Evaluator proxy for acceptance; not a real-customer survey or success rate. Mean 0–1 score across the pilot
Mean judged quality
Judged quality is separate from persona acceptance. Mean 0–1 score across the pilot
Identity trust · S1
Evaluator proxy for acceptance; not a real-customer survey or success rate. Mean persona acceptance for the named scenario across eight personas and five turns
Memory/action realism · S2
Evaluator proxy for acceptance; not a real-customer survey or success rate. Mean persona acceptance for the named scenario across eight personas and five turns
Place fidelity · S3
Evaluator proxy for acceptance; not a real-customer survey or success rate. Mean persona acceptance for the named scenario across eight personas and five turns
Text trust · S4
Evaluator proxy for acceptance; not a real-customer survey or success rate. Mean persona acceptance for the named scenario across eight personas and five turns
Commerce trust · S5
Evaluator proxy for acceptance; not a real-customer survey or success rate. Mean persona acceptance for the named scenario across eight personas and five turns
Eight simulated personas and one pilot, not a general model benchmark.
Trust/frustration are derived state; mean acceptance can conceal high-risk failures.
Only the two saved models are compared; three judge families provide separate evaluations.
THE SAVED OBSERVATIONS
What the evidence shows.
GPT
GPT Image 2
openai / gpt-image-2
0.478Mean persona acceptance
Broader acceptance coverage in this pilot; strongest in text and commerce relative to the other model.
Observed answer
Scenario changes the read
GPT Image 2 leads four of five scenario means; Nano Banana leads identity trust.
Study scope
Pilot scope
Five turns across eight personas and five scenarios, not an overall deployment decision.
NB
Nano Banana
google_vertex / gemini-3-pro-image-preview
0.452Mean persona acceptance
Leads identity trust in this pilot; the aggregate result does not describe every photo task.
Observed answer
Scenario changes the read
GPT Image 2 leads four of five scenario means; Nano Banana leads identity trust.
Study scope
Pilot scope
Five turns across eight personas and five scenarios, not an overall deployment decision.
METHOD & SCOPE
How this study was run.
Each persona repeatedly edits the same seed photo from its own feedback and memory; eighty model journeys create 400 generated turns.
Two models were actually saved: GPT Image 2 (gpt-image-2) and Nano Banana (gemini-3-pro-image-preview). This edition uses those saved outputs and retains three independent judge families.
Acceptance is the calibrated 0–1 evaluator reaction. Judged quality is separate. Trust and frustration are derived longitudinal states, not survey answers.
GPT Image 2 leads four of five scenario averages. An earlier session-win headline was removed because its win convention is unreconciled with the tie-aware scenario cells: twenty GPT-favored, six Nano-favored and fourteen ties. Scenario means and tie-aware cells are separate views; overall session wins are not ranked here.
LIMITATIONS
What this does not establish.
Eight synthetic personas, five scenarios and one saved pilot provide directional debugging evidence.
Results do not establish an industry-wide model ranking or release decision.
Raw deltas and relative lifts use underlying precision; do not recompute them from rounded displayed means. The earlier headline session-win count has not been reconciled with detailed tie-aware cells and is omitted as a ranking metric.
Identity, text and commerce trust differ; one aggregate must not hide scenario risk.
FOLLOW THE SOURCE
Study material.
Explore the available sources behind these findings.