research: record frozen multilingual confirmation

Author: Labeeb2339

README.md

@@ -3,20 +3,24 @@
 RecurQuant is a reproducible research harness for **persistent recurrent-state
 quantization in Gated DeltaNet language models**.
 
-> **Current status:** research infrastructure and diagnostic pilot. There is no
-> validated quantization method, memory reduction, or speedup result yet.
+> **Current status:** the frozen diagnostic candidate passed its untouched
+> confirmation trace. This is not yet a validated general method, realized
+> memory reduction, speedup, or novelty result.
 
 The first calibration/development pilot found substantial layer heterogeneity.
 At a 4.22-bit average payload, retaining only Gated DeltaNet layer 0 at INT8 and
 using INT4 for the other 17 layers reduced worst-5% token KL by 79.8% on a
 retrieval-style trace and 62.2% on a code-style trace relative to uniform INT4.
-These are short synthetic traces, not a benchmark or generalization result.
+On the untouched multilingual trace, the same frozen plan reduced worst-5%
+token KL by 77.8% and increased top-1 agreement from 25.0% to 59.4%. These are
+short synthetic traces, not a benchmark or generalization result.
 
 ## Research question
 
-Can sub-8-bit storage of Gated DeltaNet's fixed recurrent matrix state use its
-decay and write dynamics to preserve difficult long-context behavior better than
-uniform quantization at the same modeled bit budget?
+Can sub-8-bit storage of Gated DeltaNet's fixed recurrent matrix state allocate
+precision from query-weighted read sensitivity to preserve difficult
+long-context behavior better than uniform quantization at the same modeled bit
+budget?
 
 [Qwen3.5-0.8B-Base](https://huggingface.co/Qwen/Qwen3.5-0.8B-Base) is the first
 target. Its language model repeats three Gated DeltaNet layers followed by one
@@ -52,7 +56,9 @@ The narrower hypothesis under investigation is **precision allocation for the
 persistent Gated DeltaNet matrix state**, conditioned on Gated DeltaNet dynamics
 and compared at an equal bit budget. See
 [the claim boundary](research/CLAIM_BOUNDARY.md) and
-[pilot protocol](research/PILOT_PROTOCOL.md).
+[pilot protocol](research/PILOT_PROTOCOL.md). The documented experiment trail
+preserves the [failed signals and replacement](research/EXPERIMENT_001_SIGNAL_PIVOT.md)
+and the [untouched confirmation](research/CONFIRMATION_001.md).
 
 The user-suggested
 [Gated DeltaNet-2 paper](https://arxiv.org/abs/2605.22791) reinforces why erase,
@@ -75,6 +81,24 @@ uv pip install --python .venv\Scripts\python.exe -e ".[dev]"
 The model experiment is intentionally separate from the unit-test suite because
 it downloads approximately 1.75 GB of public model weights.
 
+## Reproduce the frozen confirmation
+
+The script pins the model revision and records the environment, token digest,
+state layout, metrics, and canonical evidence hash:
+
+```powershell
+.venv\Scripts\python.exe scripts\run_qwen35_smoke.py `
+  --upgrade-layers 0 --low-bits 4 --high-bits 8 `
+  --group-size 128 --rounding nearest `
+  --prefill-tokens 32 --decode-tokens 32 `
+  --prompt-profile multilingual `
+  --output artifacts\multilingual-confirmation.json
+```
+
+This reruns the already disclosed confirmation profile; it is a reproducibility
+check, not a new held-out test. The recorded result and its limitations are in
+[Confirmation 001](research/CONFIRMATION_001.md).
+
 ## Research discipline
 
 - Model and tokenizer revisions are pinned in evidence artifacts.