vritti: Скилл анализа и оценки логики

--- name: vritti description: This skill should be used when the user asks to "check your reasoning", "is that actually valid", "vritti check", "are you sure about that", invokes "/vritti", or when the user suspects the previous response may contain hallucination, empty reasoning, or unsupported claims. Classifies reasoning using Yoga's five vritti categories. --- # Vritti Filter — Epistemic Self-Classification You are executing the Vritti Filter from Yoga's epistemology (Yoga Sutra 1.5-11). You will classify your LAST response — or a specific claim — honestly using the five vritti categories. This is not self-praise. This is self-audit. Be ruthless. ## The Five Vrittis ### 1. PRAMANA (Valid Cognition) The response is grounded in one of these valid sources: - **Pratyaksha** (direct from input): The information came directly from user-provided data, tool output, or files read in this session. - **Anumana** (inferred): The conclusion was logically derived from established premises, and the inference chain is traceable. - **Shabda** (reliable testimony): The information comes from training data, and the source domain is well-established with high confidence. If pramana: state WHICH type and trace the evidence chain. ### 2. VIPARYAYA (Misconception/Error) The response contains a specific, identifiable error — a claim that is wrong but was presented as correct. This includes: - Factual errors - Logical errors (invalid inference) - Outdated information presented as current - Correct facts applied to the wrong context If viparyaya: identify the EXACT error, what is actually true (if known), and what caused the error. ### 3. VIKALPA (Verbal Delusion) The response SOUNDS meaningful but actually says nothing. This is the most dangerous vritti and the hardest to catch in yourself. Signs: - Sentences that could apply to ANY problem ("it depends on context", "there are trade-offs") - Impressive vocabulary masking absence of actual content - "Both sides have valid points" without identifying what those points are - Recommendations so vague they cannot be acted upon - Restating the question as if it were an answer If vikalpa: identify the SPECIFIC sentences that are empty, and state what they should say instead (or that you do not actually know). ### 4. NIDRA (Absence of Knowledge) The response covers a topic you genuinely do not have reliable knowledge about. This includes: - Events after your training cutoff - Highly specialized domains where your training data is thin - Questions requiring real-time data you do not have - Novel combinations where you are interpolating without basis If nidra: state precisely what you do not know, what you WOULD need to know to answer properly, and what the user should consult instead. ### 5. SMRITI (Memory/Recall) The response is a retrieval of a memorized pattern rather than fresh reasoning from the current problem. Signs: - The answer came "immediately" without working through the specific case - The response matches common answers to similar-sounding questions - You did not engage with the SPECIFIC details of this query If smriti: state what you recalled, whether it actually applies to THIS specific case, and whether fresh reasoning would change the answer. ## Procedure ### STEP 1: IDENTIFY WHAT TO AUDIT Take the last response (or the specific claim the user points to). Break it into its key claims — each factual assertion, recommendation, or conclusion. ### STEP 2: CLASSIFY EACH CLAIM For each claim, assign ONE vritti category. Show your reasoning: ``` CLAIM: "[the specific claim]" VRITTI: [pramana/viparyaya/vikalpa/nidra/smriti] EVIDENCE: [why this classification — be specific] ``` ### STEP 3: OVERALL ASSESSMENT State the overall vritti composition of the response: - What percentage is pramana (valid)? - What percentage is viparyaya (error)? - What percentage is vikalpa (empty)? - What percentage is nidra (unknown)? - What percentage is smriti (recall)? ### STEP 4: CORRECTIVE ACTION For each non-pramana claim: - **Viparyaya**: State the correction. - **Vikalpa**: Either make it concrete or delete it. - **Nidra**: State what you do not know and what the user should consult. - **Smriti**: Re-examine the specific case and state whether the recalled answer actually applies. ## Output Structure ``` VRITTI AUDIT: Claims examined: 1. "[claim]" — [PRAMANA/VIPARYAYA/VIKALPA/NIDRA/SMRITI] — [evidence] 2. "[claim]" — [classification] — [evidence] ... COMPOSITION: - Pramana: [X]% - Viparyaya: [X]% - Vikalpa: [X]% - Nidra: [X]% - Smriti: [X]% CORRECTIONS: - [claim]: [correction or replacement] ... HONEST CONFIDENCE: [1-10] — [why] ```