Skip to content

Latest commit

 

History

History
97 lines (79 loc) · 4.81 KB

File metadata and controls

97 lines (79 loc) · 4.81 KB

Unattended reservation comparison

Both Luna Max implementations were accepted without a correction worker. The search-guided arm cost 24.15% more this time. Across all three pairs, the worker costs are effectively equal; the earlier 13% advantage did not hold. No operational skill change follows from this result.

Worker results

Pair A: assignment only B: with search guidance B saving vs A
Original $0.03897336 $0.03376816 13.36%
Frozen repeat $0.03868776 $0.03353020 13.33%
Unattended repeat $0.04262936 $0.05292576 -24.15%
Total $0.12029048 $0.12022412 0.055%

These are frozen historical API-equivalent estimates, not bills or Codex allowance measurements. All attempts used Luna Max, confirmed in native usage records; no Sol correction ran. Both arms retained component ownership.

The unattended protocol kept the task, prompts, configuration, checks and B-first order unchanged. The host launched the workers serially, ran checks and stopped at ready_for_review. The original orchestrator ended its launch turn and returned only when the user requested acceptance. No reviewer worker, automation or live-model polling loop was added.

Current pair A B
Responses 15 16
Input tokens, including cache 349,102 442,986
Cached input tokens 295,168 357,888
Output tokens, including reasoning 21,616 23,957
Tool-output characters 14,902 20,698

B used more input and output, not merely a more expensive cache mix. Neither trace showed truncation or compaction. This does not establish that the added paragraph caused the difference: sampling, cache and fixed order remain uncontrolled. Retain useful search discipline, but withdraw a measured implementation-cost advantage on this fixture.

Acceptance review

The original Astra High orchestrator reviewed all three changed source files and the added tests in each arm against the frozen contract. It revalidated the saved receipt, protected files and artifact hashes, without rerunning unchanged tests or editing either implementation.

  • Both passed 16 independent/public test methods, including validation subcases and the fixed-seed 100-operation sequence. Saved added-test results: A 6/6, B 5/5.
  • Validation and duplicate aggregation precede all stock mutation; failures leave stock unchanged. Tenant/SKU identities, first-occurrence order and iterator exception propagation are preserved.
  • Preview calls the same core with dry_run=True. Submission stores independent journal rows, including for empty success. API dispatch preserves the result.
  • No scope violations, artifact-list gaps or unresolved material defects. Worker notes that the external check was pending are resolved by host receipts.
  • Shared active ownership was released. Original attempt receipts remain ready_for_review; the separate acceptance record supplies the human-facing final decision.

This accepts the written contract, not arbitrary callbacks, allocation failures, concurrency or production suitability outside that contract.

Cost through review: recorded subtotal, not a final total

Component Recorded API-equivalent cost
Both workers $0.09555512
Setup/launch turn, completion event observed $1.923442
Review turn snapshot: 8 completed responses $0.864220
Recorded subtotal $2.88321712

The review snapshot stops at 2026-09-12T17:43:11.209Z; it excludes later responses, including report writing and final delivery. Setup is selected by the exact turn identity saved in the manifest, not the latest conversation turn. Original fixture development/audit and machine costs are excluded. Setup and review are shared costs, not arbitrarily assigned to A or B.

The earlier live-supervision snapshot alone was $4.49812, also incomplete and covering more than waiting. These unequal scopes do not establish a percentage saving. What this run demonstrates is that host execution can finish the pair without repeated model wakeups. Even so, setup and acceptance dwarf the roughly ten-cent worker bill. There is still no matched solo baseline proving net savings.

Next useful experiment

Stop repeating this tiny fixture to optimize fractions of a cent. Reuse the host runner on a larger held-out, settled component and compare a solo owner with one component-owning worker under equivalent acceptance requirements. Freeze the acceptance packet and account for setup, corrections and review on both sides. Do not introduce a cheaper reviewer or change search guidance in that same comparison.

The ignored local-unattended.json points to reservation-c2b6b91e, containing original receipts, checks, trace metrics, unattended-supervision.json and the separate acceptance-review.md. No new paid attempt was launched for this review.