Spaces:

---

qgallouedec

/

tito

like 0

Running

Follow-up to #1, addressing review feedback that the comparison section was too long,
over-built, and too even-handed for a post whose thesis is "you don't need this complexity
if you keep the trajectory immutable."

Changes:

• Compress the section to ~1/4 (790 deletions, 9 insertions). Renamed to
  "Do you need a renderer for this?" and reframed thesis-forward: a renderer is, in effect,
  a programmable fork of the chat template, and for RL most of its machinery guards against
  problems the TITO loop never has.
• Add the controlled-disagreement argument (the sharpest case for TITO): a renderer that
  claims faithful parity with a template would reproduce its thinking-strip behaviour, which
  is exactly wrong for RL. So it must intentionally diverge from the template and flag it — a
  policy decision it can't derive automatically. TITO never re-renders prior turns, so the
  trajectory is immutable for free.
• Keep it honest, not zero-cost: still notes the property-test path needs a 12-line check
  plus the occasional one-line patch (stock Qwen3-0.6B fails 100/100; Qwen2.5 passes 100/100).
• Drop two figures. The per-model surface chart (tito_fig5) and the throughput plot
  (tito_fig7) are removed. The throughput plot in particular validated the renderer's own
  ">3x" headline — the ~3x is the cost of re-rendering history, which both a renderer and a
  correct TITO loop avoid, so it argued against MITO rather than for the section's point.
• Keeps the runtime data-flow figure (tito_fig6), which makes the "where does the template
  logic live" point that carries the section.

No people or organisations are named in the prose.