Methodology for public calibration.

Expert-designed tasks cover Accounting Knowledge, HGB Reporting, IFRS Reporting, Financial Analysis, and Accounting Capstone. The active evaluation material stays hidden; the public sample task is synthetic and excluded from scoring.

Each task fixes its input, required output shape, allowed tools, and scoring basis before a run begins.

Deterministic checks are used where exact fields, calculations, and structured outputs can be evaluated directly. Accounting Knowledge uses hybrid scoring; deterministic checks and AI-judged rubric components contribute to its suite result.

No overall score or cross-suite aggregate is published.

A public result represents one completed, versioned run under a recorded model and provider configuration. The current public runs use no tool access.

Comparisons remain configuration-specific. Single runs; no statistical ranking.

Hybrid runs use an AI judge only for rubric-scored components that are not reduced to deterministic checks. Judge identity, usage, and retry aggregates are attached to the sanitized run evidence where the contract provides them.

AI-judge output does not change the provisional release status.

Hidden evaluation material reduces direct task exposure but cannot prove that a model has never encountered related concepts. Public demonstration tasks are kept separate from the active hidden set and are not scored.

Current results cover 4 suites and 240 represented tasks. Accounting Capstone is defined but has no public results. The release does not establish statistical rank certainty or standardized comparability.

It does not claim complete coverage of professional financial work or performance outside the recorded configuration.

The current scope does not test XLSX or workbook artifacts.