Concert ticket study: the original 40 attempts

Results at a glance

Summary values are shown to two decimal places. Downloads retain full precision.

I ran 40 concert-ticket negotiations: 20 between Haiku pairs and 20 between Opus pairs. Thirty-eight completed. Two runs were interrupted, so their outcomes are unknown. Each pairing has 19 completed agreements and one unknown; both interrupted runs were buyer-first. There were no known no-deals. I report the interrupted outcomes as unknown, not zero.

Source: pairing-summary.csv. No known outcome left either owner below fallback, and no known run used a format repair.

Difference and uncertainty

Among the 38 completed negotiations in this collection, OO mean joint gain is 25.44 points higher than HH. Because one assigned outcome in each pairing is unknown, the prespecified rule reports no sampling confidence interval. Instead, the bounds let each unknown take any value from 0 to 90 points. Over all 40 assignments, the OO minus HH difference is then between 19.67 and 28.67 points. These are missing-outcome bounds, not confidence intervals. The sample SD describes spread among known runs; it is not a sampling interval.

Who gained: buyer and seller, by opener

Source: opener-summary.csv and pairing-summary.csv.

Buyers gained about the same in both pairings. Most of the joint difference is on the seller side. Within HH, buyers received more gain than sellers under either opener, most clearly when the buyer opened, while OO splits were close to even. By opener, the known-case OO minus HH difference is 25.84 points with a buyer opener and 25.08 with a seller opener. These role and opener comparisons are exploratory: they carry no significance test, do not resolve missingness and say nothing about motives.

What this does and does not show

• I used one constructed ticket scenario and a fresh context for each run. These results describe that setup; they do not establish how the models would negotiate in other markets.
• I did not test human negotiators, general intelligence or improvement over time. This study does not establish a general model ranking or a causal effect of capability.
• “Net value creation” means combined synthetic owner gain above both no-deal fallbacks. It is not dollars, net compute cost or social welfare.
• Models: claude-haiku-5-5 and claude-opus-5-5, adaptive thinking at low effort, a 4,096-token output allowance, standard_only tier. Collected 27–28 October 2026. No immutable dated weight snapshot was returned.

Data and reproduction

RESULTS.md · per-assignment.csv (all 40 run IDs, unknown values blank) · REGISTRATION_SCIENTIFIC.json · AUDIT_SUMMARY.json · TRANSCRIPTS.jsonl (exact delivered public texts) · README.md

The supported reproduction command is python reproduce.py; see README.md. prespecified_analysis.py is archived source only, not an entrypoint. A deterministic replay and accounting audit passed 1,983 checks, including all 38 known final offers and utilities.