Research

Data files and what the columns mean

These files are exports from the studies, not new studies. Each CSV is UTF-8 with one negotiation per row, or one group per row for the summaries. Status and denominators are always explicit.

The tables on the study pages are the main reading view. The CSVs are here if you want to check or recompute.

Concert ticket study files

The 40-completed study has its own folder with its own dictionary: DATA_DICTIONARY.md. In short: per-assignment-data.csv has one row per attempt (42), with cohort marking original40 or extension2_post_result and unknown outcomes left blank; reader-summary.csv and reader-opener-summary.csv are the two-decimal tables on the study page; ASSIGNED_SCORES.json keeps exact fractions; public-calls.jsonl has one record per call slot (323 rows, not 323 negotiations).

Ticket pilot outcomes (`c3-outcomes.csv`)

run_id: Identifies the run. A -cont1 suffix means the run completed an original assignment that had never started.
pairing: HH (Haiku–Haiku), HO (Haiku–Opus) or OO (Opus–Opus).
assignment: 1 to 4; codes the original balanced design of agent role and opener.
opener: Which side made the first move: buyer or seller.
buyer_model, seller_model: Requested and returned model aliases for each role, checked against each raw call.
status, known: Distinguish observation status: whether the run ended in an agreement and whether its outcome is known.
price_usd_per_ticket: Agreed price in fictional US dollars per ticket.
cancellation_fee_pct: Agreed cancellation fee in percentage points.
name_change_notice_hours: Agreed hours of advance notice required for a name change.
buyer_utility_points, seller_utility_points: Each side's score for the agreement in synthetic planning points.
buyer_gain_points, seller_gain_points: Utility minus the 20-point no-deal fallback for each side.
buyer_gain_exact, seller_gain_exact: The same gains as exact rational numbers (for example 3669/80), avoiding repeating-decimal ambiguity.
net_value_creation_points: Combined utility minus both fallbacks (40 points in total): the sum of the two gains.
net_value_creation_exact: The same value as an exact rational number.
efficiency_fraction: Net value creation divided by 90, the maximum possible joint gain.
below_fallback: Flags any final owner below 20 points.
phase: Keeps the original smoke or main label. Unchanged smoke assignments were included prospectively.
raw_unit_state: The run system's service classification, preserved as recorded.

Freight study outcomes (`environment-outcomes.csv`)

All dollar fields refer to the entire fictional 26-week season.

run_id: The original episode identifier.
arm: disclosed (certified emissions numbers shown) or undisclosed (numbers not provided).
opener, table_order: Together define the four balanced strata: who sends the first message (buyer_first or seller_first) and the order of table rows (F or R).
status, known: Scientific deal status and whether the outcome is known.
plan: Agreed plan as vehicle-schedule: D or E for diesel or electric; O, N or C for overnight, next-day or two-day consolidated (for example E-C).
price_usd_season: Agreed total season price.
buyer_money_usd, seller_money_usd: Buyer money is plan value minus price; seller money is price minus cost.
joint_money_usd: Buyer money plus seller money.
buyer_gain_vs_fallback_usd, seller_gain_vs_fallback_usd: Money minus the fallback: 3,250 for the buyer, 1,950 for the seller.
emissions_kg_co2e_season: Fictional certified season emissions of the agreed plan in kg CO2e. These are scenario facts, not API compute emissions.
lowest_emissions_package: Flags the electric/two-day package (E-C).
buyer_below_fallback, seller_below_fallback: Floor indicators, using strict less-than against each fallback.
acceptor_below_fallback, proposer_below_fallback: Distinguish whether the side below its fallback accepted the deal or proposed it.
invalid_public_turns: Count of recorded turns with no valid action.
negotiation_turns: Number of negotiation turns in the episode.
service_terminal_used_for_known_deal: Flag: this agreement was taken from an already-known service-terminal record after a later technical fault. The agreement is kept.
postcheck_answered_roles: Technical qualification: how many roles answered the separate postcheck. Missing postchecks do not delete known agreements.
raw_unit_state, raw_scientific_status: Preserve the run system's service outcome independently of the scientific deal status. For example, a row can have status DEAL (a known agreement) while raw_scientific_status reads INCOMPLETE, because the service record's completeness is not the same thing as whether the economic agreement is known.

Summaries and spread

c3-summary.csv: One row per pairing in the pilot, n = 4: counts of assigned runs, known agreements and technical unknowns; means of net value creation, buyer gain, seller gain and efficiency; and the sample SD (n − 1) of net value creation. These are not confidence intervals.
environment-summary.csv: One row per arm with n = 160: counts, lowest-emissions count, mean and sample SD of emissions and of joint money, and buyer and seller below-fallback counts.
environment-contrasts.csv: Disclosed-minus-omitted means for emissions and joint money, with a descriptive stratified standard error computed as sqrt(sum over 4 strata of (1/4)^2 * (s_dis^2/40 + s_und^2/40)), and a units column and uncertainty note.

No p-values, confidence intervals or pooled cross-study effect are produced.

Transcript exports (`*-public-transcripts.jsonl`)

Each line is one complete assigned negotiation, encoded as a JSON object.

run_id: Matches the run_id in the outcomes CSV.
registration_versions, protocol: The registration version(s) and protocol under which the run was conducted.
owner_prompts: The exact private brief for buyer and seller.
configuration: The recorded request configuration: thinking type, effort, maximum output tokens, service tier and temperature setting.
unit_service_status: The raw service state, outcome and scientific status recorded for the unit.
public_turns: One entry per public turn, containing only: turn, role, local call_id, requested_model, returned_model, at_utc timestamp, message, binding action, stop_reason, invalid_raw_output_excluded and message_sha256.
source_ledger, source_table: Locate the untouched original record, relative to the original research root, together with the local call ID. The transcript reader does not display them.

No hidden reasoning, raw reasoning preambles, authorization headers, credentials, provider request IDs or account data are exported. A failed-format turn carries the counterpart-visible no-valid-action placeholder and is not silently rewritten as valid language. Original source files remain unchanged.

Downloads

Concert ticket study

reader-summary.csv: results by pairing.
reader-opener-summary.csv: results by opening order.
per-assignment-data.csv: all 42 attempts; unknown outcomes left blank.
ASSIGNED_SCORES.json: exact scores as fractions.
public-calls.jsonl: recorded public requests and responses.
DATA_DICTIONARY.md: column meanings.

Ticket pilot

c3-summary.csv: per-pairing counts, means and sample SDs.
c3-outcomes.csv: all 12 outcome rows.
c3-public-transcripts.jsonl: exact private briefs and public turns.

Freight emissions study

environment-summary.csv: per-arm counts, means and sample SDs.
environment-contrasts.csv: differences in means with descriptive stratified SE.
environment-outcomes.csv: all 320 outcome rows.
environment-public-transcripts.jsonl: exact private briefs and public turns (large).

Code, registrations and delivered requests are on the code page.