Reading a report

Every report — web, PDF, JSON — renders from one assembled model, so the three artifacts never disagree. This page walks the sections in reading order.

Header and executive summary

The header states the audit id, generation date, observed window (days of traffic), and call count. The summary leads with one number: estimated monthly savings, with the percentage of current spend it represents. Around it, four cards: current monthly spend (your observed window scaled to 30 days), the optimized projection, raw observed spend, and the findings count.

Every "monthly" figure is the observed window × 30 ÷ observed days. A 3-day sample scales by 10 — the report says so on the card, not in a footnote.

Savings waterfall

Findings ranked by monthly dollar impact, largest first, with proportional bars. This is the triage view: the top one or two rows usually carry most of the recoverable spend.

Spend charts

Observed spend by model and by UTC day — where the money went, and whether spend is trending or spiky. Chart values are observed dollars, not projections.

Findings in detail

Each finding card carries:

Field	Meaning
id	stable finding id (D2-...), deep-linkable
severity	high ≥ $500/mo, med ≥ $50/mo, else low (D4: cluster-based; D5: informational)
confidence	how the estimate was grounded — verified (hash evidence), estimated, conservative, informational
monthly impact	the dollar estimate, computed per the class formula
fix	the concrete change we recommend
evidence	up to 20 sample calls: timestamp, model, token counts, note — never text

Pricing provenance

The report states the pricing-table version and human-verification date used to price it, and lists any unpriced models excluded from totals — count and ids, so you know exactly what the totals do not include.

Methodology and data handling

The final sections print the methodology statement (every haircut and conservatism disclosed) and the data-handling policy — the same text on every report, so the report is self-explaining when forwarded to finance.

The JSON artifact

report.json carries the same model: audit metadata, totals, spend breakdowns, findings with evidence. It is deterministic — the same upload produces byte-identical JSON (generated_at excluded) — so you can diff two audits of the same log and expect zero noise.