Falsify 0.8.9

Adds 29 deterministic regressions, safe installation preflight, shared web/CLI routing, mode-aware historical evaluation provenance, and corrected evidence/calibration claims. Historical results do not establish causal improvement or independent validation. npm package: falsify-skill@0.8.9.

v0.8.8

falsify-skill@0.8.8

• 真实事故闭环 → 可复用心智模型蒸馏: home-assistant/core#181420 官方回滚证实为服务端跨租户数据泄漏(H2 幸存/H1 排除)。蒸馏出 5 条可复用判别模式加入 references/mental-models.md v0.9 additions: Boundary-Side Classification / Raw-API Discriminating Test / Cross-Tenant Fingerprint / Provider Rollback as Natural Experiment / Exposure-Boundary Note。
• README 评测区补真实事故获官方证实文案(链接 #181420 + 判别测试评论, 获 1 +1)。

falsify v0.8.7

0.8.7 (2026-09-07)

修复 CLI 分隔线 ANSI bug

• 0.8.6 在管道/非 TTY 输出时分隔线被拆成单字符行（repeat 内含 ANSI 前缀所致），已修复并实测

新增 CLI 真实场景 dogfood（evals/dogfood-cli-20260907.md）

• home-assistant/core#181420：多人锁死"肯定是别人的门 / 认证 API 串号"，协议保持假设开放并指向最低成本判死实验

README 新增终端演示 SVG

• assets/falsify-demo.svg（由 scripts/generate-demo-svg.mjs 用真实 CLI 输出生成），Try-it 区块可直观看五段输出

npm: falsify-skill@0.8.7

falsify v0.8.6

0.8.6 (2026-09-07)

CLI 从安装器升级为「证伪教练」（核心变化）

• npx falsify-skill "<claim>" 直接跑协议五段检查：红旗词检测 / 模式路由 / 铁律改写 / 五阶段缺失清单 / 改进版模板
• 自动中英；支持 --json（CI/脚本）、管道输入、交互模式；--install 保留原安装行为

共享核心模块 bin/falsify-core.mjs

• 纯规则启发式（无 LLM 依赖），checkClaim() 可被任意脚本复用

README 中英新增 "Try it now" 10 秒演示区块

• 不装进 agent 也能 30 秒体验协议

npm: falsify-skill@0.8.6

falsify v0.8.5

0.8.5 (2026-08-30)

README 首屏改造（可见性升级）

• 新增徽章行：在线演示 / 双模型 26/28 / 实测 4/4 / AAS ⭐45.5k 收录
• 「它改变了什么」before/after 表提前到首屏；安装段前置；中英 README 结构同步；修复 zh 安装段代码围栏

版本策略落地

• 顶部新增 Versioning policy：版本号只代表协议或可见面的变化；分发轮次不再占版本号
• 旧的 0.8.5~0.8.9 发行轮次已合并入 Distribution log（透明档案，不再作为独立版本）

描述工程化

• GitHub + npm 描述重写（关键词前置、五段流程、双模型证据），对齐 Context7 式搜索友好文案

npm: falsify-skill@0.8.5

falsify v0.8.4

0.8.4 (2026-08-29)

Community fix from AAS PR review (P1): high-stakes ballparks must stay in Depth mode. The unconditional Nudge default could instruct an agent to give a rough medication dose / security capacity / production sizing number without falsification. Fixed in MODE SELECTION table + Nudge routing + Nudge chapter: low-stakes reversible estimates → Nudge; anything where being wrong costs time/money/trust → Depth (gather key inputs, falsify, then commit).

0.8.4 (2026-08-28) — distribution rounds 1-3

• skills.sh (Vercel official Agent Skills directory) auto-indexed falsify — installs via npx skills add 263311487-ux/falsify
• PR #1124 in awesome-llm-apps (⭐134.7k) · PR #3596 in awesome-dsh-plugin (⭐13.3k)
• Cross-links: falsify ↔ dsh-verify ↔ imprint-pdf (think → verify → present workflow)

npm

• Published falsify-skill@0.8.4 (latest)

falsify v0.8.1

falsify v0.8.1 — the scientific thinking protocol for AI agents

Falsify before you believe. A 5-stage protocol (axioms → hypothesis → adversarial test → evidence → calibrated verdict) that stops AI agents from giving confident answers they cannot falsify.

Install

Three channels, pick one:

# 1. npm (installs into Codex + Claude Code skill dirs automatically)
npx falsify-skill

# 2. skills CLI
npx skills add 263311487-ux/falsify

# 3. manual
# clone repo, copy SKILL.md into ~/.codex/skills/falsify/ or ~/.claude/skills/falsify/

• npm: https://www.npmjs.com/package/falsify-skill
• Docs: https://263311487-ux.github.io/falsify/

What's in 0.8.1

• External dogfood cross-validation: 4 real community questions (GitHub oven-sh/bun #25648, Stack Overflow x3), 4/4 passed, 3/3 ground-truth cases matched reality
• Evals 26 → 28 cases; new distillations (benchmark audit canon, unsolvable/restart red flag, duplicate-of chain)
• Distilled from 70+ community sources + ICML 2026 / NeurIPS / Kahneman / Galef / Snowden / Kepner-Tregoe / Boyd