Spaces:

---

qgallouedec

/

tito

like 3

Running

Fills in the §6 "Prefix preservation" TODO that the merged comparison section forward-references.

• States the property precisely: rendering a conversation with appended tool messages must
  extend the render without them, token-for-token — and only for tool messages. User /
  assistant / system turns are under no such obligation.
• Includes the 12-line is_prefix_preserving property test.
• Adds a real audit table run fresh across nine open-weights families: Qwen2.5, Qwen3,
  Qwen3-VL, Llama 3.1, Llama 3.2, DeepSeek-V3, GLM-4.5, GPT-OSS, SmolLM3. Eight of nine pass
  out of the box; Qwen3 is the lone failure.
• Explains the Qwen3 cause ({%- if loop.last or (not loop.last and reasoning_content) %} flips
  a prior assistant turn's rendering once a tool message makes it non-last) and gives the verified
  one-line Jinja patch. Confirmed: with the patch the property holds, and ordinary chat
  rendering stays byte-identical (the conditional only governs assistant turns after the last
  user message).
• Adds scripts/audit_prefix.py to reproduce the table.

No people or organisations are named in the prose. Remaining TODOs (History rewriting, The honest
edges, The right primitive) are unchanged.