#8081·ironclaw

Daily ironclaw failure taxonomy — 2026-09-07

Author: pranavraja99Created Sep 7, 2026Updated Sep 7, 2026

Daily ironclaw failure taxonomy — 2026-09-07

Suites analyzed

  • officeqa (42 non-pass) — This run's 42 officeqa failures are overwhelmingly genuine model-quality numeric errors: DeepSeek-V4-Flash navigates the OCR-digitized Treasury Bulletins with healthy, self-recovering trajectories (read_file/grep/shell all succeed) but extracts the wrong cell, picks the wrong year/quarter, or applies a wrong formula, so the final number misses the ±1% tolerance while nothing in ironclaw actually breaks. Every 'Hard' capability signature is incidental to a document-grounded task: the Network/OperationFailed/builtin.http[.save] entries are all off-document external fetches (FRED, stlouisfed.org) that correctly fail because officeqa never needs the web, the operation_failed/policy_denied/resource entries are self-recovered arg errors or a missing-numpy fallback inside otherwise model-error trajectories, and UID0012 even printed the correct 36080 in its own shell before submitting the wrong 35532. No harness defect drives any failure here (the prior run's InputEncode-mislabel and apply_patch-message issues do not surface as failure causes), so there is no ironclaw fix and the PR order is empty. Two caveats worth a separate bad_eval audit but unconfirmable from this bundle: UID0013 is an exactly-1000x unit mismatch that could be a golden-answer units convention, and UID0041's Theil index (0.0118 vs expected 0.011) may reflect a data-selection/methodology difference.

ironclaw rev(s): 0280dd1a48a4f813f123735b31558c459d1ce032

Harness fix candidates (ranked by blast radius)

No harness-attributed failure modes today.

Per-suite breakdown (every task links to its trajectory)

officeqa (42 non-pass)


Generated by the bench-taxonomy workflow. Categories are dynamic — discovered from each run's failures and root-caused via codegraph/source.