# Concert ticket study: the original 40 attempts

This folder records the original 40-attempt collection. The current results, with 40 completed negotiations, are in [the expanded study folder](../c3-expanded-40-completed-20261028/index.html).

Start with index.html for this collection’s results and conversations, then RESULTS.md for the detailed tables. I kept all 40 assigned IDs, including the two interrupted runs. I used one fixed ticket scenario, continuous contract terms and deterministic scoring; the prompts and request settings are recorded alongside the data. Exact requested/returned aliases and per-call dates are recorded; immutable provider weight snapshots are not available.

Units: combined synthetic utility gain above two 20-point no-deal fallbacks, maximum 90. These are not dollars, net-of-compute gains or actual welfare. Unknown numeric fields are blank, never zero-imputed. Known no-deals would be recorded as zero gain. Agreement counts, missingness, mean/median/SD/range, owner gains and opener strata are separated. Missing-outcome identification bounds are not confidence intervals. The prespecified analysis does not report a sampling interval when an assigned outcome is unknown.

Reproduction: run `python reproduce.py` in this directory (Python standard library only). It recalculates utilities from negotiated continuous terms, verifies per-assignment identities and reproduces pairing means, medians, SD, missing bounds and contrasts. candidate.mjs is the unchanged original renderer/scorer; prespecified_analysis.py is the archived pure-analysis implementation. The delivered request and public-message corpora expose the actual synthetic inputs and messages; hashes and audit report support the source provenance. No SDK, network, account access or API key is needed for reproduction.

Transcript coverage: all assigned negotiations, with every available public response and initial role instruction. Interrupted calls have no invented output. Failed raw responses or hidden reasoning are not supplied. Public-role prompts are synthetic research, not actual owners' private information. API headers, credentials, account settings and billing integration are excluded. The author directs the research; Claude/Codex assisted research software and deterministic verification. The scoped recovery was implemented locally; the frozen economic protocol was unchanged.

Data dictionary: per-assignment.csv contains original run ID, pairing, nominal assignment, opener, economic/technical status, known flag, agreement flag, gain and utility points, three agreed terms, format counts and an HTML transcript reference. Pairing/opening summary CSVs describe known outcomes and separate all-assigned bounds. per-call-usage.csv retains scientific call IDs, actual token usage/charges and unresolved bounds without credential or provider-token details. All monetary accounting is USD; economic scores are synthetic points. COLLECTION_ACCOUNTING.json preserves settled and outstanding totals, not assumed charges for uncertain requests. Historical snapshots remain in the archive and are dated.

## Reproduction entrypoint

Run `python reproduce.py`. `prespecified_analysis.py` is archived source, not a runnable entrypoint in this supplement; its original direct-run path expects `UNIT_RESULTS.json`, which is not shipped. No credentials or inference are needed by the supported command.