xuan/memory/inference-always-cfpatch.md

---
name: inference-always-cfpatch
description: "Reporting rule — inference always uses cf-patch fast-weight update; report only the cf-patch score, never the \"vanilla\" no-TTT column"
metadata: 
  node_type: memory
  type: feedback
  originSessionId: a902e50d-bd1f-422b-8298-552e3fb0a73f
---

For this TTT project, **inference is always done with cf-patch** (closed-form ridge ΔW written into each TTT layer's W_down = the fast-weight update). That IS our inference mode.

**Why:** The user standardized on cf-patch as the one inference method (live-TTT paper-rule cumsum is unstable on these ckpts; cf-patch's one-shot ΔW is the fast-weight update we use).

**How to apply:**
- When reporting cf-patch RULER/eval results, report ONLY the `patch` (cf-patch) score. Do NOT present the `vanilla` column as a baseline or result.
- The `vanilla` column in `eval_cf_ruler.sh` output is a Phase-1 artifact: OpenCompass workers do NOT import `inference_model`, so they run stock Qwen3 with no TTT and no cf-patch — it is not a mode we use. Ignore it / don't surface it.
- Cross-checkpoint or cross-config comparisons should compare cf-patch vs cf-patch.

See [[ttt-inference-findings]] for why live-TTT is off and cf-patch is the chosen path.