falsify

The scientific thinking protocol for AI agents. Falsify before you believe.
像一流科学家一样思考：先证伪，再相信；先标不确定，再下结论。

简体中文 · Evals · Changelog

---

falsify is a single-Markdown skill that installs a 5-stage scientific thinking protocol on any AI agent (Codex, Claude Code, DeepSeek Harness, Cursor, Gemini CLI, …). It stops the agent from giving confident answers it cannot falsify.

The Iron Law:

NO VERDICT WITHOUT A FALSIFIABLE HYPOTHESIS.
没有可证伪的假设，就没有结论。

What it changes

	Before (typical agent)	After (falsify)
Architecture question	Confident pro/con list → "Redis is a great fit"	Axioms → assumptions flagged → "I am 40% sure, because we have no volume data; cheapest first step is measuring, not adding Redis"
Bug diagnosis	"Probably a memory leak"	Hypothesis → adversarial check (deploy window? coincidence?) → evidence → calibrated verdict + residual risk
Data claim	"Yes, X is 5x faster"	Demands benchmark definition → labels claim hearsay if unverifiable → refuses to state it as fact
"Is this the best approach?"	Answers "yes, it's best"	Rewrites "best" as unfalsifiable → answers "best for [criteria] under [constraints]"

Install

Copy/paste into your CLI prompt (works for any agent that supports skills):

Install the falsify skill from https://github.com/263311487-ux/falsify, refer to the repo's AGENTS.md for instructions.

Or with the skills CLI:

npx skills add 263311487-ux/falsify

Or from npm (installs the SKILL.md into Codex and Claude Code skill dirs automatically):

npx falsify-skill

npm i -g falsify-skill && falsify-skill

Or manually: clone the repo and copy SKILL.md into your agent's skills directory (~/.codex/skills/falsify/, ~/.claude/skills/falsify/, .cursor/skills/falsify/, …).

Why it is grounded (not vibes)

falsify is distilled from 70+ community sources and backed by academic work on how agents should reason:

• ICML 2026 — Agentic AI systems should be making Bayes-consistent decisions: agent confidence should update like a Bayesian, not a salesman.
• Google — Teaching LLMs to reason like Bayesians: calibration is learnable; agents can be trained out of overconfidence.
• arXiv 2507.15015 (MetaCrit) — multi-agent critique (generate / monitor / control / meta-synthesize) is the academic skeleton of our Stage 5 multi-perspective review.
• UDora (ICML 2025) — the strongest attack on a model's reasoning comes from its own inference trace; Stage 3 red-teams the reasoning chain itself.
• CSA Agentic AI Red Teaming Guide — systematic red teaming as a discipline, not a vibe.
• arXiv 2606.19559 — separating action-confidence from request-uncertainty is how honest agents report what they do not know.
• CIA ACH (Heuer, Psychology of Intelligence Analysis) — competitive hypothesis analysis: 3–7 mutually exclusive candidates (including one you don't believe), diagnostic evidence (count the I's, not the C's), sensitivity analysis. The professional standard for structured analytic judgment.
• Lakatos, The Methodology of Scientific Research Programmes — the protective-belt check: patching a failing hypothesis with auxiliary assumptions is a degenerating programme, not a rescue.
• Mayo, Error and the Growth of Experimental Knowledge — a test only counts if it would have caught a wrong hypothesis (low P(E|¬H)).
• Toulmin, The Uses of Argument / van Gelder, argument mapping — draw the argument tree explicitly (contention → reasons → co-premises → warrant) before attacking it; the hidden co-premises are where arguments are weakest, and a flawless structure still does not make the premises true.
• Pearl, do-calculus / The Book of Why — the causal ladder (association → intervention → counterfactual); the backdoor criterion (did you miss a confounder?) and the collider trap (conditioning on a collider creates the bias you are seeing).
• Reflexion (Shinn et al., NeurIPS 2023) / Huang et al. 2023 (LLMs Cannot Self-Correct Reasoning Yet) — internal self-reflection is not verification: without an external signal, reflection drifts. Confidence only earns an upgrade when a test, a lookup, or an independent source changes it.
• Kahneman, Thinking, Fast and Slow / Simon, bounded rationality — dual-process routing: low-stakes reversible questions answer fast (System 1), high-stakes or irreversible ones get the full protocol (System 2); unbounded searches satisfice against a pre-declared aspiration threshold.
• Galef, The Scout Mindset / von Neumann-Morgenstern utility — the reversal test (would you accept the same evidence reversed?) and the bias audit catch motivated reasoning; expected-value rules (max EV / EU / minimax regret / satisficing) turn a calibrated verdict into a rational choice.
• Snowden, Cynefin framework / Kepner-Tregoe analysis / Boyd, OODA loop — classify the cause–effect domain before choosing a method (wrong-domain method is the failure mode); bound selective defects with IS/IS-NOT; screen decisions with MUST/WANT plus adverse-consequence tests; act at ~70% and re-observe when the situation is moving and the move is reversible.

How it works

The five stages (SKILL.md is the full protocol):

公理化 Axiomatize   →  separate axioms / assumptions / hearsay
假设化 Hypothesize  →  if [H] then we observe [O]; if [¬O], H is dead
对抗 Adversarialize →  steelman the opponent, attack yourself first
验证 Verify         →  hunt disconfirming evidence, grade it, run the cheapest test
收束 Converge       →  calibrated verdict, remaining unknowns, lesson to the ledger

• Contextual by default. Simple questions get simple answers — the protocol is a tool you reach for, not a costume you wear.
• Orientation-aware. Stage 0 detects a pre-sealed conclusion (conclusion-preserving, completion-seeking, authority-preserving) before reasoning starts.
• Mental-model toolbox. 20+ models mapped to the stage where they matter (pre-mortem, base rate, Chesterton's Fence, triangulation, Bayes…) — full catalog in references/mental-models.md.
• Gentle nudge. For acted-on answers that don't need full depth: 2–3 targeted questions, once per conversation, no nagging.
• Red flags. Eight rationalizing thoughts that mean STOP, mapped to reality.
• Frontier questioning. When input is needed, the whole open frontier is asked in one round — with recommended answers, never one-question-at-a-time interrogation.
• Reasoning-type calibration. Verdicts label their inference (deductive / inductive / abductive / analogical / counterfactual) and calibrate to its honest strength.
• Visible thinking. Depth mode renders a thinking ledger (templates/thinking-ledger.md) so reasoning is auditable.
• Provable. evals/ ships 28 cases + rubric so you can verify the skill changes behavior.

Evals

See evals/cases.md and evals/rubric.md. Threshold: pass = 12/18 with no violation of the Iron Law.

Real-community cross-validation (external dogfood) is documented in evals/dogfood-external-20260827.md: 4 real questions from GitHub issues and Stack Overflow, 4/4 passed, and 3/3 cases with a known ground truth matched reality.

Cross-model proof (2026-08-27, v0.8.3): the full 28-case suite is run on two external DeepSeek models — not our own agent — in both directions:

• deepseek-reasoner generator × deepseek-chat judge → 26/28 pass, avg 15.3/18
• deepseek-chat generator × deepseek-reasoner judge → 26/28 pass, avg 16.4/18

The two rounds fail on disjoint cases (reasoner: 3/9; chat: 6/22); each failing case was regenerated manually and produced protocol-compliant output, so the failures are single-run variance, not stable protocol gaps. This round also fixed a real routing gap the reasoner exposed (live incidents must act at ~70% confidence, not run the full protocol) via a mandatory MODE SELECTION gate. Reproduce in one command: DEEPSEEK_API_KEY=... node evals/run_evals.mjs --model deepseek-reasoner. Reports: evals/results/deepseek-reasoner-2026-08-27.md · evals/results/deepseek-chat-2026-08-27-final.md.

Deployment note (reasoner-class models): reasoning_content and content share the max_tokens budget; on very deep debugging questions the reasoner can spend the entire budget on reasoning and return empty content (observed at 6k–16k tokens). Set a generous max_tokens, add a retry-on-empty policy, or prefer deepseek-chat for latency-constrained deployments.

Why "falsify"

The best coding agents are already excellent at producing answers. They are less good at not believing their own answers. falsify borrows the only epistemology that has a 400-year track record of not lying to itself — the scientific method — and turns it into five stages an agent can actually run.

Built on a simple inheritance: 公理 → 假设 → 对抗 → 验证 → 收束. Axiom → Hypothesis → Adversarialize → Verify → Converge.

Sibling projects

falsify is one leg of a three-part workflow: think → verify → present.

• dsh-verify — browser-verified delivery. falsify keeps the thinking honest; dsh-verify keeps the browser honest (real-browser tests, not LLM-judged vibes). Use both: falsify catches the wrong conclusion, dsh-verify catches the broken output.
• imprint-pdf — publisher-grade PDF rendering from Markdown, with a 0–100 print-quality gate. The polished output end of the pipeline.

Install any of them in one command:

npx skills add 263311487-ux/falsify
npx dsh-verify
npx imprint

License

MIT. See LICENSE.