PRICE-02: the price-framing replication with Sol and Terra

What was asked

The scenario and the order of the six participant turns were retained from PRICE-01. A new infection is expected to affect 900 people in a region where the participant lives. A pharmacy will sell Treatment A, a one-season preventive course. Similar courses sell for $25. If all 900 people take it, the gain frame says 300 will be protected; the loss frame says 600 will be left unprotected. These describe the same outcome.

The two-by-two design crosses Frame (gain, loss) with Model (GPT-6 Sol, GPT-5.6 Terra). The primary outcome is the maximum price for one course for oneself, in whole US dollars from 0 to 60. The secondary outcome is likelihood of buying at $25, from 1 (not at all likely) to 7 (very likely). A short price rationale and protected/unprotected count checks complete the instrument.

Two samples each scheduled 20 sessions per condition. The 40-session measurement pilot and four-session execution pilot are excluded from every main-study summary on this page. Complete six-turn sessions are eligible; seven protocol failures remain missing and were not replaced. This whole-session rule is stricter than PRICE-01’s per-outcome availability.

Numerical results

Gain minus loss is $10.28 for Sol and $9.58 for Terra. Both frame comparisons are significant in the planned exact tests: Holm-adjusted p = 2.700 × 10⁻⁹ for Sol and 1.823 × 10⁻⁷ for Terra. The sample-blocked price interaction is not significant (p = .732), so the difference between their framing effects is not established.

Likelihood to buy at $25

Sol’s mean likelihood was 4.90 under gain and 2.56 under loss, a difference of 2.33 points. Terra’s means were 4.70 and 1.47, a difference of 3.23 points. Both gain-over-loss comparisons are significant: Holm-adjusted p = 4.923 × 10⁻⁹ for Sol and 1.586 × 10⁻⁷ for Terra. Sol and Terra do not differ significantly under gain (p = .691), but do under loss (p = .00356). These ratings concern a stated intention in this scenario, not an observed purchase.

What the rationales say

The five original rationale codes are retained. The final coding package contains 154 captured rationales: 153 from eligible sessions and one for audit only. Final Claude Sonnet 5 coding has not run; there are no code counts, code charts or agreement estimates to report. A pending code is not a zero.

Fresh calibration and both pilot coding stages were completed. The final transfer remains on hold pending explicit approval for the prepared rationale payload. The page and its downloads contain only numerical summaries, not rationale text.

Manipulation checks

Every eligible complete session returned 300 protected and 600 unprotected. Both checks were correct in 153 of 153 observed eligible answers. The seven protocol failures remain missing against the scheduled denominator of 160.

The two samples

The two samples are blocks within the same planned collection. They are separate groups of sessions, not repeated measurements or independent studies. The table and charts retain each sample’s actual count. No sample-specific hypothesis verdict is supplied.

Inferential statistics

These are the frozen plan’s numerical tests on 153 eligible sessions. All four cells meet the 90% valid/scheduled threshold, and all 160 scheduled sessions are accounted for. The code outcomes and final signed study record remain pending; these tables are a partial numerical analysis.

Each outcome has four two-sided exact permutation tests, permuting within the two sample blocks. Their statistic is the absolute sum of the two within-sample mean differences. The displayed pooled mean difference is a descriptive effect size and need not equal half that statistic when counts differ. Holm adjustment includes all four comparisons within each outcome, separately for price and likelihood. No comparison was dropped from either family.

Differences are in US dollars for price and rating points for likelihood. The exact test compares absolute values of the signed sample sum; signs are retained here to make direction readable. The permutation method does not supply a conventional standard error, t statistic, degrees of freedom or contrast confidence interval, so none is invented. Full-precision p values and exact allocation totals are in the numerical JSON download.

The declared price model includes a sample intercept, Frame, Model and their interaction. Gain and Sol are coded +0.5; loss and Terra −0.5. The three Wald t tests use 148 residual degrees of freedom. Their p values are unadjusted, as declared. The coefficient intervals are supplementary, unadjusted 95% Wald t intervals.

The price interaction is b = 0.680, t(148) = 0.343, p = .732. This provides no significant evidence that the two models’ price-framing effects differ; it does not establish equivalence. No omnibus model for likelihood was registered, and none is added here.

All four numerical hypotheses meet their frozen criterion: a significant effect in the predicted direction at Holm-adjusted α = .05. This is the mechanical numerical result, not a claim that the full study has been finalized or admitted into the OpenPsy app. The original Opus-versus-Sonnet interaction hypothesis H1c was not transferred to these models.

Missing-data sensitivity

The original grammar-only sensitivity rule covers invalid, refused and provider-truncated outcomes. None of the eligible numerical answers had those classifications. Its bounds and four decision-rule results therefore equal the observed analysis. Native whole-session failures are a different classification and must not be relabeled as provider truncation.

The separate native sensitivity includes all seven unavailable scheduled outcomes. It evaluates all missing values at the scale minimum, all at the maximum, and the assignment against the gain-over-loss prediction: missing gain outcomes at the minimum and missing loss outcomes at the maximum. Every scenario reruns both complete four-comparison Holm families. These are endpoint assumptions, not imputed observations or confidence intervals.

All four criteria remain met in these endpoint scenarios. The least-favourable Terra price comparison is closest to the threshold, with Holm-adjusted p = .039. These bounds address the seven missing numerical outcomes under the stated range assumptions; they do not address route differences, unverified provider controls or other sources of bias.

What these results support

The four retained gain-over-loss hypotheses meet their numerical decision rule for both maximum price and likelihood to buy. That statement is restricted to the collected sessions, wording and native route. The price interaction is not significant; a non-significant interaction does not establish equal effects.

The missing-data endpoint scenarios retain all four numerical criteria, although the least-favourable Terra price comparison is close to the threshold at Holm-adjusted p = .039. No rationale-based explanation is supplied until the planned coding is complete.

The native campaign remains outside the completed standard Library/app workflow. This partial numerical report does not substitute for completed coding, a fresh signed independent analysis review or final study recording.

How the run went

The 40-session measurement pilot and four-session execution pilot completed, with fresh calibration, collection review and coding. The fixed main schedule then processed all 160 sessions. There are 153 eligible complete sessions and seven known protocol failures, with no unknown or unattempted main-study sessions and no replacements. Protocol rejection is not by itself evidence of a scientific model refusal.

This was a source-pinned native successor with fresh plan, execution and independent-review records. It was not the unchanged original OpenPsy app import-and-run path. The differences below remain part of its interpretation and cannot be repaired by renaming or reformatting the page.

How this relates to PRICE-01

PRICE-01 is the original Claude Opus 5 / Claude Sonnet 5.5 study, run on 6 October 2026. It has a saved OpenPsy study, recorded analysis and a published results page. Its price outcome used 157 available responses, its likelihood outcome 156, and its coded rationales 154; those denominators are specific to the original study.

PRICE-02 is the review label for the GPT-6 Sol / GPT-5.6 Terra replication. The retained campaign identifier is PRICE01-SOL-TERRA-20261007. Historical paths, signed records and source hashes continue to use that original identifier. This presentation mapping does not create a new registered app study or retroactively change collection.

Read the two studies separately: participant models, route controls, context, concurrency and eligibility differ. A comparison between the pages is descriptive and does not isolate a model or provider effect.

Tables and downloads

Review and application status

This page presents the partial numerical results for review on OpenPsy. PRICE-01 remains the separate original study. Rationale coding, final signed recording and standard app integration remain pending.

The corrected page reuses the established site’s typography, section structure, two-by-two tables, model colours, figure grammar and accessible bar/line controls. It includes the planned scalar tests and their uncertainty/sensitivity tables. It does not claim that the native study has completed coding, final recording or the canonical app workflow.