Which document is actually current?
A larger context window can hold more contradictory material.
Retrospective field note for 2026-03-09, reconstructed on 16 November 2026. The archive date is not a claim of publication at the time.
Upload the whole folder and the model gets three versions of the brief, two abandoned plans and a transcript explaining why the first version was wrong.
More context can help. It can also give the wrong document more room to look authoritative.
Mark the current source. Keep the decision and its reason together. Separate unresolved questions from settled requirements. That little bit of organisation saves the next worker from treating every sentence in the archive as equally current.
For a long task, maintain a short handoff: what changed, what is decided, what is blocked and what happens next. The transcript can remain available without becoming the only map of the project.
Retrieval and larger windows can work together. Find the relevant material first, then give the model enough room to compare it properly. Finally, point the reviewer to the passage behind the consequential claim.
GPT-5.4 belongs in that evaluation. The test is whether it handles the actual collection more accurately, including conflicting versions. Filling the available window is not an achievement by itself.
OpenAI released GPT-5.4 on 5 March. Source.