# Concert ticket study: results

I compared 20 completed negotiations between Haiku pairs with 20 between Opus pairs. Each pairing includes 10 buyer-opening and 10 seller-opening negotiations.

| Pairing | Completed | Mean /90 | Median | SD | Buyer gain | Seller gain | Efficiency |
|---|---:|---:|---:|---:|---:|---:|---:|
| HH | 20 | 57.74 | 57.84 | 3.51 | 41.38 | 16.35 | 64.15% |
| OO | 20 | 83.59 | 83.91 | 2.49 | 41.80 | 41.79 | 92.87% |

Net value creation means the two owners' combined synthetic planning-point gains above their no-deal alternatives. Each owner gets 20 points without a deal; the most the two can gain together is 90. This is not money, net of compute costs, or measured social welfare.

Opus pairs averaged 25.85 more points. The lead shows up under both openers. All 40 known outcomes were agreements. No known agreement fell below either owner's fallback.

## Collection history and uncertainty

After seeing the initial results, I added one new buyer-opening negotiation per pairing. These were new runs, not replays or replacements. Across the full history, 42 assignments produced 40 known outcomes and two interrupted unknowns. The original 40-assignment cohort is still 38 known / two unknown; its raw records have not changed. The full assigned set has 11 buyer-opening and 10 seller-opening assignments per pairing. These unknowns are not scored as disagreements or zero gains.

This extension was chosen after viewing results. It is a descriptive comparison in one constructed scenario, not a confirmatory study, a general model ranking, a human comparison, or evidence of improvement over time. The missing-outcome bounds for the all-assigned Opus-minus-Haiku mean difference are 20.33 to 28.90 points. Those bounds cover the possible missing values; they are not a confidence interval. No new sampling interval is claimed.

The exact aliases were `claude-haiku-5-5` and `claude-opus-5-5`. Both used adaptive thinking at low effort, standard-only service, a 4,096-token output allowance and a 12,000-input-token admission bound. Requests and returned aliases were checked; these aliases do not establish an immutable model snapshot across dates. Collection began October 27 and the two-run extension finished October 28, 2026.

## What they negotiated

A fictional company bought a block of 40 concert tickets from a ticket agency. Both sides could negotiate the same continuous price ($60–$100 per ticket), cancellation fee (0–100%), and advance notice for name changes (0–96 hours), with at most two decimal places. Each side saw its own private utility function and the common rules. There were no scorecards and no added information about the counterpart's utility. The buyer weighted price most; the seller weighted advance notice most. That difference in priorities created opportunities for mutually beneficial trades.

There were at most 12 alternating turns, plus a final response to a standing last-turn offer. Agreement required a valid acceptance by the other side. The frozen format contract allowed one repair of an invalid response; economic no-deals, format failures and technical unknowns are distinct outcomes. No failure or unknown was silently excluded from the assignment log.

## The two additional outcomes

| Pairing | Price per ticket | Cancellation fee | Name-change notice | Buyer gain | Seller gain | Combined gain |
|---|---:|---:|---:|---:|---:|---:|
| HH | $75.50 | 14.00% | 16.00 hours | 53.09 | 0.28 | 53.38 |
| OO | $64.00 | 38.00% | 96.00 hours | 43.60 | 43.40 | 87.00 |

These two follow-ups were selected by missing opening-order strata, not by desired results. They should not be treated as representative anecdotes by themselves. All available public messages, private role instructions and input contexts are included for every assignment.

## Files and reproducibility

- [Browse all 42 assignment transcripts](index.html#transcripts), including partial unknowns.
- [Reader summary CSV](data/reader-summary.csv) and [opening-order summary](data/reader-opener-summary.csv).
- [Per-assignment data](data/per-assignment-data.csv), [exact scored results](data/ASSIGNED_SCORES.json), and [public API call records](data/public-calls.jsonl).
- [Original scientific registration](registration/ORIGINAL40_REGISTRATION_SCIENTIFIC.json) and [dated extension note](registration/EXTENSION2_SCIENTIFIC_NOTE.json).
- [Offline reproduction instructions](README.md#reproduce) and [data dictionary](DATA_DICTIONARY.md).

Reader tables use two decimals. Original public messages, exact fractions, timestamps and machine-readable data retain their recorded precision. Hidden reasoning, credentials, account details, financial integration and raw request headers are excluded.