Older ChatGPT Work Completion Can Make Newer Conversation Turns Disappear

Evaluation scope

ChatGPT Work conversations where an older asynchronous task completes after newer conversation turns have already occurred. Evaluates preservation of visible timeline order, newer-turn retention, branch visibility, reload/reconnect consistency, and deterministic recovery. Excludes unsupported permanent-deletion claims, ordinary message editing, unsupported root-cause attribution, and the separate partial-execution completion-ledger problem evaluated by HXR-EVAL-0053.

Question being evaluated

When an older ChatGPT Work task completes after newer turns have already been added to the same conversation, does the product merge that completion into the conversation without hiding, replacing, or making newer turns inaccessible?

Current findings

What works well or deserves recognition

The missing exchanges remained available to the assistant's active context after they disappeared from the visible UI, and external Notion changes made during those exchanges remained saved. This indicates that some underlying state survived and enabled recovery, even though the original visible history was not automatically restored.

Difficulties and opportunities to improve

Newer conversation turns became inaccessible from the visible timeline after an older asynchronous Work task completed late. This created apparent data loss, broke conversational continuity, and forced manual transcript recovery. A visible conversation that can regress after late completion is not a reliable record of the user's recent instructions and decisions.

Mixed findings and limits

The evidence supports visible-history disappearance or regression, not permanent server-side deletion. A delayed completion being reconciled against an older conversation state is one plausible explanation, but branch replacement, client rendering state, synchronization, or another mechanism remain unverified. The recovered transcript is a reconstruction from assistant context, not a server-side conversation export, and may not contain every missing turn.

Organization information and response status

OpenAI Work reliability case 15955492 has been notified of this late-completion/history-regression occurrence. HXR attached the private late-completion screenshot and recovered transcript, described the failure as materially different from the stream/partial-execution problem, and asked Support to split or reroute it if a different engineering owner or support case is appropriate. No substantive provider response specific to HXR-EVAL-0054 is recorded yet.

How HXR would test performance

Start a long-running Work task, allow its stream to interrupt, then continue the same conversation with several newer user/assistant turns while the older task remains active. When the older task later completes, verify that every newer turn remains visible in original order, the late completion is inserted or linked without overwriting newer history, reload/reconnect preserves the same timeline, no silent branch switch occurs, and an export or conversation readback matches the visible sequence.

Evidence coverage and limits

The focal occurrence is supported by a private screenshot of the older task's delayed completed response, a private recovered transcript preserving the substantive newer exchanges after they disappeared from visible history, contributor testimony about the disappearance, and readback confirming external Notion changes from those exchanges remained saved. These sources share one session and do not establish system-wide prevalence.

Next observation or verification

Await OpenAI's response on whether the late-completion/history-regression failure should be split into a separate support or engineering case. Preserve any explanation of branch/state reconciliation and test future late-completion scenarios for turn retention and visible-history stability.