How the 48-Hour LLM Spend Diagnostic works.

Fit-check: scope and field names, no data

The first email should contain only the business context, confirmation that both OpenAI and Anthropic exports exist, the intended 31-day period, file formats, approximate row count and size, the diagnostic question and a plain list of column or JSON field names. Do not attach exports, sample values, screenshots, API keys, passwords or credentials.

VANTYX Labs confirms which of two supported paths applies: documented provider-native usage/cost JSON as complete pages, or the published cost-bearing normalized schema. The downloadable CSV headers define that schema; an array of JSON objects with the identical keys is accepted too. Unsupported, partial or ambiguous shapes are identified before purchase. A fit-check is not DATA ACCEPTED and does not start delivery.

Download the usage schema · Download the billing schema

Purchase and transfer agreement

After scope is confirmed, the customer may use the applicable one-time Stripe checkout. Purchase does not authorize data transfer by ordinary email.

The default supported method uses one customer-controlled, expiring HTTPS URL per accepted file, with one structured message per file and no more than four files for the engagement. Each URL must be self-contained and require no separately emailed password, bearer token or credential; its embedded access token stays inside the protected transfer boundary, and the customer revokes the URL after confirmed download. Before files move, both sides agree in writing on the protected transfer channel, processing location, permitted files, retention arrangement and deletion expectation. See the data-handling policy.

Acceptance gate

Received files are checked for authorization, supported provider, documented shape, complete pagination where applicable, readability, period, combined size and row limits. Files outside scope, corrupted or partial native pages, credentials, raw prompts, customer content or unnecessary personal data are rejected.

The delivery window begins only when VANTYX Labs sends written confirmation containing the exact state DATA ACCEPTED. Checkout, upload or partial delivery does not start the clock.

Source inventory and deterministic mapping

Accepted files receive a source reference and SHA-256 integrity hash. Original source files remain distinct from working copies. The engine records the source schema, endpoint, format, pagination state, mapped fields, unavailable fields and limitations, then retains a pointer to each originating record.

Native OpenAI and Anthropic JSON is mapped only from documented fields. Normalized templates require an explicit source-supplied cost_usd for monetary reconciliation. Unknown or absent values remain unknown; aliases or costs are never invented merely to force totals to agree.

Reconciliation and supported findings

The engine totals canonical billing rows and canonical usage rows with explicit cost. It reports provider arithmetic only when every canonical usage row has a source-supplied USD cost. Otherwise monetary reconciliation is NOT_COMPARABLE.

Exact duplicates use the source-supplied identifier in the normalized schema, or a deterministic content fingerprint for provider-native records, plus identical normalized fields. They are excluded from the canonical total. Evidence lists are bounded in summary outputs while full usage-row lineage remains in the normalized dataset. Absent model, project or workspace dimensions become explicit unmapped-dimension exceptions. Retry and sequential context-growth observations run only where every required metadata field is present.

No variance, duplicate, retry or context observation is automatically waste, overbilling, savings, a recoverable charge or evidence of causation.

Native versus cost-bearing inputs

Quality checks

• Verify input and output hashes, accepted-source inventory and enforced scope limits.
• Recalculate canonical totals after exact duplicate exclusion.
• Check that every variance, duplicate and unmapped-dimension finding has source references.
• Label unsupported monetary comparisons NOT_COMPARABLE.
• Document unavailable fields and unsupported analyses rather than filling gaps with invented values.

Delivery within 48 hours

The package includes the coverage manifest, source mapping, normalized dataset, reconciliation workbook, variance register, cost-driver evidence, executive report, deterministic methodology, up to five prioritized next checks and a structured written findings handoff.

The 48-hour promise applies after written DATA ACCEPTED. It does not promise that every source difference can be resolved, that native usage can be repriced, that savings exist or that a provider will reimburse any amount.

Method limitations

• The diagnostic is limited to the supplied, complete and accepted exports.
• It does not connect to production accounts or independently verify provider systems.
• It does not independently reprice source-supplied costs or infer costs absent from native usage records.
• Provider beta or grouping-specific dimensions outside the published accepted schema are rejected during fit-check rather than silently ignored.
• It excludes non-LLM cloud infrastructure costs.
• It is not legal, tax, accounting or financial-statement assurance.
• It does not include remediation implementation or ongoing monitoring.
• It does not guarantee savings, recovered charges or any commercial result.