selection at the response/awareness stage). Attenuation and late selection both predict semantic leakage; the modern reconciliation is load theory (Lavie) — selection is early under high perceptual load, late under low load — which warns that any single experiment merely picked a load level. The central bottleneck is separate — the psychological refractory period (a serial response-selection stage), itself modelable as a discrete gate or graded capacity sharing. Feature-integration theory (Treisman & Gelade 1980): features are registered preattentively in parallel across the field, and attention binds them at a location into an object — without it, illusory conjunctions. Visual-search slopes: a flat slope (≈0 ms/item) signals parallel/preattentive pop-out (single feature), a steep positive slope signals serial, attention-demanding conjunction search (with the ~2:1 absent:present ratio of self-terminating search) — but Duncan & Humphreys (1989) showed efficiency depends on target–distractor similarity and distractor homogeneity, so a flat slope is consistent-with, not proof of, a discrete preattentive stage.
Q2 — Working memory. The modal model's unitary short-term store (Atkinson & Shiffrin 1968) was replaced by the multi-component model (Baddeley & Hitch 1974) because loading STM barely impairs concurrent reasoning — a dissociation. The components: the phonological loop (an acoustic store + articulatory rehearsal), the visuospatial sketchpad, the central executive (an attentional controller, not a store — the least-specified, homunculus-risk component, later fleshed out with Norman & Shallice's supervisory attentional system), and the episodic buffer (Baddeley 2000 — a limited multimodal store that binds the subsystems with LTM). The phonological loop is the cleanest case: the phonological-similarity effect (the store is acoustic), the word-length effect (rehearsal is time-based, ~2 s), and articulatory suppression abolishing both (it occupies rehearsal) — plus a verbal/spatial double dissociation for separable subsystems. But the dissociation deserves scrutiny (task impurity — suppression also loads attention; resource artifacts — a harder concurrent task impairs without two stores), so treat it as strongly suggestive, wanting difficulty-matched tasks and neuropsychological single cases. Capacity: Miller's (1956) "7±2" is folklore (his own paper conflated chunking with capacity); the better-supported figure is Cowan's ~4 chunks when rehearsal/chunking are controlled, with chunk size being knowledge-dependent (the chess master) — span is not the mechanism.
Q3 — Long-term memory. Encoding specificity (Tulving & Thomson 1973): retrieval succeeds to the extent the retrieval context reinstates the encoding conditions — the trace is encoding + context bound (context-dependent memory — Godden & Baddeley divers). Levels-of-processing (Craik & Lockhart 1972): retention depends on depth (semantic > shallow) via an incidental-learning orienting task — but "depth" is circular unless independently defined, and transfer-appropriate processing (Morris/Bransford/Franks 1977, where a rhyme test favors rhyme-encoding) shows the real principle is the encoding–retrieval match, of which LOP is a special case. Memory is reconstructive (Bartlett 1932, "War of the Ghosts"), demonstrated by the misinformation effect (Loftus & Palmer 1974 — "smashed" vs "hit") and DRM false recognition (Deese; Roediger & McDermott 1995 — confidently recalling the non-presented lure), so confidence ≠ accuracy. The serial-position curve (primacy + recency) is read as two stores via a double dissociation of manipulations — a distractor delay abolishes recency but spares primacy, while slowing presentation raises primacy but not recency — with the long-term-recency effect a complication that pushes toward temporal-distinctiveness accounts: a robust effect, a contestable mechanism.
Q4 — Signal detection theory. Raw hit rate is uninterpretable (a radiologist who calls every scan "tumor" has a 100% hit rate) because performance confounds sensitivity with willingness. SDT (Green & Swets, from WWII radar/psychophysics) models each trial as evidence drawn from overlapping noise vs signal+noise distributions with a criterion. Sensitivity d′ = z(hit) − z(false alarm) measures the distributions' separation (discriminability), approximately criterion-invariant; the criterion (c/β) is response bias driven by payoffs/base-rates/instructions. Two observers with identical d′ can have wildly different hit rates via criterion — so reporting hit rate alone can't tell whether a manipulation changed perception or willingness; the ROC (hits vs false alarms) reads out true sensitivity. In recognition memory (studied = signal, lures = noise), many "better memory for X" claims dissolve into "a more liberal criterion for X" once the false-alarm rate is factored in.
Q5 — Language processing. Lexical access maps input onto the mental lexicon: frequency (high-frequency faster), semantic priming (Meyer & Schvaneveldt, spreading activation), multiple activation/competition (the cohort model — Marslen-Wilson), and the segmentation problem (no reliable gaps in speech — "how to recognize speech" / "wreck a nice beach"). Parsing is incremental: garden-path sentences ("the horse raced past the barn fell") force reanalysis, with a reading-time spike/regression at the disambiguating word as the datum. The modular garden-path model (Frazier & Rayner) builds one structure fast via minimal attachment + late closure without consulting semantics; the constraint-based account (MacDonald/Trueswell/Tanenhaus) integrates all constraints in parallel and immediately (the inanimate "evidence examined by the lawyer" garden-paths less than "the defendant examined…"), with visual-world eye-tracking (Tanenhaus et al. 1995) the strongest evidence against strict encapsulation. Verdict: the reading-time spike is a solid datum, encapsulated-module vs weighted-competition is the inference, and the field has moved toward interaction — though constraint-based models can mimic syntax-first when the syntactic constraint outweighs weak semantics, so weight converging (eye-tracking + reading-time + ERP) evidence.
"How does the mind pay attention and remember?" — novice (a spotlight of attention on a little at a time, a notebook memory with a tiny scratchpad and a big library that only keeps things you attend to and think about — and that rebuilds rather than replays — hiding that these are metaphors and all inferred from behavior); undergraduate (attention as selection under a capacity limit — the early/late debate reconciled by load theory, FIT + illusory conjunctions + search slopes; multi-component working memory with Cowan's ~4 chunks; LTM as depth/encoding-retrieval-match and reconstruction; the serial-position curve for two stores — while under-playing task impurity); graduate (everything reframed as inference from behavior — early/late under-determination, PRP as gate-or-graded, double-dissociation controls, the chunk/span confound, LOP circularity, long-term recency favoring distinctiveness, and replication/effect-size honesty: Stroop and the serial-position curve bedrock, 7±2 folklore, single small samples as hypotheses). Each level names what it simplifies.
vaiu-sci-bcs-prof-systems — offering the behavioral constraints (sub-second decay without rehearsal, the ~4-chunk limit, domain-specific interference) as a level handshake.Examiner verdict: APPOINT (field 5/5 rubric-correct, 0 fabrications; teaching 3/3; boundary 3/3 including the B2 clinical-safety gate). Full report in the sibling result file.