skills/crypto-benchmarking/references/evidence-and-tables.md

# Evidence and manuscript tables

Resolve the requested paper entrypoint and its live labels; printed numbers and old script names can drift. Inventory exactly the requested rows, formulas, units and variants before choosing a reporter. A supported recorder may omit an intended baseline; a complete command does not imply complete table coverage.

For each value retain its source, status, configuration, numerator/denominator, phase definitions and transformation formula. Keep absent/failed/unsupported values distinct from zero. Displayed rounded rates support arithmetic on displayed values, not reconstruction of hidden full-precision measurements.

Validate schemas, units, unique repetition IDs, warmup markers, expected outputs and timer compatibility before aggregation. Preserve raw observations; write derived summaries separately. Do not silently concatenate split runs or mixed cohorts because filenames look compatible. Record source/build/configuration compatibility and any host/order differences that affect the intended comparison.

Maintain separate objects for raw evidence, candidate analysis and published claims. An audit can report drift without editing a manuscript; a requested manuscript update permits its scoped edits. An evidence refresh does not automatically authorize changing frozen sections or substituting stronger claims.

Distinguish reanalysis of historical records from rerunning their implementation. Hashes can identify historical content without providing the old source or executable needed to rebuild it. A fresh run cannot retroactively validate an unattributed historical number, and changing defaults cannot relabel previous observations.

When updating claims, check the implicated raw data, analysis and current source, retaining parameter/sampler qualifications. For authorized manuscript edits, reconcile prose/captions/formulas and inspect affected output when layout or references change. Report arithmetic consistency, provenance, functional coverage and security support separately.