# Concert ticket study — methods and evidence

Start with [the concert ticket study](index.html) for the setup, results and every conversation. [RESULTS.md](RESULTS.md) is the editable summary. The files below contain the data and checks behind the results.

The reader-facing sample is 40 completed outcomes, 20 per pairing. The full assignment log is 42 attempts, including the two original technical unknowns. The two new runs were requested after the original results were seen. Original and extension IDs remain distinguishable.

## Reproduce

1. Download or extract this study folder, preserving its subfolders.
2. With Python 3.10 or later, run `python reproduce.py` from this folder. It verifies hashes, exact payoffs, denominators, opening balance, and summary arithmetic using only the standard library.
3. With Node.js available, run `node source/replay_all.mjs`. This independently reconstructs all recorded public exchanges through the frozen scorer and verifies the final mutual acceptance, prompt contexts, and utilities for all 40 completed agreements. Unknowns remain unknown.

Neither command contacts a provider, loads credentials, runs new simulations or changes the source files. The included candidate is the scientific scorer/prompt source, not the original financial/API runner. The original frozen registration and dated extension note are separate.

## Reading guide

- `index.html`: results, opening-order table, links to all transcripts.
- `transcripts/`: exact public responses and role instructions; unknowns show their available partial history.
- `data/`: full-precision source data and two-decimal reader tables.
- `registration/`: scientific inclusion and model/configuration records.
- `source/`: frozen candidate and offline replay verifier.
- `AUDIT.json`, `QA.json`, `MANIFEST.json`: validation and integrity records.

This package contains public agent messages and synthetic owner prompts, not hidden chain-of-thought. Some usage metadata records thinking-token counts; no hidden reasoning text is included. Original quotes are not rounded or rewritten. All monetary-looking contract terms are scenario inputs; outcomes are synthetic planning points.

The essay is maintained separately. Four unrelated whole-day spending, exposure and operational cap fields are omitted from the public RESULTS_FULL_PRECISION.json projection. All scientific statistics, source trials and study-specific costs are retained; the source ZIP is preserved privately.

Provider response identifiers (`public.id`) are omitted from the public call-record projection. Exact scientific messages, request contexts, actions, model/configuration records and usage remain unchanged. The original archive is retained privately.