ai-hyp-gate
Skill9 starsChanged 6 months agoMITView on GitHub
[STATUS: review] [CONFIDENCE: high] [REVIEWED: 2026-06-02] 7-шаговый orchestrator проверки AI-сгенерированной гипотезы ДО включения в research design. Цепочка skills: source trace → novelty → falsifiability → independent expert → causal structure → pilot → documentation. На каждом шаге явный KILL signal — если шаг fail, цепочка останавливается и сообщает причину. Защита от: hallucinated sources, pseudo-novelty, unfalsifiable claims, confirmation bias, causal/correlational confusion, premature scaling, irreproducibility. Базируется на rules/falsification-ladder.md + rules/estimand-ops.md + existing skills (поиск литературы, /sci-evidence, /hypothesis-arbiter, Agent(skeptic) с Context Asymmetry). Triggers: /ai-hyp-gate, /hyp-gate, /verify-hypothesis, "проверь гипотезу через gate", "AI hypothesis gate", "7-шаговая проверка гипотезы", "kill-gate гипотезы", "верифицируй гипотезу до эксперимента", "ai hypothesis verification". НЕ для: уже подтверждённых гипотез в active experiments (→ /scientific-research), одной идеи без AI generation (→ /skeptic Mode 1), code review (→ /code-review).
In this repository164 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .codex/AGENTS.mdAGENTS.md
- agents/CLAUDE.mdCLAUDE.md
- CLAUDE.mdCLAUDE.md
- claude-md/CLAUDE.mdCLAUDE.md
- hooks/CLAUDE.mdCLAUDE.md
- insight-architectSkill
- agent-teamsSkill
- brainstormingSkill
- claim-decomposerSkill
- context-engineeringSkill
- dispatcherSkill
- git-worktreesSkill
- hd-mavp-routerSkill
- mentor-modeSkill
- reference-registrySkill
- routing-policySkill
- skill-self-updateSkill
- tdd-workflowSkill
- ab-testSkill
- academic-researchSkill
- advanced-methodology-extensionsSkill
- agent-governanceSkill
- algorithm-designSkill
- analystSkill
- archcode-genomicsSkill
- arm-cortex-microcontrollersSkill
- atomizeSkill
- backward-traceabilitySkill
- boyko-constraint-engineSkill
- boyko-epistemic-auditSkill
- boyko-goal-expansion-100Skill
- boyko-knowledge-architectureSkill
- boyko-leverage-mapSkill
- boyko-methodSkill
- boyko-preflightSkill
- boyko-project-radarSkill
- boyko-rippleSkill
- boyko-scientific-consortiumSkill
- boyko-specialistSkill
- boyko-toolcheckSkill
- boyko-triangle-auditSkill
- boyko-why-ladderSkill
- browser-actSkill
- build-autoSkill
- c4-architectureSkill
- career-prepSkill
- changelog-genSkill
- citation-managementSkill
- code-debuggingSkill
- code-materialitySkill
- codex-skepticSkill
- codex-solverSkill
- codex-tracySkill
- combinatorial-creativitySkill
- common-groundSkill
- conductorSkill
- consilienceSkill
- construct-measurement-gateSkill
- convertSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
9 rules. About 0 are common, 7 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 46
MUST complete все 7 шагов ИЛИ explicit fail с указанием шага
- Distinctive: line 47
MUST stop при первом kill signal — не пытаться "обойти" failed step
- Very specific: line 48
MUST output decision document — markdown report недостаточен, нужен decision.md с verdict
- Distinctive: line 51
Default mode: read-only audit — не запускаем эксперимент сами
- Distinctive: line 125
Write kill criterion в виде: "Если <metric> <comparator> <threshold>, гипотеза опровергнута"
- Distinctive: line 199
SUTVA: no interference, no hidden treatment versions
- Distinctive: line 202
Check noncollapsibility: если outcome heterogeneous population, prefer risk difference (RD) или restricted mean survival time (RMST) over …
- Very specific: line 231
Step 6: Run minimal test → metrics/run.json
- Distinctive: line 403
Default mode: read-only audit — выдаёт verdict + artifacts, НЕ запускает реальный эксперимент сам.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
<!-- BSV — Brief Skill View | поиск: BSV
Скил : ai-hyp-gate
TL;DR : 7-шаговый kill-gate для AI-сгенерированных гипотез до research design
Вызов : /ai-hyp-gate, "проверь гипотезу", "kill-gate гипотезы"
НЕ для : Уже подтверждённых гипотез, ручных идей без AI, code review
-->/ai-hyp-gate — AI Hypothesis Verification Gate
Purpose
Цепочка 7 проверок которая работает как защитный gate ДО того как гипотеза попадёт в research design / эксперимент.
Принцип: дешёвые проверки (Шаг 1-3, минуты) идут ПЕРЕД дорогими (Шаг 4-6, часы-дни). Каждый шаг имеет explicit kill signal — если fail, останавливаемся и сообщаем причину.
Базируется на:
rules/falsification-ladder.md(FL Steps 0, 3, 6)rules/estimand-ops.md(L0 gate)Agent(skeptic)с Context Asymmetry HARD RULE- Existing skills: поиск литературы,
/sci-evidence,/hypothesis-arbiter
Hard Rules
- MUST complete все 7 шагов ИЛИ explicit fail с указанием шага
- MUST stop при первом kill signal — не пытаться "обойти" failed step
- MUST output decision document — markdown report недостаточен, нужен
decision.mdс verdict - Каждый шаг = отдельный sub-process (изолированный контекст)
- Token budget: ~30-50 turns на полный gate
- Default mode: read-only audit — не запускаем эксперимент сами
Input Contract
Принимает на вход:
- Гипотеза в одном предложении (что должно быть истиной)
- Источник — кто сгенерировал (Claude / ChatGPT / Gemini / комбинация)
- Целевой эксперимент — куда планируется включить (project / experiment ID)
- Domain — биомед / физика / ML / экономика / другое
Если что-то не указано — спросить до начала Step 1.
7-Step Protocol
Step 1 — Source Trace (Трассировка источника)
Вопрос: существует ли реальная первичная публикация для каждого fact-claim в гипотезе?
Actions:
- Извлечь все specific factual claims из гипотезы (имена, числа, механизмы, ссылки)
- Для каждого claim:
Agent(explorer, prompt="Find primary source for: <claim>")ИЛИ- поиск литературы запрос с конкретными keywords
- Если AI указал DOI/PMID/arXiv ID — verify через WebFetch
- Mark каждый claim:
[VERIFIED-REAL]— primary source confirmed (URL/DOI)[INFERRED]— derivable из verified facts[UNVERIFIED]— claim не найден в literature
KILL signal:
≥1 claim helper-pivotal для гипотезы со статусом
[UNVERIFIED]после reasonable search (≥3 поисковых запросов).Verdict: HALLUCINATED SOURCE → STOP. Report which claim, what was searched, что не найдено.
Output: source_register.md (claim / status / source URL / search trail)
Step 2 — Novelty Check (Проверка новизны)
Вопрос: не является ли гипотеза перефразировкой известного результата?
Actions:
- Семантический поиск:
- поиск литературы с paraphrased queries (3-5 вариаций)
- Semantic Scholar для recent (последние 5 лет)
- Google Scholar для broader coverage
- Pre-work check (наш собственный архив):
- поиск по своему архиву прошлых экспериментов и отклонённых гипотез (например,
grep -ri "<keyword>" experiments/ null_results/ parked/), если такой архив есть
- поиск по своему архиву прошлых экспериментов и отклонённых гипотез (например,
- Если есть похожие найденные работы — оценить delta:
- Same method, same population, same outcome → pseudonovelty
- Same idea, different domain → transfer hypothesis (может быть valid)
- Same outcome, different mechanism → alternative explanation (valid)
KILL signal:
Найдена published работа с same idea AND same approach AND comparable rigor.
Verdict: PSEUDONOVELTY → STOP. Report найденный source, объяснить почему gate kill.
Если в null_results/ или parked/:
Verdict: REJECTED PREVIOUSLY → STOP. Прочитать decision.md из archive прежде чем re-attempt.
Output: novelty_check.md (queries / matches / delta analysis / verdict)
Step 3 — Falsifiability (Критерий Поппера)
Вопрос: какой конкретный экспериментальный результат опровергнет гипотезу?
Actions:
- Применить FL Step 0 из
rules/falsification-ladder.md:- Write the claim ОДНИМ предложением
- Identify измеримые quantities
- Write kill criterion в виде: "Если
<metric><comparator><threshold>, гипотеза опровергнута"
- Write Micro 4-liner:
Question type: [ ] descriptive [ ] predictive [ ] causal Claim: <one sentence> Check: <how to verify in 7 days> What this does NOT mean: <boundary> - Sanity check:
- Kill criterion должен быть в abstract numbers, не в "если эксперимент не подтвердит"
- Threshold должен быть указан ДО эксперимента
- Должно быть possible получить kill result — иначе гипотеза unfalsifiable
KILL signal:
Невозможно сформулировать concrete numeric kill criterion. ИЛИ kill criterion = "expert judgment" / "if it doesn't work" / "if results are inconsistent" (subjective).
Verdict: NOT FALSIFIABLE → STOP. Гипотеза не научна в текущей формулировке.
Output: claim.md (FL Step 0 артефакт + Micro 4-liner)
Step 4 — Independent Expert Review (с Context Asymmetry)
Вопрос: что говорит независимый skeptic не знакомый с историей вопроса?
Actions:
- Подготовить минимальный input для skeptic:
source_register.md(Step 1 output)claim.md(Step 3 output)- НИЧЕГО ИЗ session reasoning chain
- Spawn skeptic с Context Asymmetry HARD RULE enforced:
Agent(skeptic, prompt=" Falsification test for claim: <claim.md content> Source register: <source_register.md> Task: Generate 3 test cases that would FALSIFY this claim. Specify exact commands or experiments. Report: CONFIRMED / FALSIFIED / NEEDS-REAL-DATA / WEAKENED. Do NOT consider how the hypothesis was generated. Only: does the claim hold? ") - Для biomed/clinical/engineering domains добавить domain-specific skeptic:
- (опционально) отдельный domain-specific reviewer, если он есть в вашей установке
KILL signal:
Skeptic verdict =
FALSIFIEDИЛИ skeptic находит fundamental flaw в methodology.Verdict: SKEPTIC REJECTED → STOP. Report skeptic's reasoning + suggested falsification tests.
WEAKEN signal:
Skeptic verdict =
WEAKENED— гипотеза держится но с serious caveats.Action: добавить caveats в
claim.md, NARROW scope гипотезы, продолжить с уменьшённым confidence.
Output: skeptic_review.md (skeptic's verdict + 3 falsification tests)
Step 5 — Causal Structure (EstimandOps L0)
Вопрос: заявляет ли гипотеза причинность? Если да — обоснован ли механизм?
Actions:
- Apply
rules/estimand-ops.mdL0 gate:- Descriptive: "What is X in population P?" → no causal interpretation
- Predictive: "What will X be for new case?" → no causal interpretation
- Causal: "What would change if we did A?" → требует causal layer
- Если classification = causal:
- Draw DAG (nodes + edges)
- Check 4 identifiability assumptions:
- Consistency: Y = Y^a when A=a
- Positivity: P(A=a|L) > 0 для всех a и L в support
- Exchangeability: Y^a ⊥ A | L (no unmeasured confounders)
- SUTVA: no interference, no hidden treatment versions
- Name identification strategy (randomization / IV / RD / DiD / g-formula / TMLE)
- List unmeasured confounders
- Check noncollapsibility: если outcome heterogeneous population, prefer risk difference (RD) или restricted mean survival time (RMST) over hazard ratio (HR) или odds ratio (OR)
- LLM-specific check: causal parrots warning — LLMs systematically conflate correlation with causation. Re-read original claim — действительно ли он implies causation?
KILL signal:
Заявлена causation НО:
- Identification assumptions нарушены AND cannot be recovered by design ИЛИ
- Нет DAG ИЛИ
- Нет identification strategy named
Verdict: UNIDENTIFIABLE CAUSAL CLAIM → STOP. Choices: redesign as descriptive ИЛИ change study design.
DOWNGRADE signal:
Гипотеза заявляет причинность но evidence supports только correlation.
Action: rewrite claim.md как associational, добавить "what this does NOT mean: causation".
Output: estimand.md (full canvas: population/intervention/comparator/endpoint/summary measure/MCID/ICE) + optional dag.md
Step 6 — Pilot Experiment (Минимальный тест)
Вопрос: воспроизводится ли expected behavior на small real data перед полным запуском?
Actions:
- Apply FL Standard Ladder Steps 3-6 из
rules/falsification-ladder.md:- Step 3: Positive control (known-good input) →
controls.md - Step 4: Negative control (known-bad input) →
controls.md - Step 5: Define baseline (existing method или null model) →
baselines/<id>.json - Step 6: Run minimal test →
metrics/run.json
- Step 3: Positive control (known-good input) →
- Pilot rules:
- На REAL data (НЕ synthetic) —
[VERIFIED-REAL]mandatory - Sample size минимум для statistical power detection effect ≥MCID
- Pre-registered analysis plan (что считаем success ДО запуска)
- На REAL data (НЕ synthetic) —
- Result classification (FL Step 8):
- PROMOTE — success, full experiment justified
- REPEAT — inconclusive, iterate с modifications
- REJECT — pilot fail → kill gate
- ARCHIVE — valid but deprioritized
KILL signal:
Pilot не воспроизводит expected behavior на real data. ИЛИ negative control unexpectedly succeeds.
Verdict: PILOT REJECTED → STOP. Это model boundary / data issue. Report что обнаружено, save в null_results/.
SCOPE-REDUCE signal:
Pilot succeeds только на subset (specific population/condition).
Action: narrow гипотезу до tested scope, добавить boundary conditions в caveats.md.
Output: pilot_result.md + metrics/run.json + controls.md + baselines/<id>.json
Step 7 — Documentation & Audit Trail
Вопрос: есть ли воспроизводимый аудит всей цепочки?
Actions:
- Создать experiment directory:
experiments/<YYYYMMDD-slug>/ - Зафиксировать:
- Original AI prompt (с timestamp)
- AI model + version (название модели и версия)
- Generation date
- Все 6 предыдущих outputs (
source_register.md,novelty_check.md,claim.md,skeptic_review.md,estimand.md,pilot_result.md) - Final decision (
decision.md)
- Final decision template:
# Decision — <hypothesis slug> **Date:** YYYY-MM-DD **Question type:** descriptive / predictive / causal **AI source:** <model + version + prompt> ## Gate Results | Step | Verdict | Artifact | |---|---|---| | 1. Source trace | PASS / FAIL | source_register.md | | 2. Novelty | PASS / FAIL | novelty_check.md | | 3. Falsifiability | PASS / FAIL | claim.md | | 4. Skeptic | PASS / WEAKEN / FAIL | skeptic_review.md | | 5. Causal | PASS / DOWNGRADE / FAIL | estimand.md | | 6. Pilot | PROMOTE / REPEAT / REJECT | pilot_result.md | ## Final Decision PROMOTE TO RESEARCH DESIGN / REJECT / ARCHIVE ## Caveats <list any WEAKEN/DOWNGRADE/SCOPE-REDUCE signals from gate> ## What this does NOT mean <explicit boundaries>
KILL signal:
Нет originating prompt OR нет AI model+version OR нет gate decisions.
Verdict: UNAUDITABLE → STOP. Reproducibility broken at methodology level. Fix documentation before promoting.
Output: experiments/<id>/decision.md + all gate artifacts
Flow Diagram
AI-generated hypothesis
↓
Step 1: Source trace ────[FAIL: HALLUCINATED]──> STOP
↓
Step 2: Novelty ────[FAIL: PSEUDONOVELTY]──> STOP
↓ ────[REJECTED in null_results/]──> STOP
Step 3: Falsifiability ──[FAIL: NOT FALSIFIABLE]──> STOP
↓
Step 4: Skeptic ────[FAIL: FALSIFIED]──> STOP
↓ ────[WEAKEN: scope narrow]──> add caveats, continue
Step 5: Causal ────[FAIL: UNIDENTIFIABLE]──> STOP
↓ ────[DOWNGRADE: not causal]──> rewrite as associational, continue
Step 6: Pilot ────[FAIL: REJECTED]──> save null_results/, STOP
↓ ────[SCOPE-REDUCE: subset only]──> narrow, continue
Step 7: Documentation ──[FAIL: UNAUDITABLE]──> STOP
↓
PROMOTE TO RESEARCH DESIGNFinal Report Format
# /ai-hyp-gate Result — <hypothesis>
## Verdict: PROMOTE / REJECT / ARCHIVE
## Gate Trail
1. Source trace: PASS — verified N claims via lit search
2. Novelty: PASS — no prior work found in N queries
3. Falsifiability: PASS — kill criterion: <numeric threshold>
4. Skeptic: WEAKEN — added caveats X, Y
5. Causal: DOWNGRADE — rewritten as associational
6. Pilot: PROMOTE — pilot N=100 confirms effect at p<0.01
7. Documentation: PASS — audit trail in experiments/<id>/
## Artifacts
- experiments/<id>/source_register.md
- experiments/<id>/novelty_check.md
- experiments/<id>/claim.md
- experiments/<id>/skeptic_review.md
- experiments/<id>/estimand.md
- experiments/<id>/pilot_result.md
- experiments/<id>/decision.md
## Next Step
<one concrete action — design full experiment OR pivot OR archive>What NOT to produce
- ❌ Markdown отчёт без artifacts на disk
- ❌ "Probably falsifiable" — нужен concrete numeric kill criterion
- ❌ Skip skeptic step "потому что и так понятно" — Context Asymmetry mandatory
- ❌ Pilot на synthetic data, claim PROMOTE —
[VERIFIED-REAL]required - ❌ Skip Step 5 causal check для "не causal hypothesis" без явной classification
Anti-loopholes
EXPLOIT 1 — Skip source trace: "AI is reliable, sources implied". Fix: Mandatory verification. Если AI не указал sources, generate specific claims и search для каждого.
EXPLOIT 2 — Fake novelty: "I haven't seen it" = novel. Fix: 3+ search queries обязательны. Pre-work check в null_results/ обязателен.
EXPLOIT 3 — Soft kill criterion: "если не работает". Fix: Numeric threshold ОБЯЗАТЕЛЬНО. "Не работает" = not falsifiable.
EXPLOIT 4 — Friendly skeptic: дать skeptic полный контекст. Fix: Context Asymmetry HARD RULE — skeptic видит только artifacts, не reasoning chain.
Companion Skills (mapping на шаги)
| Step | Skill | Mode |
| 1 | поиск литературы | source verification |
| 1 | Agent(verifier) | claim verification |
| 2 | поиск литературы | semantic novelty search |
| 2 | аудит источников | если AI cited specific papers |
| 3 | rules/falsification-ladder.md | FL Step 0 spec |
| 4 | Agent(skeptic) | Context Asymmetry red team |
| 5 | rules/estimand-ops.md | L0 classification + DAG |
| 6 | /scientific-research | pilot + controls + baseline |
| 6 | /sci-evidence BUILD | pilot design |
| 7 | rules/falsification-ladder.md | FL Steps 8-11 documentation |
Reference
rules/falsification-ladder.md— FL protocol (Step 0 → Step 11)rules/estimand-ops.md— L0 gate + causal layerrules/audit-verification-gate.md— verification thresholdsagents/skeptic.md— Context Asymmetry HARD RULEexperiments/_template/— directory skeleton
Final Notes
Default mode: read-only audit — выдаёт verdict + artifacts, НЕ запускает реальный эксперимент сам.
Implementation: если PROMOTE verdict → user явно запускает следующий шаг через /scientific-research или Agent(builder). Gate сам не scales до full experiment.
Time budget: typical full gate = 30-90 минут (Steps 1-5 быстро, Step 6 pilot занимает основное время).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

