Thinking Hash System

Description

Server-side reasoning-preservation mechanism: Anthropic computes a hash of the true thinking trace, stores the trace in a database, and returns only the hash to the client. On subsequent turns the client echoes the hash; the server looks up and re-injects the real reasoning for LLM continuation. Consequence: local JSONL transcripts contain only hashes, so any analysis of local transcripts systematically undercounts thinking depth. Zero-hash turns indicate genuinely zero-reasoning regardless of effort setting.

Key claims

Relations

Sources

src-20260419-be936b1b7d4c