Archive — read records past the retention window

Find, open and download archived incidents and changes; download a month's fine-tuning data or turn it into Test Lab cases.

How-to · Updated 2026-10-19 · Describes release 0.1.164

Archive

Archive (/portal/archive, under Insights) lists the incidents and changes that went past the workspace's retention window: each one was written to the platform's archive storage, checked byte for byte, then removed from the live database. Here you find them again, open them read-only, and download them. How the retention window and the archive work is in Guardrails and cost limits → Data retention.

Who sees what:

Part of the page Permission
Archived records — the list, opening a record, Download JSON incidents.view, the permission of the live record
Archive batches — one row per archived month, Download training export members.manage
Create back-test cases on a batch members.manage and benchmark.manage

Before you start

Find and open an archived record

  1. Open Archive under Insights in the sidebar, or select Open the archive on the Data retention card.
  2. Narrow the Archived records list: type part of a title in Search titles, pick Kind (incidents or changes), or pick the month the record closed in Closed in. Clear filters brings the whole list back. The list reads the archive index only, 50 records per page, newest close first.
  3. Select a record's title. The record is read back from its month's archive file through the control plane, checked against the file's fingerprint, and shown read-only — an incident with its facts, signals, analysis reports, approvals and timeline; a change with its facts, plan, reported executions, analysis runs and timeline. Credentials stay masked as they are on a live record.
  4. Download JSON saves what is shown. The row's own Download JSON in the list saves the same record without opening it.
  5. Back to the archive returns to the list.

You should see the record's title, an Archived badge and the line Archived on … from the … batch. Read-only: archived records cannot be changed. A record that cannot be read back (the archive storage is unreachable, or its file no longer matches its fingerprint) says The archived record could not be read instead of showing anything partial.

Links keep working too. An incident or change link from a ticket, a chat message or an old notification opens the live page; when the record has been archived, the page says Archived incident or Archived change with the date it was archived and the date it closed, and View archived record opens the same read-only view. A link to a record that was never archived still says it was not found.

Download a month's fine-tuning data

Requires members.manage.

  1. On Archive, find the month in Archive batches. Each row shows the month, its state (Archived, Failed or In progress), what it removed from the database, how many records were Held back for the next batch because they changed after their export, and any error.
  2. Select Download training export on an archived month. It is available once the batch is archived; on a failed or unfinished batch the button is disabled.

You should see a training-<month>.jsonl.gz file download. It is built through the control plane from that month's archive file — never a direct link to the storage — and runs the full PII scrub with the workspace's salt before a byte leaves the platform. Every download writes one audit row.

Turn a month into back-test cases

Requires members.manage and benchmark.manage. A month that left the database can no longer feed Back-test on history in the Test Lab; this builds the same cases from its archive file instead.

  1. On Archive, select Create back-test cases on an archived month.
  2. Pick the Agent: all incidents, the incident agent, the change agent or the compliance agent.
  3. Optionally tick Validated fixes only (only records whose fix a validation later reported cleared) and set a Min human score (only records someone rated at least that highly on the Feedback page).
  4. Select Create cases.

The cases follow the same rules as a back-test on live history: the run whose fix landed, feedback across every run of the record, a human correction wins over the recorded answer, an answer rated wrong with no correction is left out, incidents closed as noise are left out, and each case carries the evidence the original run recorded so the replayed agent can read it. Change cases follow the change-agent rules in Benchmarks and scores → Back-test the change agent; a change whose origin incident was archived in another month gives no case. Every case records where it came from: the batch id and the SHA-256 of the batch's manifest.

You should see Created "archive 2025-09 (12 cases)" with 12 cases. Run it from the Test Lab. (with the skipped count when some records did not qualify) and Open Test Lab. The new golden dataset is listed under Replay (recorded cases) with a from archive badge; its name carries the month, the agent and the options you chose. Nothing runs and nothing is charged until you select Run on the dataset — then it runs like any dataset, sandboxed by default. When nothing in the month meets the rules, no dataset is made and the card says why. Every request writes one Test cases built from the archive audit row, also when it built nothing.

What the archive does not do

Next steps