# Additive exploratory subgroup uncertainty and spread

Owner request on November4,2026 after observing results: uncertainty intervals for the four displayed category-by-starter averages and an interactive individual-outcome view. This is posthoc, not a revised preregistered primary analysis. External budget zero; saved local data only. No retries, collection, changes to frozen primary, new subgroup selection or publication.

Input: accepted PRIMARY-v1.json in main-design/analysis-combined-after-audit-v1, including all960 scheduled outcomes. Verify its prior review hash 2c9e3f108e8b974b90e3c6257a6b6b2704732387e00f3064fa11c15fe3f92d46. Check SECONDARY-RESULTS-v1.json hash 8dd5327ff11d12671792202805168d539e0f813be59989ddd52335fdd58657c2.

Four fixed cells: tradeoffs/buyer-first, tradeoffs/seller-first, narrow-zone/buyer-first, narrow-zone/seller-first; both economic orientations pooled. Each contains30 independent families and6 assigned negotiations per family. Report completed-episode mean gap sum(gap)/completed count to match existing chart. Its target is the ratio of expected completed-gap totals to expected completed counts under the study completion process, not the full scheduled-population mean with unknown outcomes filled in.

Compute approximate pointwise95% whole-family percentile bootstrap confidence intervals:20000 resamples with replacement of30 family IDs within each category; all rows belonging to a selected family travel together, including both starters/orientations/repetitions and missingness indicators. Use same draws for both starters within a category; numpy PCG64 seed2026110401; sorted families and fixed category order tradeoffs then narrow_zone. Compute ratio of total completed gaps to completed counts per draw. Percentiles2.5/97.5 using linear interpolation. Stop if any resample has zero completed count, malformed join, inconsistent totals or source hash mismatch. No resampling individual negotiations independently. Approximate coverage assumes independent representative scenario families and bootstrap regularity;30families small, no formal coverage guarantee. Pointwise intervals are not simultaneous across four rows, not selection-adjusted and not a first-mover-effect contrast test. No p-values or claims based on whether a row excludes zero.

Missing outcomes are unknown, never zero/ties or successful walkaways. Retain counts per cell. To expose the separate missingness issue, compute full-schedule legal sample bounds per cell as (sum observed gaps + sum missing lo)/180 and corresponding hi. These are fixed-sample identification bounds, not95%CIs and not the frozen primary. Do not replace completed-mean intervals with these or call the completed intervals bounds on unobserved outcomes. Do not give a complete-case replacement primary estimate.

Spread: for all completed negotiations in a selected cell show exact integer buyer-minus-seller gap histogram/dot bins with fixed5point increments, and counts buyer higher/seller higher/equal; unknown count separately. Equal includes walkaways and equal-gain agreements; expose their separate counts. Include selected gap's individual episode IDs, buyer/seller gains, economic orientation and completion type on selection or in a details table. Use stable shared x-axis across four cells, include zero, no smoothing. Default tradeoffs/buyer-first but allfour choices visible; intervals only for fixed full cells (no unreviewed filtered CIs). Label 'earned more', never wins/losses. Agreement may benefit both parties even when unequal.

Success: all4means exactly reproduce previous unrounded summaries, counts30families/180assigned percell, all720feasible outcomes accounted for including8unknown; independent methods and numerical review accepts; UI clearly distinguishes sampling uncertainty, individual spread, and missing-outcome uncertainty. On material objection repair within bounded three-loop protocol. Preserve source files/earlier article versions and all primary outputs.