Receipts

Three formats

Every guarded run produces a receipt in three shapes:

• Terminal — a concise summary printed at the end of the run.
• HTML — a single, self-contained file you can open, save, or share.
• JSON — machine-readable, for CI gates and dashboards.

What's in a receipt

• Budget vs spent — what the run was allowed, and what it used.
• Waste averted — the projected uncapped spend minus the actual spend (what the budget saved).
• Hops — each step, its model, tokens in/out, and cost.
• Trip reason — if the run paused or was killed, why.
• Side effects — any side-effecting tools that fired before the run stopped.

Honesty

Token counts are self-metered per model family — exact tiktoken for OpenAI, a conservative o200k-based approximation for Anthropic/Google (whose exact tokenizers are API-only, which offline metering can't call) — then reconciled against the provider's reported usage, so the billed number reflects what actually happened, not an estimate. Prompts and keys are never part of the event stream.

Audit artifact & blast radius

Every receipt is an audit artifact: it carries the run_id, the start timestamp, the attribution tags (team/customer/env), and exactly what was spent — offline, owned, no backend.

When a run trips with side-effecting tools already fired, the receipt leads with a blast radius block — a compensation checklist of what hit the world before the stop:

 ⚠ blast radius — these side-effects already fired, verify/compensate:
     ☐ send_email
     ☐ post_to_cms

"If it stopped mid-run, what did it do to the world?" — answered in the artifact, not a war-room.