A tool for reading tax filings, carefully.
AYVES is a research tool that reads publicly filed Country-by-Country tax reports (PCbCRs) and turns them into a six-indicator risk read with source citations for every number. AYVES is open-source financial intelligence, operated by BIRD BV, a Netherlands private company.
AYVES is operated by BIRD BV. Corporate governance and legal responsibility sit with the BV. The tool is maintained by a small operating team inside BIRD BV under a single owner. The methodology paper carries no named authors and no named endorsers; §9.2 of the methodology explains the design choice.
A public filing goes in. The engine reads the tables, the notes, and the fine print as one document.
Every figure is pinned to the page, row, and character span it was read from. The citation is part of the number.
Each jurisdiction is scored on six indicators against published, versioned thresholds. The rules are named and citable.
A read raises questions worth asking. The verdict is left to a qualified professional who can follow the citations.
AYVES operates commercially under BIRD BV. Infrastructure and language-model costs are governed by a configurable monthly ceiling with a kill-switch that pauses the anonymous tier if spend crosses that ceiling. A free account tier gives genuine research access; a paid tier adds advanced features. Tier scope and pricing are not permanent promises and are subject to change.
AYVES does not lean on endorsements or on named reviewers. The tool proves itself through published results a reader can check.
Every value on screen traces to a character span in the filed PDF. Open the citation, read the source.
Every threshold cites a published academic or regulatory source. The derivation and citations are held in an internal methodology reference that pins each analysis to a specific commit hash.
Every rating is a deterministic pure function of the extracted values and the published thresholds. Walk the code at src/server/analysis/.
Every deploy candidate runs against a fixture set of real filings under six ship-block metrics. The corpus is public at /samples.
Every analysis writes a reproducibility manifest with the per-agent prompt hash, model, temperature, and seed. Immutable via an append-only Postgres trigger.
- Language is hedged in six languages by a regex blocklist and a second-pass narrative scorer. Accusatory framing is refused before it reaches a public surface.
- Every threshold change requires a linked methodology commit in the same session. The admin surface refuses a write whose supplied hash matches the currently stored one.
- Cross-provider divergence sampling runs on a random slice of analyses plus on any analysis carrying a RED jurisdiction rating. A divergence above threshold routes the analysis to a needs-review path before it reaches the user.
- Corrections carry a ten-business-day response target from the moment the requesting entity is verified. The corrections page sets out the process; the queue and outcomes are audit-trailed.
For a numeric correction to an AYVES analysis, use the corrections process. For anything else, email hello@ayves.ai. Governance enquiries (methodology references from external work, press) go to the same address.