hippoagent-validate-claimlisted

Before asserting any verifiable factual claim (year/number/attribution/citation/date), call `hippo_validate_claim(claim)` to check it against persistent memory. Triggers when about to say things like "X was born in YEAR", "Y won the Nobel in YEAR", "Z released VERSION in MONTH", "W said EXACT_QUOTE", "this library/paper/repo is from ORG", or any other claim containing a Capitalized name + a specific year/number/version/org. The tool returns verdict ∈ {supported, contradicted, unknown} + an advice string in Italian (e.g. "in memoria 1987 (fact f_id), NON 2014 — controlla prima di affermare"). If verdict=contradicted, STOP and re-read the evidence_facts; do NOT assert the claim verbatim. If verdict=unknown, hedge ("non ricordo con certezza", "credo ma verifica"). If verdict=supported, proceed. Zero LLM cost (deterministic, sub-100ms). Anti-confabulation gate — born from live confabulations 2026-05-14 (Tonegawa Nobel 1987→2014, Anthropic Skills Oct 2025→Oct 2026, LightRAG HKUDS→HKUST).
aureliocpr-ctrl/hippoagent · ★ 0 · AI & Automation · score 72

Install: claude install-skill aureliocpr-ctrl/hippoagent

# HippoAgent — anti-confabulation gate (hippo_validate_claim) Prima di asserire un fatto verificabile, chiama `hippo_validate_claim`. È un gate **deterministico, zero LLM call, sub-100ms** che intercetta i pattern di confabulazione più frequenti: cambio di anno, numero, versione, attribuzione, organizzazione. ## ⚠️ Quando ATTIVARE (trigger) Il tool va chiamato **prima** di pronunciare una frase factual che contiene almeno **2 token salienti** (nome Capitalized + anno/numero, o due nomi Capitalized). Esempi tipici: - "Tonegawa won the Nobel Prize in **1987**" - "Anthropic released Skills in **October 2025**" - "LightRAG is from **HKUDS** (HK Data Science group)" - "Newton wrote **Principia** in **1687**" - "HippoRAG paper appeared at **NeurIPS 2024**" - "GPT-4 was released in **March 2023**" - "Sonnet 4.6 said X" ← attribuzione di citazione - "Cycle #51 introduced narrative episodes" ← versione/release interna **NON serve** per claim generiche ("Tonegawa is a researcher", "Python is a language"): meno di 2 token salienti = il tool risponde `unknown` di default. ## Flow ``` 1. Pre-flight: hippo_validate_claim(claim="<la frase factual completa>") 2. Read verdict: payload.verdict ∈ {supported, contradicted, unknown} 3. Act: ┌─ supported → procedi, asserisci la frase ├─ contradicted → STOP. Leggi payload.advice + payload.evidence_facts. │ Re-leggi i fact con hippo_facts_list o hippo_facts_search. │ Riformula con il valore