← TERMS

Chain of Thought (CoT)

Linear legible reasoning stream before bind — one reasoning topology, L-β assert

Chain of thought (CoT) is a linear reasoning topology — the legible symbolic stream an agent emits step by step before a commitment boundary crossing: inner speech, token trace, plan steps, tool calls in order. In Thinking by Writing it is L-β: tier-one assert when captured honestly, not tier-zero truth about the world and not commit until send or act.

When reasoning branches, merges, or broadcasts to other agents, the topology is no longer a chain — use graph-shaped assert types and honest topology metadata; see Reasoning Topology.

Each projected token conceptually indexes a micro-decision; production systems batch the stream — see Thinking by Writing (Capturing L-β on the log).

Envelope: correlation, causation, context

See shared tier-one envelope and Thinking by Writing: correlationId, causationId (links conversation messages), contextRef (MCP / harness snapshot).

Why capture it

A preserved CoT trace is explainability — audit, negative feedback for training and harness tuning, compressible patterns over time. Non-linear shapes: Reasoning Topology. Literature anchor: Wei et al. 2022 CoT.

Failure modes

Deleting CoT before ResponseSent or a human signature is stream-to-summary collapse (second-hand commitment). Treating CoT as measure or commit without a boundary crossing is inference-as-fact. Visible CoT may not reflect sub-symbolic computation — capture-chain distortion at the narration layer.

Pre-bind freeze at automated binds: ReasoningAsserted.