Worked example¶

Our test suite audits an engineered fixture — the "waste pack" — that plants exactly one specimen of each waste class in ~3 days of synthetic traffic. The suite pins each finding to a hand-derived dollar figure; if the engine ever computes a different number, the build fails. These are those numbers, with the arithmetic.

Every derivation below was computed independently of the engine (spreadsheet Decimal arithmetic first, engine second) and founder-verified on 2026-07-17 against the provider rate pages.

The fixture spans exactly 3 UTC days, so monthly = observed × 10.

D2 — missing cache: $0.246784/mo¶

30 uncached claude-sonnet-5 calls, identical 5,400-char prefix, 2,000 prompt tokens, 130s apart. Intro rates: input $2, read $0.20, write $2.50; TTL 300s.

cache windows = 13  →  13 writes, 17 cached reads
cacheable     = min(2000, 1024) = 1024 (hash-verified cap)
gross         = 17 × 1024 × 1.80 / 1e6 = 0.0313344
write penalty = 13 × 1024 × 0.50 / 1e6 = 0.0066560
observed      = 0.0246784  →  monthly 0.246784

D4 — retry storm: $0.0510/mo¶

5 identical gpt-5.4-mini calls 20s apart (500 prompt / 200 completion):

per call = (500 × 0.75 + 200 × 4.50) / 1e6 = 0.001275
wasted   = 4 × 0.001275 = 0.0051  →  monthly 0.0510

D1 — oversized model: $1.35/mo¶

25 claude-opus-4-8 tag-extraction calls (1,500 prompt / 60 completion), downgrade claude-sonnet-5:

0.009 − 0.0036 = 0.0054 per call × 25 = 0.135  →  monthly 1.35

D3 — prompt bloat: $0.50/mo¶

Bloated route: 20 calls at 6,000 prompt tokens vs corpus bin median 1,000:

excess 100,000 tokens × $1 input × 0.5 factor / 1e6 = 0.05  →  monthly 0.50

D5 — unbounded max_tokens: $0.00 (informational)¶

12 calls declaring 8,192 max_tokens against completion p95 of 120. Flagged (8192 ≥ 4 × 120); impact $0.00 because no reserved billing applies — the report refuses to invent a number.

D6 — chatty agent loop: $0.096/mo¶

10-call run, 65s spacing, 80-token completions, 6/10 calls sharing a prefix:

saved_calls 8 × overhead 1200 × $1 / 1e6 = 0.0096  →  monthly 0.096

The false-positive guard¶

The same suite audits a clean_optimal fixture — traffic engineered to be efficient — and asserts zero findings. A detector that fires on clean traffic fails the build. Precision is tested, not assumed.