Interrupted runs and outputs¶
Start with the state you can observe, then use the matching check before deleting or resubmitting anything.
The terminal stopped but processes remain¶
Check whether the experiment manager and child processes still appear in the run logs or process list. If a normal run monitor is still active, let its termination path finish. When shutdown has failed and the processes are left behind, run:
kill forcibly terminates all matching GPTNT interactive entry points and KTANE processes, not
only one manifest's cluster.
status reports running after the run stopped¶
status overlays queued and running attempt names from GET /active when the experiment manager
is reachable. If the manager is stale, stop the leftover cluster and rerun status. The local
footer ledger itself reports only done, failed, and not attempted.
status reports failed or not attempted¶
Point --output-dir at the exact recorder directory printed by run. not attempted means the
ledger found no matching footer group. failed means files exist but the group is incomplete,
hard-crashed, or lacks a valid solved, timeout, or strikeout outcome.
Inspect output/logs/run_<run-output-name>/ for the process that stopped. Correct the runtime or
provider condition, then rerun the manifest. Completion filtering skips valid attempts and queues
the remaining ones.
A .parquet.tmp file remains¶
The recorder writes experiment-*.parquet.tmp before an atomic rename. A remaining temporary file
is not a completed record. Preview cleanup:
Only after checking every listed file, apply it with --execute.
Cleanup deletes source evidence
cleanup-outputs --execute deletes invalid record groups and orphaned temporary writes. Keep
completed Parquet files through submission validation.
build-db rejects the database schema¶
The existing database has a v1 or otherwise incompatible base-table structure. Rebuild from the retained Parquet:
gptnt build-db <recorder-directory> \
--output output/experiments.duckdb \
--delete-existing-db
Static predictions exist without metadata¶
GPTNT refuses to assign current provenance to older prediction_*.json files when run_meta.json
is missing. Move the incomplete output directory aside or remove it after preserving anything
needed for diagnosis, then rerun the static task from the beginning.