# Appendix B — Evidence and Claim Language
# Book: Agent Architectures: Advanced Strategies for Intelligent LLM Systems
# Chapter URL: https://programmer.ie/books/agent-architectures/81-chapter/
# Prose fingerprint (sha256, normalized): 09e533a402d13b5d4ab3709a98ef3605ea2ae3ca9d0177de5b98c3217bdd38ff
# Snapshot generated at build time by Hugo. This is the complete rendered prose
# of the chapter; it is not a transcript of the source file.

The ladder E0 conceptual / illustrative — a sketch. Says "consider", never "measured". E1 sourced external claim — a primary paper or doc, cited. Says "X reports". E2 executable example — code or schema that ran once. Says "we ran". E3 measured experiment — protocol written and frozen before the run, qualified observer, frozen result. E4 replicated result — E3 repeated across conditions, with negative arms kept. E5 mechanically verified property — deterministic check over a fixture set. Wording rules A critic preferring a candidate is not “better”. The required form is: under rubric R, candidate B scored higher than A in N seeded trials. Agreement is not verification. Retrieval is not exposure; exposure is not influence; influence is not utility. Each link needs its own evidence. Revision is a candidate until adjudicated. Scores are reported with disagreement, never as a single accuracy number. Allowed shorthands: observed (E2+), measured (E3+ with numbers and budget), supported (E3+ beating a baseline), suggests (single-condition E3), consistent with (alternatives open), did not distinguish, inconclusive, not tested, speculative (E0, labeled). Forbidden without E3 or higher: improves, helps, better, verifies, proves, preserves intent, earns overhead. Failure is a result Inconclusive and negative runs are retained, not re-rolled. Every excluded run carries a category (target / non-target / agent / environment / observer / evaluation failure, protocol violation, inconclusive), a reason, and an evidence pointer. Infrastructure failure is never classified as model failure.