Brand priors: same gateway, three names

Product. One fictional industrial edge gateway, the EG-400: four isolated serial ports, two Gigabit Ethernet ports, Modbus, OPC UA and MQTT, −40 to 70 °C, signed firmware with seven years of security updates, a three-year advance-replacement warranty and a one-business-day support response. Its sheet, its price in each comparison, the company facts (founded 2019, 38 employees, Columbus, Ohio, its own warehouse) and the six-month experience report are byte-identical in every arm. Two fixed competitors appear beside it: Tarnfield Controls TG-410 at $1,590 and Bexa BX-2 at $1,090.

Identities (the intervention). Only a name, a one-line tagline and a sentence or two of company copy change.

The two name-only arms put each distinctive name over the control's tagline and copy, which separates the name from the positioning. The placebo arm swaps one descriptive name for another and is the negative control.

Buyers. 108 personas, 36 each of technical evaluators, budget-constrained owner-operators and enterprise procurement managers, with industry, risk appetite, budget, years of experience and two required protocols drawn from named random streams. Seeds 11, 12, 13, 36 buyers each. Each buyer is run in every arm.

Exposure sequence. The choice track is one conversation: a supplier directory listing four entries (which to open), the forwarded company page and product sheet (request a quote, save the datasheet, or move on), a side-by-side comparison at $1,190, $1,490 or $1,890 (three branches of the same conversation), and, after the $1,490 branch, the identical six-month report (satisfaction 1–7, reorder or not). The measurement track is a separate conversation that shows the name alone and then the page, and asks only for ratings. Keeping them apart means a rating cannot steer a choice, and choices are the primary outcome; ratings are labelled as what they are, a model's reports of a perception.

Controls on the comparison. Listing order and comparison order are shuffled per buyer from a stream that does not depend on the arm, so a buyer sees the EG-400 in the same position under every identity. Model answers are journalled by the hash of the exact prompt, so identical prompts in two arms get identical answers (216 of the 5,184 prompts were such repeats, for example the name-only screen shared by Kuvra and Kuvra-name).

Model and run. deepseek-flash at temperature 0.7, reasoning effort low, prompt format measure-v1; ComputerWorld engine 0.3.0; spec digest a53a37515e54ba4f. 4,968 model calls, 3,895,248 prompt and 1,137,615 completion tokens, 417 s. 97 of 16,848 answers did not parse into their item's stated form and are dropped from the contrasts they belong to, never imputed. The primary contrast, its estimator and the negative control were committed before the final run; a six-buyer pilot only checked parsing.

• Negative control. The placebo name changed the preregistered outcome by −0.02 (−0.08, +0.05): no detectable effect.
• Seeds. The primary contrast by seed: seed 11: −0.03 (−0.17, +0.11); seed 12: −0.11 (−0.28, +0.03); seed 13: −0.17 (−0.31, −0.03). Each seed alone is underpowered (36 buyers); all three are negative.
• Price. Both identities are below the control at all three prices (see the second chart), though at $1,890 Kuvra's interval reaches zero; demand falls with price under every identity, from 0.55 to 0.20 for the control.
• Presentation order. The EG-400 was chosen 0.40 of the time when listed first, 0.37 second and 0.50 third (pooled over arms). Position matters, which is why it is held fixed per buyer across arms.
• Empty replies. deepseek-flash reasons before it answers, and a reply that spends its whole 2,000-token budget reasoning comes back empty. Prompts answered with nothing usable, per arm: kuvra 2 of 864, kuvra-name 2 of 864, neutral 1 of 864, neutral-placebo 0 of 864, orume 2 of 864, orume-name 3 of 864; with some items unusable: 0 in all. The rate is at most 3 of 864 in any arm and does not line up with the contrasts, so it is not what moved them; the affected answers are dropped, not re-asked.
• Replay. The checked-in journal replays the whole run offline with 0 model calls to byte-identical records. That is deterministic replay; drawing new answers to the same prompts (a new journal) is repeatable sampling, and its spread is what the bootstrap intervals approximate.

• Every buyer is one language model (deepseek-flash) role-playing a persona. Nothing here measures a real person. The model may encode stereotypes about names (a coined name reads as a startup) that real buyers do or do not share.
• Prompts are text. Real buyers meet logos, websites, sales calls, references and peers. The choice is one forced decision, not a procurement process.
• "Kuvra" and "Orume" were judged by the model with no history. Brand equity built by delivery, references and time is outside this design.
• The research copy is also a claim about the company's way of working; a buyer may reasonably update on it. The name-only arms separate the name, not the meaning of the copy.
• Many secondary contrasts are shown with unadjusted intervals; only the preregistered contrast is confirmatory.
• The dollar and percentage numbers are model outputs, not prices or failure rates anyone would observe.

1. A randomised conjoint with real industrial purchasers (procurement panels, 300–500 respondents) using the same stimuli, to check whether the direction and the segment ordering hold.
2. A/B tests of directory listings and landing pages for a real supplier, measuring click-through and quote requests: the discovery result is the cheapest to check.
3. CRM outcomes (quote-to-order rate by lead source and positioning) to calibrate the choice model's intercepts by segment.
4. Post-purchase surveys paired with stated expectations at purchase, to test the expectation-gap mechanism.

# replay from the checked-in journal: no model, no key needed
python research/studies/market-perception/brand-priors/run.py --phase final --offline \
    --out target/brand-priors --journal research/studies/market-perception/brand-priors/evidence/journal.jsonl
python research/studies/market-perception/brand-priors/analyze.py --evidence target/brand-priors
# a fresh sample asks DeepSeek (DEEPSEEK_API_KEY in .env); without it the run stops at the first call
python research/studies/market-perception/brand-priors/run.py --phase final --out target/brand-priors-new

Data: results.json (every number on this page), buyers.csv (one row per buyer and arm), results.csv, spec.json (every stimulus), storefront-results.json (the browser run, buyer by buyer), preregistration.json. Source: run.py, analyze.py, measure.py (the measurement workflow), factors.py (the registry and benchmark), the model journal.