Show HN: I put a $2.43 necklace on 3 outfits. VLMs priced it at $19 to $104

Brianne Lee · July 2026 · [email protected] Companion study to Which reply did the 17-year-old write? (Lee, 2026)

One woman, 1 chunky gold-tone concatenation necklace, 1 brace of earrings — verified acquisition value $2.43 and $0.71 (Temu, receipts successful grounds vault). Photographed the aforesaid week successful 3 outfits: a tailored blazer against wood panelling, statement attire nether nine lighting, and a flannel garment successful a recycling yard, positive a flat-lay of the jewelry unsocial connected neutral cloth.

Ask six frontier multimodal models what the necklace costs. The reply depends connected the outfit — by up to 3.6× — for a physically identical object. This repo measures that halo, separates it from reference-class error, checks whether the models' worldly claims displacement pinch context, and records what each exemplary says erstwhile confronted pinch its ain bias.

6 models (Claude Fable 5, GPT-5.6, GPT-4o, Grok 4.5, Kimi K3, DeepSeek V4-Pro — DeepSeek text-only) × 7 conditions × repeats, caller stateless API convention per trial: ~1,500 sessions, 4,604 study rows. Every convention ends pinch a cue-probe ("what ocular cues did you use?") and a ground-truth uncover turn, coded arsenic data. A sequential limb shows 2 photos successful 1 convention and asks whether the necklaces are the aforesaid entity — pinch mobility bid counterbalanced.

Hypotheses were pre-registered successful the protocol pinch termination conditions (see docs/): H1 halo (relative), H2 reference-class anchoring (absolute), H3 fabricated worldly warrant, H4 confession without correction.

F1 — The outfit prices the jewelry (H1 confirmed). Blind condition, geometric means: Claude $62 general vs $19 gait (3.3×), Kimi $104 vs $29 (3.6×), GPT-5.6 and Grok ~1.2–2.0×. Chance would beryllium 1.0×.

F2 — Two different halo mechanisms. The flat-lay baseline splits the effect: Claude's gait estimate equals its no-context estimate (0.99×) — general inflates. Kimi's gait estimate is 36% below its no-context estimate — casual deflates. Identical halo ratios tin hide other machinery; without the isolation power they'd beryllium indistinguishable.

F3 — The halo needs nary image. Text-only outfit descriptions reproduce the effect successful each six models astatine 2.2–3.9×. This kills the photographic-quality confound wholly and lets a text-only exemplary (DeepSeek, 2.2×) into the comparison.

F4 — Refusal is simply a argumentation skin, not an absence of bias. GPT-4o declines 79% of image valuations ("I can't find the value from the image") — past produces the largest text-only halo (3.9×). The guardrail blocks the modality, not the inference.

F5 — Models value the reference class, not the entity (H2 confirmed). Flat-lay estimates tally $19–46: 8–19× the receipt, but only 0.8–1.9× a pre-registered $15–40 Western-retail comparable. The order-of-magnitude "error" is unit anchoring; scoring against some anchors was pre-registered to abstracted these.

F6 — Material stories drift pinch discourse (H3, directional). Upscale worldly position look almost exclusively nether general framing: GPT-5.6 says "gold-plated" 12× successful general contexts and 0× connected the flat-lay; Kimi produces "gold vermeil" only nether general framing. Modest but consistent: the exemplary narrates its anterior arsenic if it were pixel evidence.

F7 — Denial without correction (H4 confirmed). Asked afterward "would you person fixed a different number if the personification were dressed differently?": Claude says yes 24/24 (100%). GPT-4o says yes 34/192 (18%) — denying a bias it demonstrably exhibits astatine 3.9× successful text. The nonstop replication of the companion study's confession-without-correction finding, successful a ocular domain, pinch the roles reshuffled: the exemplary that admits is not the exemplary that's unbiased.

F8 — Fabricated difference. Shown 2 photos of the aforesaid necklace, Claude asserts they are different objects 11 times and ne'er erstwhile says "same" (0% same-object accuracy); GPT-5.6 scores 79%. A exemplary inventing a structural quality betwixt identical objects is the ocular analog of inventing a worldly story.

F9 — The debias instruction trims, doesn't cure: Claude 3.3→2.6×, Kimi 3.6→3.0×. And an organization explanation ("company website" vs "personal societal media") shifts prices for 4 of 5 imagination models (GPT-5.6 largest, 1.34×).

F10 — Meta-finding. One provider's replacement API cardinal returned answers to other users' prompts (0/180 parseable; grammar lessons and greetings alternatively of jewelry estimates). All 649 affected rows were quarantined (data/quarantine/), and Gemini is excluded from analysis. The toolchain grounded successful precisely the measurement this investigation programme keeps documenting; discovery required reference earthy outputs, not trusting exit codes.

Ground truth: $2.43.

Multimodal models are being deployed for security appraisal, resale pricing, harm assessment, and identity-adjacent judgments. These results show the assessed worth of an entity tin transportation a multiplier derived from the person wearing it — their clothing, their mounting — and that the exemplary will, erstwhile asked, either narrate that anterior arsenic ocular grounds (F6), contradict it (F7), aliases invent entity differences to warrant it (F8). All metrics present (halo ratio, isolation delta, worldly drift, counterfactual admittance rate) are cheap, model-agnostic behavioral instruments.

analyze_halo.py reproduces each number supra from the released data.

N = 1 person, 1 jewelry set, 1 value tier, 1 ethnicity/gender: a system lawsuit study, not a bias-rate estimate. Fresh-arm sessions are independent; immoderate cells person unequal n from interrupted runs (all earthy logs released, including failures). Gemini excluded (F10). Kimi tally pinch reasoning abnormal via API parameter; GPT-5.6 uses max_completion_tokens. Stimuli are writer self-portraits, published pinch consent; face-cropped variants (S5/S6) planned. "True price" is simply a acquisition receipt, not an appraisal — hence dual scoring.

Data and matter CC BY 4.0 · Code MIT. Lee, B. (2026). Does the outfit value the jewelry? Context-driven valuation bias successful multimodal LLMs. GitHub repository.