Rewrite all three READMEs in plain language

[mboss37]

Summary

Follow-up to #1 and #2 (both merged): a readability pass on the root README and both example READMEs.

The previous version explained the safety mechanisms accurately but assumed the reader already knew terms like Bonferroni deflation, sign test, reward-hacking, and holdout discipline before they hit the first screen. This rewrite:

• Opens every README with a jargon-free explanation of what the loop actually does, before any technical vocabulary.
• Adds a glossary in the root README that defines the recurring terms once (gate, holdout, candidate budget, fails closed, reward-hacking, guardrail/floor, sandbox, statistical significance test) instead of re-explaining them inline everywhere.
• Reworks the failure-mode table and example write-ups into plain sentences.
• Keeps the same technical content, structure, and section order; this is a wording pass, not a scope change.
• No em dashes, no filler/AI-buzzword phrasing (checked with a targeted grep).

Test plan

• Repo-wide grep for em/en dashes and forge/anvil/jail/blacksmith language: zero matches outside .git/__pycache__
• Targeted buzzword grep (leverage, seamless, robust, delve, elevate, etc.): zero matches
• All relative links in the three READMEs verified to resolve (../../README.md cross-links, template/sandbox/, template/INSTANTIATE.md, LICENSE, etc.)
• examples/prompt-technique-tournament: python -m pytest eval/ -q (129 passed)
• examples/latency-tuner: python -m pytest sim/ -q (90 passed)
• python3 template/tools/verdict.py self-test (all 6 checks pass)
• Guardrail Floor 2 parity check reproduced locally (byte-identical across template and both examples), matching what CI enforces

https://claude.ai/code/session_011t5gmHpHh68Jcigy5uPPEx

---

Generated by Claude Code