Incumbent computer-use diagnostic campaign

Incomplete benchmark. Candidate: Incumbent plus lifecycle repair, private v14 (experimental). Identity complete: true.

Measured 2026-10-08T11:41:06.846164+00:00 to 2026-10-09T01:18:37.640384+00:00. Sealed 2026-10-09T01:31:37.950692+00:00. Revision 1.

Latest revision and correction history · All campaigns · RSS feed

Results and coverage#

65 attempts recorded. The denominator includes every attempted outcome. Unexecuted attempts are listed separately.

Qualification and execution modes

65 of 192 frozen cells observed. Execution modes have separate populations and are never pooled into one performance score.

• An attempt has an unknown independent goal outcome.
• An attempt lacks one exact artifact-backed independent outcome and cleanup witness.
• An attempt lacks original native observations and measured budget, intervention, effect and data-integrity counters.
• An attempt required external intervention and is not an independent comparison trial.
• The frozen matrix has missing, duplicate or unplanned attempts; valid failures remain counted.
• The registered execution harness is contradicted or unverified by scoped post-recovery ancestry evidence; fresh-agent timings are withheld.

Scenario, model and environment#

Each dimension is a separate partition of the campaign. Do not sum overlapping dimensions together.

Workflow, computer and execution mode#

Every planned cell stays assigned to its workflow and computer environment. Rows aggregate fixture seeds within each execution mode. Successful duration uses passed attempts only; missing measurements remain visible. Ranges are descriptive, with no percentile or improvement claim.

Workflow results CSV

Baseline comparison#

No matched baseline was supplied. This report makes no improvement claim.

Additional measurements#

Method and provenance#

Suite: Twelve complete browser, native and mixed workflows; version: 12-workflows-v1.

Frozen matrix of twelve workflows on four physical computers: 48 scripted reference trials and 144 fresh-agent trials with seeds 17, 29 and 43. Fresh agents use GPT-6 Astra with ultra reasoning in separate contexts. Each trial has a 600-second and 60-call budget plus a separate 60-second cleanup deadline. The campaign stopped early after execution-harness identity drift was discovered; the original six-hour maximum was not reached or reset. All 192 planned slots remain: 19 scripted and 46 fresh-agent attempts were recorded, and 127 slots were not started. Native Ibara input was required. Independent fixture files, application events and retained native evidence determine original task outcomes. Workflow rows separate computer and execution mode. All fresh-agent timing is withheld because execution-harness identity was contradicted or unverified. Workflow successful-duration rows include passed attempts only; the overall scripted attempt-duration mean includes all 19 measured attempts. These timings are descriptive. Original outcomes and evidence remain unchanged.

Evidence type: warehouse_snapshot.

snapshot_sha256: 799ecc24f7725c6ac903591e8762d5aa738f1d9d4fc68eb5926e1ccff3d689b2

Failures and limitations#

• This incomplete campaign is diagnostic evidence, not an accepted comparison baseline. Its 42 passed, 16 unknown, 4 failed and 3 blocked original outcomes describe 65 attempted slots; 127 planned slots were not started. No improvement or general reliability claim is supported.
• After a host crash, the root and coordinator execution ancestry used Codex CLI 0.162.0 instead of registered 0.161.0. Fifteen resumed fresh-agent contexts are affected. Direct child executable witnesses were not captured, and the process start time does not establish when the harness changed. All fresh-agent timing aggregates are withheld; original task results remain recorded.
• The displayed measured_at evidence interval includes preparation and later assessment records; it is not the timed trial window. Primary execution began on 8 October 2026 at 22:24:14 UTC and had stopped by 9 October at 01:11:39 UTC, as recorded by the early-stop receipt. The campaign stopped early for the harness identity issue. The maximum cutoff was unchanged but not reached. Unstarted slots were not replayed or replaced, and no new campaign window was inferred.
• The host crash lost four active agent contexts. Original bounded collectors and cleanup receipts were retained, but interruption coverage and exact crash timing are incomplete. Two earlier active trials were also affected by an explicit root interruption.
• One editor save prompt required a person to save and close the test document before hand-back. That later recovery does not turn the original timed cleanup failure into success. An early-stop trial remains partial with its original outcome.
• Three fresh editor trials were blocked before acquisition by a missing verification host field and a frozen retry fence despite a retry-safe validation response. Scripted OCR misses also caused original unknown outcomes; later visual diagnosis does not replace them.
• Two CSV trials produced the correct filtered rows with CRLF line endings but failed an exact-byte oracle expecting LF, although that distinction was not stated in the public task goal. Two browser trials used native context menus to open the required tab but failed the frozen click-route oracle. These original failures remain visible.
• Interactive-input measurement had a collector payload-shape defect in one scripted trial and missing scoped input-release witnesses in fresh trials. An independent later review verified one file-manager task, but its original unknown score remains unchanged. Post-task assessments never substitute for original scores.
• Three planned fresh variants per workflow and computer are insufficient for precise reliability estimates. Missing coverage and early stopping prevent balanced comparisons across workflows or hardware. Scripted and fresh-agent results must remain separate.
• Model tokens and cost were unavailable from the actor harness; unavailable values are not zero. Supplemental browser-capability checks are separate diagnostic evidence and are not included in this 192-slot primary matrix.

Data and charts#

• Public report JSON
• Outcome counts CSV
• Measurements CSV
• Report schema
• Data dictionary and reuse notes
• Checksums
• Vector chart
• Chart alt text

When citing these measurements, credit ibara and link this dated report revision. Private trajectories and third-party application content are not distributed here. This report does not change stable product documentation or qualify an untested release.