Fidelity

Providers attach data to what a model produces and expect it back: a signature on a thinking block, a phase on a message, an encrypted copy of the reasoning. Your code has no use for it, and the next request fails or degrades without it.

MMSP keeps that data in one place, the fidelity field of a content item.

{"type": "thinking.done", "thinking": "Let me check", "fidelity": {"signature": "EuYB..."}}

What it holds

Provider	Item	Fidelity
Anthropic	thinking.done	The signature of the thinking block, or the data of a redacted one
OpenAI Responses	thinking.done	The channel the reasoning arrived on, encrypted_content, and what else the server demands back
OpenAI Responses	text.done	The phase of the message
Chat Completions	thinking.done	The reasoning_field the server used: reasoning_content or reasoning
Chat Completions	thinking.done	A gateway's encrypted reasoning block: encrypted_content, with its format and the block's reasoning_id
Gemini	any item	The thought signature

The rules

• It is opaque. fidelity is an arbitrary JSON object. Pass it back unchanged; never modify it and never drop it.
• It belongs to the item. Keep it with the item it arrived on, in the order the items arrived. The stateful call does both.
• It appears once per item in a stream. At most one .delta of a group carries it, and the .done item carries the same value. The fragment that brings it may be otherwise empty.

Moving a conversation between providers

The content of a history reads the same for every provider, so a conversation can continue on another model. Fidelity is the part that does not travel: a signature means something only to the provider that issued it.

Two limits come from the providers themselves:

• Gemini signatures are bound to the endpoint that issued them. A history recorded with a Gemini API key cannot move to a Vertex AI key with its signatures, or the reverse.
• Gemini expects a signed thought in a turn that called tools. When the turn came from another provider, MMSP opens it with the placeholder signature Google documents for thoughts the model did not produce, so the request is accepted.