Take one question and make it harder.
You do not need an account, agreement with the project, or a particular AI model. Pick a task, keep its evidence rules intact, and send back the result—even when the anomaly disappears.
The guide comes before the prompt.
A task-specific master guide defines what can count as evidence, what must remain unresolved, and which shortcuts would accidentally build the answer into the run.
Evidence guide
Loading…
Run prompt
Loading…
Independence is useful.
Use another ChatGPT session, another model, your own archive search, a numerical solver, or a human analysis. Record tool/model versions and deviations. Do not quietly improve the prompt until it produces the expected result.
Keep
Raw output, URLs and archive identifiers, page numbers, hashes, code revision, parameters, convergence/failure information, and the exact master guide/prompt used.
Especially keep
Negative retrievals, failed replications, contradictory sources and simulations that make one of our own anomaly claims weaker. Those are often the most valuable donations.
Optional identity
identity.transductive.science can give repeated contributions a browser-local continuity label. Reading and contributing do not require it. Never put credentials or private data in a submission.
Show us the restraint as well as the result.
New submissions enter the project as pending. Approval means the stored object was reviewed; it does not mean its interpretation became canonical.