Campaign 016 preparation and launch handoff — 2026-10-04¶

The operator approved a separate 200,000-token readiness canary guard with the $0.25 allowance on codira-tests-key. Main campaign controls remain unchanged: 2,500,000 session tokens, 64,000 output tokens, $7 shared pool, $8 daily ceiling, z-ai/glm-5.3-flash with mandatory max reasoning, three repetitions and 144 paired attempts. The approved campaign 015 price revision is retained.

Campaign 015 passed its full offline and repository gate, but the proxy rejected the canary locally before forwarding a completion because its inherited token reservation could not fit $0.25. That failed evidence remains immutable. The new harness binds the separate token guard into canary admission, execution, received usage and readiness receipts, leaving main execution controls intact. Actual proxy guard admission and receipt isolation are covered by focused checks. Main execution remains separately authorized; readiness requires its current immutable receipt and all eleven checks.

Selected controls¶

Control	Value
Specification	`benchmarks/agent-efficiency/campaign-specs/codira-efficacy-campaign-016.json`
Factory directory	`.artifacts/agent-efficiency/campaigns/codira-efficacy-campaign-016/`
Factory fingerprint	`9014d8624fd75e546f621e3b637e919a203fa52e58fce34103a753266abdb505`
Execution root	`.artifacts/agent-efficiency/executions/glm016-20261004-r1/`
Panel and repetitions	`representative-v1`, 24 tasks, three repetitions
Schedule	144 attempts, 72 per arm, seed `261004`
Model and reasoning	`z-ai/glm-5.3-flash`, mandatory reasoning, `max`
Session controls	2,500,000 tokens; 64,000 output tokens; 1,800 seconds
Request controls	45 logical requests; three transport attempts per response
Main accounting	$7 shared pool; $8 daily ceiling; $1.75 attempt planning estimate
Price ceilings per million	$0.16 prompt; $0.60 completion
Canary selection	`codira-tests-key` through `--canary-key codira-tests`
Approved separate canary allowance	$0.25
Canary turn timeout	120 seconds
Separate canary token guard	200,000; main remains 2,500,000

The authenticated catalog observed on 2026-10-03 admitted max, high and low, mandatory reasoning, a 1,048,576-token context and 131,072 maximum output tokens. Observations are retained under .artifacts/agent-efficiency/analysis/glm-campaign-selection-20261003-r1/. Preparation must repeat authenticated admission for each selected key; these historical observations do not substitute for fresh readiness.

The main shared pool stops admission of new work at the observed threshold. The token and attempt figures are planning estimates for main execution, and an in-flight response can cross the remaining pool. The $7 pool does not promise completion of all 144 scheduled attempts. The canary additionally applies its own token and dollar guards and rejects usage beyond its allowance. Canary accounting is separate from the main campaign ledger; ceiling-based estimates are not actual invoiced charges.

Frozen source and environment¶

The selected clean serving checkout is commit 6a3c025e119a4113323bf3eef7b3c2a376a301fe. Its core serving fingerprint is c090198e2cd98b19472bc0cc604041c050d7249873c0afd28c81a56a43c142e5. The updated Codira task fixture uses that commit and tree 06a7593e5387f7fa20589a643873ef0de1d3fa9b. Panel references and calibration documents were regenerated through their repository generators.

The environment image is localhost/codira-glm-campaign@sha256:4353282d3fefec0e8807317053bc980a4856dbcd086d41bb37ebcbab3ff54d53. Its image-profile fingerprint is 3f9a6ab2e93095a7c0ea21217bc1990700550cf0959173d8920f93cfcb20d7f7; the distinct runtime-profile fingerprint is fa064e8d8a4d607aa67ddca2f9ab23dcc20df1494468c17bdec0b3770fe128d2. Build and offline image-qualification evidence is retained under .artifacts/agent-efficiency/environment-images/glm-20261003-r1/. The qualified native executable is version 0.159.3 and accepts max.

Harness, specification and documentation changes must be finalized before readiness. The resulting receipts bind their actual bytes. Do not modify the factory manifests or receipt-bound inputs to repair a mismatch.

Complete preparation, after authorization¶

Use the repository procedure in campaign creation instructions. The operator authorized one bounded canary on the tests key with a $0.25 allowance. Execute this command from the repository root in the host context:

uv run python scripts/launch_agent_efficiency_pilot.py \
  --prepare \
  --campaign-dir .artifacts/agent-efficiency/campaigns/codira-efficacy-campaign-016 \
  --execution-root .artifacts/agent-efficiency/executions/glm016-20261004-r1 \
  --runtime podman \
  --seed 261004 \
  --canary-key codira-tests \
  --canary-budget-usd 0.25 \
  --canary-max-total-tokens 200000 \
  --fixture-source 'click-public=/home/marco/Personalia/Progetti/Software/Python/codira/.artifacts/agent-efficiency/attempts/011/fixture-sources/click' \
  --fixture-source 'picomatch-public=/home/marco/Personalia/Progetti/Software/Python/codira/.artifacts/agent-efficiency/attempts/011/fixture-sources/picomatch' \
  --fixture-source 'codira-current-public=/home/marco/Personalia/Progetti/Software/Python/codira/.artifacts/agent-efficiency/executions/fixture-update-20261003-r1/fixture-sources/codira-current-public' \
  --fixture-source 'python-service-synthetic=/home/marco/Personalia/Progetti/Software/Python/codira/benchmarks/agent-efficiency/synthetic/python-service' \
  --fixture-source 'typescript-workspace-synthetic=/home/marco/Personalia/Progetti/Software/Python/codira/benchmarks/agent-efficiency/synthetic/typescript-workspace' \
  --fixture-source 'go-service-synthetic=/home/marco/Personalia/Progetti/Software/Python/codira/benchmarks/agent-efficiency/synthetic/go-service'

Do not wrap the parent command in SOPS. The launcher scopes the registered tests environment to the canary child. Main preflight, post-canary capacity admission and campaign execution retain the pilot environment. Preserve the complete preparation log, terminal exit status and all eleven check records. The pipeline runs the full repository gate under tmux before inference; do not run a duplicate gate simply to repeat that check. Run the independent Codira audit before preparation. A failed check requires investigation and a fresh execution root; do not retry inference automatically. Changed frozen controls require a fresh factory campaign identity.

Receipt-based launch, separately authorized¶

Require a successful readiness-receipt.json in the execution root, every required check admitted, actual successful model-requested index_status and context_for_task calls, shell-marker completion, complete usage, post-canary main-key capacity admission and unchanged receipt bindings. Inspect the actual accounting evidence; offline tests or successful HTTP transport do not qualify the model or prove task correctness.

After explicit main-campaign authorization, use only the receipt bindings:

uv run python scripts/launch_agent_efficiency_pilot.py \
  --launch \
  --execution-root .artifacts/agent-efficiency/executions/glm016-20261004-r1

Readiness expires after fifteen minutes. Expired preparation requires a fresh execution root and renewed bounded preparation authorization where needed. Preserve existing evidence. Report operational completion and deterministic task-oracle correctness separately when evaluating the campaign.