Solutions
The same corpus, three different problems.
PIT answers what the public record held at a given minute and hands back the evidence for the answer. Which part of the platform matters depends on who is asking. A reviewer suspects your benchmark is memorized; a join quietly reads a few hours of the future; a compliance officer wants the evidence a year after the run. Each page below names the products that job actually uses.
Eval builders
You publish benchmarks or gate a model rollout on one, and every public-data eval you run is contaminated by construction. You use the harness to run the agent with the network off, benchmark datasets for the redacted and date-shifted control arms, and a certificate so the reviewer can check the result without trusting either of us.
Systematic research
You run event studies over filings and know lookahead bias by name. 10.3% of filings in our contiguous window were accepted by EDGAR after the day boundary a filed-date join uses. You use the point-in-time API for cuts that name their clock and resolve tickers at the query instant, flat files when the study is faster offline, and the published benchmarks as the template for measuring what a clock choice costs.
Accountable teams
Your claim will be audited by compliance, an allocator or a customer. You use decision audit logs to receipt what a production agent read, certificates to make one run checkable by a stranger, and the registry when you want the result listed rather than merely claimed.