原始文档的离线阅读副本 · 原始文件

Finance Ops training interruption

Recorded on 2026-10-18 (Asia/Shanghai).

Native run: 8e16e14d-3151-4fbc-af8e-389fd7c0affe. Training session: d96e8b0d-b873-403f-b7bb-75f92b9d0e08. Frozen catalog family: Group 009 branch-close, train_001 to test_001. Fixed assignment tests experience effects, not retrieval accuracy.

Observations

• Training is terminal failed; no paired target rows or candidate exist.
• Four retained assistant responses account for 44,541 known task tokens.
• Eighteen tool calls have no failed results, missing-file results or business rejections.
• Finance records were retrieved and their full native spill was read.
• Subsequent queries retrieved compensation and payroll data.
• Step 5, event 61, records an assistant attempt ending with pi-ai stream idle timeout after 300000ms, failure code TIMEOUT.
• Event 62 records one scheduled retry. This does not prove the retry produced a response.
• Event 64 records a turn abort. The controller's error is TimeoutError: The operation was aborted due to timeout.
• The measured followup-to-idle interval is 1,008,246.620125 ms and is incomplete. It is not the same measure as the provider's 300,000 ms idle timeout threshold.

No final answer was written. Grade, schema validity and task success are unknown. Assessment, extraction, screening and target arms were not executed. This run cannot demonstrate memory delivery, benefit or harm. Missing usage is not zero; retained response usage is not a verified complete bill.

The upstream response failure's root cause is unknown. Successful HTTP tool responses do not establish every returned dataset was relevant or sufficient. The unrelated compensation/payroll reads are observable, not a proven cause of the timeout.

Execution And Cleanup

Current launcher host bundle was built before staging. Execution used the built-in demo controller and native Harness sessions; the host script did not call the model API directly. The normal daily profile was not changed.

After terminal failure, all native sessions were observed idle. Raw export: audit-export-68295809/. Audit FAIL reflects absent complete pair/candidate, not a finding that recorded tools failed. Frozen sources, archive identity, no production release and rejected-not-applied checks passed.

The staging host was stopped after export. web-final-status.json records STOPPED, native process exit code 0 and environmentServerStopped=true. The viewing endpoint was closed; use the static report for retained evidence.

Earlier Startup Failure

../native-rsi-web-6ua03age/web-final-status.json preserves an earlier expected_single_official_evaluator startup failure before native model execution. Group 009 exposes an official shell entry, unlike prior Python evaluator layouts. The driver now selects eval.sh only when no Python evaluator exists, preserving existing groups that include both. The host wrapper invokes the official shell entry. Shared evaluator, configuration and reference hashes are host-only dependencies checked before grading; task tools cannot access them.

The corrected infrastructure startup used this new directory. No completed or failed model attempt was overwritten or retried for a favorable outcome. Synthetic shell invocation and schema-validation regression passed; the native training interruption prevented live end-to-end acceptance of this group's grader.