DEMONSTRATION ONLY · SYNTHETIC DATA · NOT A CUSTOMER RESULT

Engine-generated proof

One deterministic run. Every boundary visible.

This page reports the exact output of the public synthetic fixture, not a hand-written customer case study.

No customer, savings or performance claim.

All names, identifiers, source rows and amounts come from deterministic synthetic fixtures. They prove that the engine executes its documented checks and produces inspectable artifacts. They do not prove demand, customer savings, recovered charges, accuracy on an unseen dataset or VANTYX Labs past performance. Provider names identify supported source types only; no affiliation or endorsement is implied.

Executable proof

Inspect the exact generated artifacts

These files are produced by the same deterministic diagnostic path summarized below. The workbook and generated HTML are human-readable; the two JSON files expose the compact proof and the source/output manifest.

XLSXReconciliation workbookCoverage, mappings, totals, exceptions and actions HTMLGenerated executive reportThe exact self-contained report produced by the engine JSONProof summaryExact coverage, findings and provider arithmetic JSONCoverage manifestSource hashes, capabilities, limits and output hashes

Coverage and arithmetic

What the synthetic fixture proves

Usage rows6 → 5Received → canonical after exact duplicate exclusion

Billing rows2 → 2Received → canonical

Supplied usage cost$150.00Canonical cost-bearing usage rows

Supplied billed amount$154.00Canonical billing rows

$154.00 billed−$150.00 usage=+$4.00 observed difference

The arithmetic is exact because this fixture uses the published normalized template with explicit source-supplied USD costs. The engine does not independently reprice those values, and the $4.00 difference is not labeled waste, overbilling or recoverable money.

Source coverage

Four fixture files, source-linked

Fixture source	Kind	Rows received	Canonical rows	What is demonstrated
synthetic-openai-usage.csv	Usage	4	3	Explicit costs, retry/context fields and exact duplicate exclusion
synthetic-anthropic-usage.json	Usage	2	2	Explicit costs and a project mapping limitation
synthetic-openai-billing.json	Billing	1	1	Customer-supplied billed total
synthetic-anthropic-billing.csv	Billing	1	1	Customer-supplied billed total

Download the exact synthetic inputs and compare their SHA-256 values with the coverage manifest. No customer data is included.

CSVOpenAI usage fixture4 source rows, including one exact duplicate JSONOpenAI billing fixture1 source-supplied billing row JSONAnthropic usage fixture2 source-supplied usage rows CSVAnthropic billing fixture1 source-supplied billing row

The generated coverage manifest contains SHA-256 hashes for every input and output. Source references in findings use the exact form sha256:<file-hash>:row:<number>.

Provider reconciliation

Exact arithmetic on supplied values

Provider	Canonical usage cost	Billed amount	Billing minus usage	Status
OpenAI — synthetic	$125.00	$130.00	+$5.00	Variance
Anthropic — synthetic	$25.00	$24.00	−$1.00	Variance
Total	$150.00	$154.00	+$4.00	Synthetic demonstration

Positive means billed amount exceeds the explicit usage cost supplied to the fixture. Negative means the supplied billed amount is lower. Neither direction establishes provider error or a recoverable charge.

Supported findings

What the engine observed—and what it did not claim

SYNTHETIC OBSERVATION

01One exact duplicate usage row

Excluded from canonical reconciliation

The repeated OpenAI record has the same stable record ID and identical normalized fields. Its explicit $10.00 source cost is reported, but it is not presented as savings.

SYNTHETIC OBSERVATION

02One unmapped dimension exception

project_id absent on two canonical Anthropic rows

The engine preserves the missing dimension as an explicit coverage limitation and asks the source owner whether a grouped export or mapping exists.

SYNTHETIC OBSERVATION

03Retry metadata present

One OpenAI retry-attempt row · $10.00 supplied cost

An attempt number greater than one is observed. The engine does not infer that the retry was unnecessary, wasteful or caused by any particular system.

SYNTHETIC OBSERVATION

04Sequential context growth present

One OpenAI conversation · maximum observed growth 16,000 input tokens

Positive input-token growth is descriptive evidence only. It does not by itself prove inefficiency, causation or avoidable cost.

Anthropic retry and context-growth analysis is N/A in this fixture because the required metadata is not present in every relevant source. The engine does not manufacture it.

Five prioritized checks

Bounded actions derived from exact source references

Verify the Anthropic variance.Review the observed −$1.00 billing-minus-usage difference against the supplied source records.
Verify the OpenAI variance.Review the observed +$5.00 billing-minus-usage difference against the supplied source records.
Confirm the project mapping.Determine whether the Anthropic source can be exported or mapped with project_id grouping.
Remove the exact duplicate before downstream use.Keep the repeated usage row excluded unless the source owner provides contrary evidence.
Review the observed retry attempt.Confirm whether the OpenAI attempt greater than one was necessary; do not assume waste.

Capability boundary

Native JSON and normalized costs answer different questions

Input path	Cost total	Usage drivers	Dollar comparison	Critical limitation
Provider-native JSON	Provider-reported only	Supported where exported	NOT_COMPARABLE	Not independently verified against an invoice; source categories not present in the report remain outside the result, and native usage lacks explicit cost.
Cost-bearing normalized template	Customer-supplied billing total	Supported where supplied	SUPPORTED	Exact arithmetic uses source-supplied costs; those costs are not independently repriced.

DEMONSTRATION ONLY · SYNTHETIC DATA · NOT A CUSTOMER RESULT

What this proof does not establish

It does not establish that any OpenAI or Anthropic invoice is wrong.
It does not establish waste, savings, recoverable charges or causal explanations.
It does not independently reprice source-supplied usage costs.
It does not show how the engine will classify fields absent from a future customer export until those fields are mapped.
No production account, API key, raw prompt, customer content or personal data was used.

Use the real scope, not the demo numbers

Have OpenAI and Anthropic exports to map?

Start with a no-file fit-check. Describe the source schemas and question; do not attach exports or credentials.

Check fit