Verification record

Calculator proof

The calculator reproduces the executed Cheapifier findings exactly. Organizational totals are transparent extrapolations from those findings.

6 / 6scenario paths reproduced

93real-model cases executed

3,447 / 3,447automated checks passed

867 / 867semantic probes passed

600 / 600preset combinations verified

Executed findings

Each workload below was run locally through the complete Cheapifier pipeline with ONNX Runtime and the real LLMLingua-2 XLM-RoBERTa model at a 95% keep rate. The calculator’s configured stage totals are compared against this source report in the automated gate.

Workload	Cases	Baseline	Measured final	Calculator final	Reduction	Checks
Balanced knowledge work	36	245,228	103,799	103,799	57.672452%	1,272 / 1,272
Software engineering	12	209,353	34,315	34,315	83.609024%	453 / 453
Customer support	9	80,033	16,037	16,037	79.962016%	324 / 324
Regulated knowledge work	9	39,630	15,098	15,098	61.902599%	318 / 318
Research and content	9	90,933	26,323	26,323	71.052313%	330 / 330
Tool-intensive automation	18	455,850	51,398	51,398	88.724800%	750 / 750

What is proven

The six calculator workload records match the local real-model report.
Every intermediate stage total and every final token total is reproduced.
The downloadable evidence matches the values compiled into the calculator.
All 600 company × maturity × usage × workload preset combinations return finite, deterministic results.
Cost calculations reconcile to the declared token mix and example prices.

Projection assumptions

Company size and usage intensity determine the baseline monthly token volume.
AI maturity determines what share of input traffic uses Cheapifier.
The executed workload path is scaled linearly across that qualified traffic.
Dollar examples use 75% input tokens and $5 input / $15 output per million tokens.
Monthly-spend mode converts dollars to baseline tokens at the resulting $7.50 blended price per million.
Observed semantic preservation is evidence for these 93 cases, not a universal guarantee.

Calculation

usage baseline = users × tokens per user per day × 22 days
dollar baseline = monthly AI spend ÷ $7.50 blended price × 1,000,000
qualified input = baseline × 75% input share × maturity coverage
scenario stream = qualified input + measured overhead per interaction
scenario stream = scenario stream × each recorded sequential stage ratio
final tokens = unchanged output + unchanged unqualified input + scenario stream
net savings = original baseline − final tokens

Example cost uses the original and final input/output totals at the displayed $5/$15-per-million price basis. Model-routing savings are excluded from the four-control calculator.

Reproduce the gate

npm run prove:calculator
npm test
npm run build

Source report SHA-256: c252393dc247a62c5e33493649aa2d0eab4c4caafcc4ac47c5992178a05043c9
LLMLingua-2 model SHA-256: 0fddee8b1e0de33216b534370b884481f033f8d5225e73ac88f38bb8f829caff
Tokenizer SHA-256: d5e1e0abdc17f93ca3c0c02a75d90c7c1f4ef556dd1a4e5c12d4ad4e8812e112

Download executed evidence Download proof receipt