Approved campaign 015 controls — 2026-10-04

Historical approved price controls. Campaign 015 was superseded after a local canary rejection; its original specification and evidence remain immutable. See the preparation history for outcomes and replacements. This record is not a preparation instruction or renewed authorization.

The specification remains benchmarks/agent-efficiency/campaign-specs/codira-efficacy-campaign-015.json.

Campaign 014 passed the image, both native tool arms, fixture freshness, rubric calibration, all 18 applied-patch cases and the full repository gate. The gate reported 1,295 passed, three skipped, 85% coverage and no blocking scan findings. Authenticated admission then rejected the public input price above its frozen ceiling. No paid model call occurred. Its execution root and complete evidence remain immutable under .artifacts/agent-efficiency/executions/glm014-20261004-r1/.

The refreshed public catalog listed z-ai/glm-5.3-flash at $0.15 input and $0.50 output per million tokens. The scoped authenticated pilot catalog observed at 2026-10-04T07:03:09Z listed $0.0352 input and $0.50 output, with mandatory reasoning and supported efforts max, high and low. The admission contract must cover both catalogs. Raw public model metadata and sanitized authenticated controls are retained under .artifacts/agent-efficiency/analysis/glm-campaign-selection-20261003-r1/ as public-controls-20261004-r1.json and controls-campaign-20261004-r1.json.

Control	Campaign 014	Approved campaign 015
Prompt ceiling per million	$0.04	$0.16
Completion ceiling per million	$0.60	$0.60
Attempt planning estimate	$1.60	$1.75
Campaign shared pool	$7	$7
Daily ceiling	$8	$8
Separate tests-key canary allowance	$0.25	$0.25

The revised planning estimate covers the conservative token reservation: 2,500,000 session tokens at $0.60 per million, plus a 1,048,576-token context reserve at $0.16 per million and 64,000 output tokens at $0.60 per million, totaling $1.70617216. Main shared-pool execution retains its existing observed pool enforcement; this estimate does not promise completion of the schedule.

The model, mandatory max reasoning, image and fixture identities, three repetitions, 24 tasks, 144 paired attempts, seed 261004, 2,500,000 session tokens, 64,000 output tokens, 1,800-second attempt timeout, 45 logical requests and three transport attempts per response remain as selected. The bounded readiness canary uses codira-tests-key and a 120-second turn timeout. Main execution continues to require separate authorization.