Research

Code and checks

The frozen scripts, the scientific registrations and the requests as delivered, as direct downloads. Paths keep the source package's layout under files/reproducibility/.

The registration exports keep the assignments, model identities, timestamps and source hashes. They are labelled projections of the original records, not byte copies of the operational databases. Billing, account and credential code isn't here.

Checks you can run offline

From research/files/ in a downloaded copy of this site:

python reproducibility/derive_tables.py --check
node reproducibility/verify_original36_replay.mjs

The source package ran these with Python 3.10.11 and Node v24.11.1 using standard libraries only: nothing to install, no API keys, no network, no model calls.
The Python check recomputes the 12 pilot payoff rows and the 320 freight rows, compares them with the published summary CSVs and verifies the public-message hashes. The Node check replays all 36 original ticket assignments against the frozen renderer and parser. Neither runs a model or fills in missing outcomes.
On 28 October 2026 both passed in an isolated copy: 12 pilot outcomes, 320 freight outcomes, 1,430 message hashes and 314 replayed requests across the 36 original assignments.

Concert ticket study (40 completed)

Results, all 42 conversations and the methods are in files/reproducibility/c3-expanded-40-completed-20261028/. From that folder, python reproduce.py verifies hashes, exact payoffs, denominators, opener balance and the summary arithmetic; node source/replay_all.mjs reconstructs every recorded exchange through the frozen scorer and checks the final acceptance and utilities for all 40 completed agreements. The two unknowns stay unknown. Neither command contacts a provider or runs a model.

Both checks passed on 28 October 2026. The replay covers 323 recorded request contexts across 42 attempts, including all 40 completed agreements. The extension audit separately reports 260 checks.

The original 40-attempt collection keeps its own results and python reproduce.py. Its raw records are retained, and its arithmetic check passed in an isolated copy on 28 October 2026.

Frozen request settings

The pilot, the freight study and the concert ticket study all used these settings.

Request settings, all three collections
Setting	Value
max_tokens	4096
thinking	adaptive
output effort	low
service_tier	standard_only
counted input gate	12,000 tokens
temperature	not explicitly set

Top-level scripts

reproducibility/derive_tables.py: Python check that recomputes outcome rows for both studies and compares them with the published summaries.
reproducibility/verify_original36_replay.mjs: Node replay of all 36 original ticket assignments against the frozen renderer and parser.
reproducibility/derive_c3.mjs: Node derivation script for the C3 ticket tables, supplied with the source package.

Ticket pilot: original records (the protocol's C3 condition)

original-c3/registration-scientific.json: complete scientific registration: all 36 original planned units plus six administrative continuation records (not 42 independent observations).
original-c3/prospective-roster-36.json: the prospective 36-case roster.
original-c3/candidate.mjs: unchanged scenario, prompt renderer and scorer.
original-c3/extract.mjs: unchanged action extractor.
original-c3/inclusion-audit.json: audit of the smoke-phase inclusion decision.
original-c3/qualification-receipt-b01.json: block-1 qualification receipt, retained with its contradictory never_analyzed_with_main: true flag. Its delivery version changes only two local path fields and records that change.
original-c3/qualification-score-b01.json: block-1 score file.
original-c3/all-original36-scores.json: scores for all 36 original assignments, including cells outside C3.
original-c3/all36-public-transcripts.jsonl: public transcripts for all 36 original assignments.
original-c3/ALL36_PUBLIC_TRANSCRIPTS.md: plain-text readable supplement of the same 36 transcripts (Markdown source file).
original-c3/delivered-requests.jsonl: request bodies as delivered to the models.

The ticket pilot page reports the 12 C3 negotiations. The all-36 files also cover the other conditions in the original protocol.

Freight emissions study (protocol environment320-v1)

environment-n320/registration-scientific.json: complete scientific registration: 324 units (320 negotiations and four separate probes).
environment-n320/frozen-manifest-scientific.json: frozen scientific manifest.
environment-n320/all-assigned-scores.json: scores for all assigned original units.
environment-n320/delivered-requests.jsonl: request bodies as delivered to the models (large: about 11 MB).
environment-n320/source/binding.py: frozen binding rules for proposals and acceptances.
environment-n320/source/env_core.py: utility and emissions tables.
environment-n320/source/env_replay.py: strict reply parser and replay.
environment-n320/source/analysis.py: analysis source.

Original prompt templates

environment-n320/source/prompts/system_buyer.txt: buyer system prompt template.
environment-n320/source/prompts/system_seller.txt: seller system prompt template.
environment-n320/source/prompts/turn_user.txt: per-turn user message template.
environment-n320/source/prompts/check_user.txt: postcheck question template.
environment-n320/source/prompts/system_probe.txt: system prompt template for the separate probes.
environment-n320/source/prompts/probe_user.txt: user message template for the separate probes.

The frozen settings file is not distributed here; its scientific values are shown in the settings table above.

Data exports

The CSV and JSONL files used on the study pages, with every column explained, are on the data page.