Skip to content

Campaign 009 selective completion and composite result

Campaign codira-efficacy-completion-009 filled the six failed or unstarted slots of codira-efficacy-campaign-006 without rerunning its 54 completed task attempts. It used a fresh factory-generated identity. The source plan, state, and all 55 original records were bound by exact byte digests in the new launch plan. The parent campaign's failed patch-002-r05-codira-mcp record remains intact; its replacement is used only in the composite analysis.

Recovery and execution

A power failure interrupted candidate 009 before its authenticated preflight. Its generated factory artifacts, prepared fixture receipt, and completed repository gate survived. Its execution state contained no attempt marker, record, log, or paid response. After restart, the factory check passed, the pinned runtime image remained cached, and authenticated route and key preflight passed again. The runner then launched in tmux and exited 0 with six settled results, no pending attempt, and no operational or oracle failure. All six records passed immutable-state validation with complete provider usage.

The six attempts charged $0.48971925 at the frozen price ceilings against the completion campaign's $2 observed-spend pool. These are accounting estimates from received provider token usage, not a provider invoice.

Pair Codira-MCP result, tokens, cost Baseline result, tokens, cost
patch-002-r05 Pass; 540,874; $0.14291300 Pass; 275,416; $0.07209500
impact-001-r05 Pass; 232,083; $0.06036375 Pass; 370,602; $0.09695250
architecture-001-r01 Pass; 307,728; $0.08041850 Pass; 142,778; $0.03697650

The patch Codira arm that previously stopped at a local token reservation completed under observed-spend accounting. Its new result passed patch application, required path checks, and the protected command. This is one successful new trajectory, not proof that the previous trajectory would have converged if continued.

Composite interpretation

Campaign 006 retains 55 original records: 54 operational passes and one failed Codira patch attempt. Its frozen grades are 37 oracle passes and 17 literal text-oracle failures among the 54 completed task attempts. The separate answer adjudication found that all 17 rejected text answers contained the required task facts; a read-only replay of all 28 completed text answers under the revised oracles passed 28/28. Those retrospective grades do not alter campaign 006's records.

Selecting those 54 completed task attempts and candidate 009's six passes fills 60 distinct logical slots in 30 pairs. All 60 have complete operational results. Under the documented revised text scoring, all 60 satisfy their applicable task oracles; this is a retrospective composite, not a single campaign executed under one harness and oracle revision. Report the parent and completion provenance separately in any comparison. The oracles check specified facts or protected behavior, not the full quality of every explanation.

Campaign 006's recorded charge was $4.09832850, including $0.18740500 for the abandoned patch attempt. Adding the completion run gives $4.58804775 total observed ceiling-priced spending. The 60 selected composite results account for $4.40064275; the difference is the preserved failed attempt's charge. Neither amount is an invoice total.

Evidence

  • Factory specification, generated manifest (.artifacts/agent-efficiency/campaigns/codira-efficacy-completion-009/campaign.json), and launch plan (.artifacts/agent-efficiency/campaigns/codira-efficacy-completion-009/launch-plan.json).
  • Ignored preflight (.artifacts/agent-efficiency/gates/completion-009-20260930/preflight.json), control checklist (.artifacts/agent-efficiency/gates/completion-009-20260930/control-checklist.json), launch receipt (.artifacts/agent-efficiency/executions/c009/launch-receipt.json), log (.artifacts/agent-efficiency/executions/c009/logs/pilot.log), exit status (.artifacts/agent-efficiency/executions/c009/pilot.exit), and checkpoint (.artifacts/agent-efficiency/executions/c009/state/budget/checkpoint-000.json).
  • The saved full gate is log (.artifacts/validation/repo-gates/completion-stage-r5-20260930/validation.log) and exit status (.artifacts/validation/repo-gates/completion-stage-r5-20260930/validation.exit): 0, 1,204 passed, three skipped.
  • Final repository validation after the report was staged is log (.artifacts/validation/repo-gates/completion-009-final-20260930/validation.log) and exit status (.artifacts/validation/repo-gates/completion-009-final-20260930/validation.exit): 0, 1,204 passed, three skipped.

Raw provider bodies, protected-command traces, and fixture locations remain only in ignored per-attempt evidence. This report contains no credential values or paths outside the repository.

Subsequent runtime qualification finding

The full audience findings and internal product assessment record the 2026-09-30 forensic follow-up. Offline inspection showed that the pinned image served the older character-budget result.context protocol, while the public Codira source fixture contained the newer item/cursor code. The saved context calls confirm the older deployed shape. These results therefore do not evaluate the current pagination implementation. Composite selection, spending, original grades, and immutable attempt records are unchanged; the follow-up adds substantive quality review and corrects protocol attribution.