Historical Campaign 010 launch prompt

Archived on 2026-10-07. This is the original authorization wording, not a current launch instruction. Its historical receipt and frozen source are not current readiness. Any future execution requires the current campaign procedure, valid receipt bindings and fresh explicit authorization; do not paste this prompt as a reusable launch template. This archive makes no claim that Campaign 010 ran.

Original prompt:

Launch the prepared OpenRouter representative campaign codira-efficacy-campaign-010.

Read AGENTS.md, docs/process/agent-efficiency-campaign-preparation-2026-10-02.md,
docs/process/agent-efficiency-campaign-factory.md and the corresponding factory
manifest and launch plan under .artifacts/agent-efficiency/campaigns/.
Use the already prepared execution receipt under
.artifacts/agent-efficiency/executions/codira-efficacy-campaign-010-20261002/;
preserve its
fixture bindings, image digest, harness identity, 24 tasks, two arms, seed,
high reasoning and limits. Do not launch subscription campaign 011.

I authorize paying OpenRouter execution using the scoped codira-agent-efficiency-
pilot key and the approved $10 observed-spend campaign pool. Allow the final
response to cross $10; do not reserve hypothetical worst-case spend or reject
extra key headroom. Recheck authenticated model availability, reasoning,
pricing and key allowance immediately before launch without a completion.

First verify the factory artifacts, current source/harness fingerprints,
successful full validate_repo gate, image/runtime qualification, 72 rubric
calibrations and 18 applied patch calibrations. If the frozen controls differ,
report the mismatch before spending; do not silently regenerate this identity.

Launch through scripts/launch_agent_efficiency_pilot.py and its prepared
factory command in durable tmux, using the approved SOPS environment only in
the credentialed child. Preserve complete provider responses, event streams,
protected checks, failure traces, log and exit status. Preserve completed slots;
investigate unfinished starts before any resume. Report operational outcomes
separately from task correctness and blinded semantic adjudication.

Give progress and cumulative observed cost, then a complete task-by-task and
paired analysis, persisting the report under docs/process. Do not replace
semantic review with identifier matching or rerun completed attempts.