See every prompt envelope repeat.
Parse owned inference JSONL, reconstruct loop-level token volume, separate cache from envelope size, and inspect the one-final-envelope comparison without pretending it is a billing guarantee.
Inference events JSONL
Dated fictionalized sample from the supplied forensic capture. Replace with equivalent owned fields.
Awaiting owned JSONL.
Loop evidence
A cache hit is displayed separately from the submitted prompt envelope.
| Loop | Prompt | Cached | Completion | Cache % | Attempts |
|---|
Observed arithmetic
Total prompt volume is the sum across inference_done events. Cache ratio is cached tokens divided by prompt tokens. Prompt-to-completion ratio compares the two totals. Additional prompt volume is observed total minus the final loop's envelope.
Forensic boundary
The source capture is dated 2026-07-12 and references client/runtime versions 0.2.93 and 0.2.82. The one-loop envelope is a volume counterfactual, not a claim about current xAI billing, cache discounts, platform behavior, or guaranteed credit savings.