Do gain vs. loss frames change what models choose?

Framing effects are one of the most replicated findings in human judgment research: people accept more risk to avoid a loss than to secure an equivalent gain. The question for machine psychology is whether language models inherit that asymmetry from human text — and whether it is uniform across models.

The study starts with two registered hypotheses in the Hypotheses stage:

Both predictions pass the platform's feasibility checklist: the manipulation is pure text, the outcome is readable from a single response, and the effect is meaningful as a distribution over many replications.

You could cross framing (gain, loss) with scenario in a plain full factorial. But the sharpest test of a framing effect is a paired one: the same scenario under both frames, compared within the pair. AIScholar's matched-pair crossing strategy builds exactly that grid and unlocks a decoupling-gap analysis later.

The expected-output declaration matters: by declaring a two-field response up front (a categorical choice and a freeform justification), extraction and validation are automatic at collection time, and the invalid-rate tracker will flag any condition where models stop following the format.

In AIScholar, factor levels are analytical labels; the actual wording lives in the snippet library. One template serves every condition, and snippet versions are bound to factor levels:

{{scenario_setup}}

Two response programs have been proposed.

{{option_description}}

Which program do you recommend? Answer with exactly
"Program A" or "Program B" on the first line, then give
a one-paragraph justification.

The variation grid shows all eight resolved prompts before anything runs, so you can verify that paired prompts differ only in the frame — the design's core validity requirement.

Execution runs against your own OpenRouter key with orchestrated concurrency and retries. The sensible order of operations:

The cost estimator prices the full 960-trial run before you commit, and pilot trials are excluded from analysis automatically.

The categorical choice extracts directly from the first line — no judge needed. The justification field goes through an LLM-as-a-judge coding pass with a two-variable scheme:

A 10% sample gets a second judge model plus human review; inter-rater agreement (Krippendorff's α) is computed before the machine codes are trusted for the rest.

The primary outcome is binary (risky option chosen or not), so the main test is a GLMM: framing as a fixed effect, scenario as a random intercept, and a framing × model interaction for H2. The matched-pair design then enables the decoupling-gap analysis — does the stance flip while the justification profile stays put, or do they move together?

A pattern like the illustration would support H1 (all shifts positive) and H2 (Model B's shift visibly smaller — testable via the interaction term). The decoupling gap adds the machine-psychology twist: if justification types barely change while choices flip, the models are re-rationalizing rather than re-reasoning — a finding you could not see without the paired design.

The write-up assistant drafts Method and Results sections directly from the design, execution metadata, and analysis outputs — including the exact model identifiers, parameters, and replication counts a reviewer will ask about. Citations to the framing literature come from the citation library, and the reproducibility package bundles the full trial-level dataset with the Method Specification Prompt so another lab can re-run the study verbatim.

Single-turn text trials measure stated choices, not decisions with stakes. The study cannot say a model "experiences" loss aversion — only that its response distribution shifts with the frame. Effects can be sensitive to prompt wording, which is why a natural follow-up is the prompt-sensitivity study: re-run the winning conditions across paraphrase variants and check whether the framing shift survives.

• Matched-pair design
• Decoupling gap
• GLMM