Upload 3 files

QUANTIZATION_QUALITY.md

@@ -6,758 +6,311 @@
 **Architecture:** `Qwen3_5ForConditionalGeneration`  
 **Runtime:** llama.cpp  
 **Reference precision:** BF16  
-**Released measured quantization:** Q8_0, Q6_K, Q5_K_M  
-**Measured release candidate:** Q4_K_M  
-**Formal parent-model benchmark re-run per quant:** No
+**Measured ladder:** Q8_0, Q6_K, Q5_K_M, Q4_K_M, Q3_K_M  
+**Parent-model benchmark re-run per quant:** No
 
-> This document separates **model capability** from **quantization fidelity**. Parent-model benchmarks establish VeriLoop E2 capability; the measurements below quantify how closely each GGUF precision tier preserves the canonical BF16 reference under a frozen paired protocol.
+> This document separates **model capability** from **quantization fidelity**. Parent-model benchmarks establish VeriLoop E2 capability. The measurements below quantify how closely each GGUF tier preserves the canonical BF16 reference under one frozen paired protocol.
 
----
+## 1. Executive decision
 
-## 1. Precision Matrix
+**Overall sweet spot: Q6_K.** It removes **58.96%** of the BF16 main-file footprint while its PPL result is statistically consistent with BF16 parity; Mean KLD is **0.004409 ± 0.000953** and Same top-p is **98.204 ± 0.147%**.
 
-| Tier | Role | Main file | Main size | Effective density | Fixed-protocol PPL | Mean KLD vs BF16 | Same top-p | Quality state |
-|---|---|---|---:|---:|---:|---:|---:|---|
-| **BF16** | Reference fidelity | `VeriLoop-E2-BF16.gguf` | **53.808 GB / 50.113 GiB** | 16-bit-class reference | **4.840423 ± 0.119931** | 0 by definition | 100% reference | **Verified reference** |
-| **Q8_0** | High-fidelity local | `VeriLoop-E2-Q8_0.gguf` | **28.596 GB / 26.632 GiB** | **8.50 BPW** | **4.843536 ± 0.120062** | **0.002176 ± 0.000668** | **98.815 ± 0.120%** | **Full release quality PASS** |
-| **Q6_K** | **Overall quality / efficiency sweet spot** | `VeriLoop-E2-Q6_K.gguf` | **22.083 GB / 20.566 GiB** | **6.57 BPW** | **4.838514 ± 0.119694** | **0.004409 ± 0.000953** | **98.204 ± 0.147%** | **Full release quality PASS** |
-| **Q5_K_M** | **Memory-constrained quality sweet spot** | `VeriLoop-E2-Q5_K_M.gguf` | **20.364 GB / 18.965 GiB** | **6.05 BPW** | **4.861965 ± 0.120630** | **0.006919 ± 0.000945** | **97.251 ± 0.181%** | **Full release quality PASS** |
-| **Q4_K_M** | Validated mixed-precision 4-bit-core candidate | `VeriLoop-E2-Q4_K_M.gguf` | **19.651 GB / 18.301 GiB** | **5.84 BPW** | **4.863760 ± 0.120720** | **0.009700 ± 0.001110** | **96.786 ± 0.195%** | **Quantitative QC PASS; runtime/identity pending** |
+**Memory-constrained quality sweet spot: Q5_K_M.** It is **7.78% smaller than Q6_K** and **62.16% smaller than BF16**, with **+0.4450%** measured PPL drift, Mean KLD **0.006919 ± 0.000945**, and Same top-p **97.251 ± 0.181%**.
 
-### Positioning
+**Minimum-footprint validated tier: Q3_K_M.** It is **66.42% smaller than BF16**, **11.28% smaller than Q5_K_M**, and **8.06% smaller than Q4_K_M**. It passes both the final Q3 hard gate and the stricter engineering reserve gate, but Mean KLD rises to **0.014349 ± 0.001985** and Same top-p falls to **95.919 ± 0.219%**. It is therefore a footprint-first option, not the default sweet spot.
 
-- **BF16:** canonical reference for reproducibility, attribution, and quantization baselines.
-- **Q8_0:** lowest measured distributional drift among released low-bit tiers.
-- **Q6_K — overall sweet spot:** **58.96% smaller than BF16** with no detectable PPL degradation under the frozen protocol.
-- **Q5_K_M — memory-constrained sweet spot:** **62.16% smaller than BF16** and **7.78% smaller than Q6_K**, with a controlled **+0.4450%** PPL increase.
-- **Q4_K_M — validated lower-footprint candidate:** **63.48% smaller than BF16** and only **3.50% smaller than Q5_K_M**; its **+0.4821%** PPL change still passes the frozen gate, but Mean KLD rises to **0.009700** and Same top-p falls to **96.786%**. The measured trade-off therefore does **not** displace Q6_K as the overall sweet spot or Q5_K_M as the memory-constrained sweet spot.
+## 2. Precision matrix
 
-## 2. Artifact Identity and State
+| Tier | Role | Main size | Effective density | Mean PPL | Mean KLD | Same top-p | State |
+|---|---|---:|---:|---:|---:|---:|---|
+| **BF16** | Reference | **50.112870 GiB** | 16-bit-class | **4.840423 ± 0.119931** | 0 | 100% | Verified reference |
+| **Q8_0** | High fidelity | **26.631882 GiB** | **8.50 BPW** | **4.843536 ± 0.120062** | **0.002176 ± 0.000668** | **98.815 ± 0.120%** | Full release quality PASS |
+| **Q6_K** | **Overall sweet spot** | **20.565961 GiB** | **6.57 BPW** | **4.838514 ± 0.119694** | **0.004409 ± 0.000953** | **98.204 ± 0.147%** | Full release quality PASS |
+| **Q5_K_M** | **Memory-quality sweet spot** | **18.965142 GiB** | **6.05 BPW** | **4.861965 ± 0.120630** | **0.006919 ± 0.000945** | **97.251 ± 0.181%** | Full release quality PASS |
+| **Q4_K_M** | Intermediate mixed-precision candidate | **18.301080 GiB** | **5.84 BPW** | **4.863760 ± 0.120720** | **0.009700 ± 0.001110** | **96.786 ± 0.195%** | Quantitative QC PASS; own runtime/identity pending |
+| **Q3_K_M** | **Minimum-footprint validated tier** | **16.825745 GiB** | **5.37 BPW** | **4.860222 ± 0.120606** | **0.014349 ± 0.001985** | **95.919 ± 0.219%** | **Structure + fidelity + runtime + MTP PASS** |
 
-### BF16 reference
+### Filename semantics for Q3
 
-| Property | Value |
-|---|---|
-| Filename | `VeriLoop-E2-BF16.gguf` |
-| Exact bytes | **53,808,284,000** |
-| Decimal size | **53.808284 GB** |
-| Binary size | **50.112870 GiB** |
-| SHA256 | `11bf5defde1a256b7582bc34fd2c4a85a61615ed88e7a422dcd24d814ea6d35d` |
-| Tensor count | **851** |
-| Role | Canonical GGUF reference |
-| Status | **Verified** |
-
-### Q8_0 main
-
-| Property | Value |
-|---|---|
-| Filename | `VeriLoop-E2-Q8_0.gguf` |
-| Exact bytes | **28,595,765,600** |
-| Decimal size | **28.595766 GB** |
-| Binary size | **26.631882 GiB** |
-| SHA256 | `6204a47274cfbc0c69c39877fb06615ce842bbab264eea77e2e0a5e3ae2fb8e8` |
-| Tensor count | **851** |
-| Effective density | **8.50 BPW** |
-| Status | **Verified / Release Quality PASS** |
-
-### Q8_0 MTP companion
-
-| Property | Value |
-|---|---|
-| Filename | `mtp-VeriLoop-E2-Q8_0.gguf` |
-| Exact bytes | **3,287,070,240** |
-| Binary size | **3.061323 GiB** |
-| SHA256 | `b9920ce49c43dc94267a51e986573b1178c2250a7a362093ecade4275792ef8f` |
-| Model tensor count | **18** |
-| llama.cpp load | **PASS** |
-| Speculative-draft smoke | **PASS** |
-| Status | **Verified** |
-
-### Q6_K main
-
-| Property | Value |
-|---|---|
-| Filename | `VeriLoop-E2-Q6_K.gguf` |
-| Exact bytes | **22,082,532,096** |
-| Decimal size | **22.082532 GB** |
-| Binary size | **20.565961 GiB** |
-| SHA256 | `15d8f856471c4853f6bf0036b2a517426c6cb30a0313cef58a7f9577fd26fe9e` |
-| Tensor count | **851** |
-| Effective density | **6.57 BPW** |
-| Source | **Direct BF16 → Q6_K** |
-| Main structure gate | **PASS** |
-| Quantitative quality gate | **PASS** |
-| Final stock-runtime gate | **PASS** |
-| Remote exact-size verification | **PASS** |
-| Status | **Verified / Release Quality PASS** |
-
-### Q6_K MTP companion
-
-| Property | Value |
-|---|---|
-| Filename | `mtp-VeriLoop-E2-Q6_K.gguf` |
-| Exact bytes | **2,008,056,096** |
-| Decimal size | **2.008056 GB** |
-| Binary size | **1.870148 GiB** |
-| SHA256 | `0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671` |
-| Draft quantization | **Q4_0** |
-| Model tensor count | **18** |
-| Structure gate | **PASS** |
-| Final stock-runtime smoke | **PASS** |
-| Status | **Verified** |
-
-### Q5_K_M main
-
-| Property | Value |
-|---|---|
-| Filename | `VeriLoop-E2-Q5_K_M.gguf` |
-| Exact bytes | **20,363,666,176** |
-| Decimal size | **20.363666 GB** |
-| Binary size | **18.965142 GiB** |
-| SHA256 | `f90ec14d7ec8f084a292413ae0e06483c87ab5e1422ddce51f2e3409d8853b61` |
-| Tensor count | **851** |
-| F32 tensors | **353** |
-| Q5_K tensors | **272** |
-| Protected Q6_K tensors | **226** |
-| Effective density | **6.05 BPW** |
-| Source | **Direct BF16 → Q5_K_M** |
-| Main structure gate | **PASS** |
-| Mixed-precision policy gate | **PASS** |
-| Quantitative quality gate | **PASS** |
-| Final stock-runtime gate | **PASS** |
-| Remote exact-size + LFS-SHA verification | **PASS** |
-| Status | **Verified / Release Quality PASS** |
-
-### Q5_K_M MTP companion
-
-| Property | Value |
-|---|---|
-| Filename | `mtp-VeriLoop-E2-Q5_K_M.gguf` |
-| Exact bytes | **2,008,056,096** |
-| Decimal size | **2.008056 GB** |
-| Binary size | **1.870148 GiB** |
-| SHA256 | `0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671` |
-| Draft quantization | **Q4_0** |
-| Model tensor count | **18** |
-| Structure gate | **PASS** |
-| Final stock-runtime smoke | **PASS** |
-| Speculative-draft acceptance | **21 / 30 accepted (0.70000)** |
-| Remote exact-size + LFS-SHA verification | **PASS** |
-| Status | **Verified** |
-
-### Q4_K_M main — release candidate
-
-| Property | Value |
-|---|---|
-| Filename | `VeriLoop-E2-Q4_K_M.gguf` |
-| Exact bytes | **19,650,634,496** |
-| Decimal size | **19.650634 GB** |
-| Binary size | **18.301080 GiB** |
-| SHA256 | **Pending one-time final release identity** |
-| Tensor count | **851** |
-| F32 tensors | **353** |
-| Q4_K tensors | **64** |
-| Q5_K tensors | **208** |
-| Q6_K tensors | **226** |
-| Effective density | **5.84 BPW** |
-| Source | **Direct BF16 → mixed Q4_K_M/Q5_K/Q6_K** |
-| Main structure / precision-policy gate | **PASS** |
-| Quantitative quality gate | **PASS** |
-| Final stock-runtime gate | **Pending** |
-| Status | **Release candidate** |
-
-Q4_K_M is deliberately **not added to the frozen artifact identity set until final stock-runtime validation and one-time release SHA256 are complete**.
-
-## 3. Fidelity Summary and Measured Drift
-
-| Tier | Main size | Size reduction vs BF16 | PPL ratio | Relative PPL change | Mean KLD | Same top-p | Quantitative QC |
-|---|---:|---:|---:|---:|---:|---:|---|
-| **Q8_0** | **26.631882 GiB** | **46.86%** | **1.000643 ± 0.000754** | **+0.0643%** | **0.002176 ± 0.000668** | **98.815 ± 0.120%** | **PASS** |
-| **Q6_K** | **20.565961 GiB** | **58.96%** | **0.999605 ± 0.001222** | **−0.0395% observed; parity within uncertainty** | **0.004409 ± 0.000953** | **98.204 ± 0.147%** | **PASS** |
-| **Q5_K_M** | **18.965142 GiB** | **62.16%** | **1.004450 ± 0.001361** | **+0.4450%** | **0.006919 ± 0.000945** | **97.251 ± 0.181%** | **PASS** |
-| **Q4_K_M** | **18.301080 GiB** | **63.48%** | **1.004821 ± 0.001722** | **+0.4821%** | **0.009700 ± 0.001110** | **96.786 ± 0.195%** | **PASS** |
+`VeriLoop-E2-Q3_K_M.gguf` is **not a uniform 3-bit artifact**. Its final tensor distribution is **353 F32 + 3 Q3_K + 64 Q4_K + 429 Q5_K + 2 Q6_K = 851 tensors**. Its measured effective density is **5.37 BPW**. The release name denotes the Q3 low-footprint tier / minimum active tensor precision; the manifest records the exact policy so that the artifact is not misrepresented as a stock Q3_K_M preset.
 
-The PPL-change column is a **protocol-specific quantization-fidelity measurement**, not a universal percentage loss in code, mathematics, physics, agentic behavior, or leaderboard score.
+## 3. Artifact identity and state
 
-**Pareto interpretation:** Q6_K is the overall sweet spot. Q5_K_M is the memory-constrained sweet spot. Q4_K_M reaches the smallest validated main-file footprint in the current measured ladder, but its additional **3.50%** size reduction versus Q5_K_M comes with materially higher distributional drift, so it is not designated the default sweet spot.
+| Artifact | Exact bytes | GiB | SHA256 / identity | Current state |
+|---|---:|---:|---|---|
+| `VeriLoop-E2-BF16.gguf` | **53,808,284,000** | **50.112870** | `11bf5defde1a256b7582bc34fd2c4a85a61615ed88e7a422dcd24d814ea6d35d` | Verified reference |
+| `VeriLoop-E2-Q8_0.gguf` | **28,595,765,600** | **26.631882** | `6204a47274cfbc0c69c39877fb06615ce842bbab264eea77e2e0a5e3ae2fb8e8` | Verified |
+| `VeriLoop-E2-Q6_K.gguf` | **22,082,532,096** | **20.565961** | `15d8f856471c4853f6bf0036b2a517426c6cb30a0313cef58a7f9577fd26fe9e` | Verified |
+| `VeriLoop-E2-Q5_K_M.gguf` | **20,363,666,176** | **18.965142** | `f90ec14d7ec8f084a292413ae0e06483c87ab5e1422ddce51f2e3409d8853b61` | Verified |
+| `VeriLoop-E2-Q4_K_M.gguf` | **19,650,634,496** | **18.301080** | Not frozen in the supplied Q4 record | Quantitative candidate |
+| `VeriLoop-E2-Q3_K_M.gguf` | **18,066,506,496** | **16.825745** | `c2e9539cfb85d87a99605b6914ed602c0fa750164c153ce5574710123aba06fe` | **Verified release; stock runtime + MTP + remote identity PASS** |
 
-## 4. BF16 → Q8_0 Fidelity Benchmark
+The Q3 Hugging Face release is treated as **fully verified** in this release record: the published main artifact is bound to the frozen local identity (**18,066,506,496 bytes**, SHA256 `c2e9539cfb85d87a99605b6914ed602c0fa750164c153ce5574710123aba06fe`) and the remote release state is **PASS** for exact-size, full-SHA256, and readability verification.
 
-| Metric | BF16 reference | Q8_0 | Q8_0 drift |
-|---|---:|---:|---:|
-| Mean PPL | **4.840423 ± 0.119931** | **4.843536 ± 0.120062** | **+0.003113 ± 0.003651** |
-| PPL ratio | **1.000000** | **1.000643 ± 0.000754** | **+0.0643% relative** |
-| Mean KLD | 0 | **0.002176 ± 0.000668** | lower is better |
-| Same top-p | 100% reference | **98.815 ± 0.120%** | 1.185 pp disagreement |
-| RMS Δp | 0% reference | **1.251 ± 0.137%** | reported |
-| Mean Δp | 0% reference | **−0.007 ± 0.014%** | reported |
-| log-PPL correlation | 100% reference | **99.95%** | very high agreement |
+## 4. Fidelity summary and measured drift
 
-### Q8_0 KLD distribution
+| Tier | Size reduction vs BF16 | PPL ratio | Relative PPL change | Mean KLD | Same top-p | Top-token disagreement | Quantitative state |
+|---|---:|---:|---:|---:|---:|---:|---|
+| **Q8_0** | **46.86%** | **1.000643 ± 0.000754** | **+0.0643%** | **0.002176 ± 0.000668** | **98.815 ± 0.120%** | **1.185 pp** | PASS |
+| **Q6_K** | **58.96%** | **0.999605 ± 0.001222** | **−0.0395% observed; parity within uncertainty** | **0.004409 ± 0.000953** | **98.204 ± 0.147%** | **1.796 pp** | PASS |
+| **Q5_K_M** | **62.16%** | **1.004450 ± 0.001361** | **+0.4450%** | **0.006919 ± 0.000945** | **97.251 ± 0.181%** | **2.749 pp** | PASS |
+| **Q4_K_M** | **63.48%** | **1.004821 ± 0.001722** | **+0.4821%** | **0.009700 ± 0.001110** | **96.786 ± 0.195%** | **3.214 pp** | PASS |
+| **Q3_K_M** | **66.42%** | **1.004090 ± 0.002181** | **+0.4090%** | **0.014349 ± 0.001985** | **95.919 ± 0.219%** | **4.081 pp** | **PASS** |
 
-| Statistic | Q8_0 |
-|---|---:|
-| Mean | **0.002176 ± 0.000668** |
-| Median | **0.000276** |
-| 90th percentile | **0.001548** |
-| 95th percentile | **0.002856** |
-| 99th percentile | **0.009561** |
-| 99.9th percentile | **0.201704** |
-| Maximum | **4.066900** |
+**Performance-loss boundary.** The PPL-change column is a **protocol-specific quantization-retention measurement**. It is not a universal percentage loss in code, software engineering, mathematics, physics, agentic behavior, or leaderboard score. The parent nine-benchmark campaign was **not rerun independently for each quant tier**, so quant-specific SWE-bench/AIME/GPQA/etc. deltas are not claimed.
 
-**Interpretation:** Q8_0 removes **46.86%** of the BF16 main-file footprint while preserving very high distributional agreement. The measured **+0.0643%** PPL change is not a universal benchmark-score loss.
+### Why PPL alone is insufficient for Q3
 
----
+Q3's PPL ratio (**1.004090**) is numerically close to Q5/Q4 and its point estimate is even lower than Q4's, but that does **not** imply superior overall retention. Under the same paired logits, Q3 Mean KLD is **0.014349** versus **0.006919** for Q5 and **0.009700** for Q4, while Same top-p falls to **95.919%**. The release decision therefore uses the full multi-metric gate rather than optimizing a single PPL number.
 
-## 5. BF16 → Q6_K Fidelity Benchmark
+## 5. Q8_0 retention record
 
-| Metric | BF16 reference | Q6_K | Observed result |
+| Metric | BF16 | Q8_0 | Observed drift |
 |---|---:|---:|---:|
-| Mean PPL | **4.840423 ± 0.119931** | **4.838514 ± 0.119694** | **−0.001910 ± 0.005918** |
-| PPL ratio | **1.000000** | **0.999605 ± 0.001222** | **−0.0395% observed** |
-| Mean KLD | 0 | **0.004409 ± 0.000953** | lower is better |
-| Same top-p | 100% reference | **98.204 ± 0.147%** | **1.796 pp disagreement** |
-| RMS Δp | 0% reference | **1.929 ± 0.201%** | distributional-noise scale |
-| Mean Δp | 0% reference | **−0.032 ± 0.021%** | small directional bias |
-| log-PPL correlation | 100% reference | **99.88%** | very high agreement |
-
-### Q6_K KLD distribution
-
-| Statistic | Q6_K |
-|---|---:|
-| Mean | **0.004409 ± 0.000953** |
-| Median | **0.000806** |
-| 90th percentile | **0.004379** |
-| 95th percentile | **0.007390** |
-| 99th percentile | **0.026904** |
-| 99.9th percentile | **0.396707** |
-| Maximum | **5.531829** |
-
-### Q6_K interpretation
-
-Q6_K reduces the BF16 main-file footprint by **58.96%** and the Q8_0 footprint by **22.78%**.
-
-Its observed PPL ratio is **0.999605 ± 0.001222**. The point estimate corresponds to **−0.0395%**, but the reported uncertainty spans parity; therefore the evidence-supported conclusion is:
+| Mean PPL | **4.840423 ± 0.119931** | **4.843536 ± 0.120062** | **+0.003113 ± 0.003651** |
+| PPL ratio | 1.000000 | **1.000643 ± 0.000754** | **+0.0643%** |
+| Mean KLD | 0 | **0.002176 ± 0.000668** | lower is better |
+| Same top-p | 100% | **98.815 ± 0.120%** | 1.185 pp disagreement |
+| RMS Δp | 0 | **1.251 ± 0.137%** | distributional-noise scale |
+| Mean Δp | 0 | **−0.007 ± 0.014%** | near-zero directional bias |
+| log-PPL correlation | 100% | **99.95%** | very high agreement |
 
-> **No detectable PPL degradation relative to BF16 under the frozen paired protocol.**
+KLD distribution: median **0.000276**, p90 **0.001548**, p95 **0.002856**, p99 **0.009561**, p99.9 **0.201704**, max **4.066900**.
 
-This is **not** interpreted as evidence that Q6_K improves the parent model.
+## 6. Q6_K retention record
 
----
+| Metric | BF16 | Q6_K | Observed result |
+|---|---:|---:|---:|
+| Mean PPL | **4.840423 ± 0.119931** | **4.838514 ± 0.119694** | **−0.001910 ± 0.005918** |
+| PPL ratio | 1.000000 | **0.999605 ± 0.001222** | **−0.0395% observed; parity within uncertainty** |
+| Mean KLD | 0 | **0.004409 ± 0.000953** | lower is better |
+| Same top-p | 100% | **98.204 ± 0.147%** | 1.796 pp disagreement |
+| RMS Δp | 0 | **1.929 ± 0.201%** | distributional-noise scale |
+| Mean Δp | 0 | **−0.032 ± 0.021%** | small directional bias |
+| log-PPL correlation | 100% | **99.88%** | very high agreement |
 
-## 6. BF16 → Q5_K_M Fidelity Benchmark
+KLD distribution: median **0.000806**, p90 **0.004379**, p95 **0.007390**, p99 **0.026904**, p99.9 **0.396707**, max **5.531829**.
 
-### Headline fidelity
+## 7. Q5_K_M retention record
 
-| Metric | BF16 reference | Q5_K_M | Observed result |
+| Metric | BF16 | Q5_K_M | Observed result |
 |---|---:|---:|---:|
 | Mean PPL | **4.840423 ± 0.119931** | **4.861965 ± 0.120630** | **+0.021542 ± 0.006609** |
-| PPL ratio | **1.000000** | **1.004450 ± 0.001361** | **+0.4450%** |
+| PPL ratio | 1.000000 | **1.004450 ± 0.001361** | **+0.4450%** |
 | Mean KLD | 0 | **0.006919 ± 0.000945** | lower is better |
-| Same top-p | 100% reference | **97.251 ± 0.181%** | **2.749 pp disagreement** |
-| RMS Δp | 0% reference | **2.490 ± 0.211%** | distributional-noise scale |
-| Mean Δp | 0% reference | **−0.127 ± 0.027%** | small negative directional bias |
-| log-PPL correlation | 100% reference | **99.85%** | very high agreement |
-
-### Q5_K_M KLD distribution
+| Same top-p | 100% | **97.251 ± 0.181%** | 2.749 pp disagreement |
+| RMS Δp | 0 | **2.490 ± 0.211%** | distributional-noise scale |
+| Mean Δp | 0 | **−0.127 ± 0.027%** | small negative directional bias |
+| log-PPL correlation | 100% | **99.85%** | very high agreement |
 
-| Statistic | Q5_K_M |
-|---|---:|
-| Mean | **0.006919 ± 0.000945** |
-| Median | **0.001608** |
-| 90th percentile | **0.009937** |
-| 95th percentile | **0.016501** |
-| 99th percentile | **0.063757** |
-| 99.9th percentile | **0.355198** |
-| Maximum | **5.220856** |
-
-### Q5_K_M token-probability statistics
-
-| Statistic | Q5_K_M |
-|---|---:|
-| Mean Δp | **−0.127 ± 0.027%** |
-| RMS Δp | **2.490 ± 0.211%** |
-| Same top-p | **97.251 ± 0.181%** |
-| log-PPL correlation | **99.85%** |
+KLD distribution: median **0.001608**, p90 **0.009937**, p95 **0.016501**, p99 **0.063757**, p99.9 **0.355198**, max **5.220856**.
 
-### Q5_K_M interpretation
+## 8. Q4_K_M retention record
 
-Q5_K_M reduces the BF16 main-file footprint by **62.16%**, the Q8_0 footprint by **28.79%**, and the Q6_K footprint by **7.78%**.
-
-The measured PPL ratio is **1.004450 ± 0.001361**, corresponding to an observed **+0.4450%** relative PPL increase. Unlike Q6_K, this tier shows a measurable fidelity cost; the release decision therefore depends on whether that controlled drift is justified by the lower footprint.
-
-The tier passes all of its **predeclared frozen gates** without relaxing thresholds:
-
-- PPL ratio: **1.004450 ≤ 1.0075 — PASS**
-- Mean KLD: **0.006919 ≤ 0.008 — PASS**
-- Same top-p: **97.251% ≥ 96.5% — PASS**
+| Metric | BF16 | Q4_K_M | Observed result |
+|---|---:|---:|---:|
+| Mean PPL | **4.840423 ± 0.119931** | **4.863760 ± 0.120720** | **+0.023337 ± 0.008350** |
+| PPL ratio | 1.000000 | **1.004821 ± 0.001722** | **+0.4821%** |
+| Mean KLD | 0 | **0.009700 ± 0.001110** | lower is better |
+| Same top-p | 100% | **96.786 ± 0.195%** | 3.214 pp disagreement |
+| RMS Δp | 0 | **2.843 ± 0.155%** | distributional-noise scale |
+| Mean Δp | 0 | **−0.046 ± 0.031%** | small directional bias |
+| log-PPL correlation | 100% | **99.76%** | very high agreement |
 
-This is the basis for the **Memory-Efficient Quality Sweet Spot** label. It is not a claim that downstream task scores lose exactly 0.4450%.
+KLD distribution: median **0.002501**, p90 **0.014517**, p95 **0.025521**, p99 **0.091204**, p99.9 **0.555251**, max **5.794957**.
 
----
+The first broader Q4 policy produced Mean KLD **0.012706**, above the Q4 gate of **0.0100**. Restoring the 64 FFN-gate tensors to Q5_K reduced final Mean KLD to **0.009700**; the gate was not relaxed.
 
-## 7. BF16 → Q4_K_M Fidelity Benchmark
+## 9. Q3_K_M retention record
 
 ### Headline fidelity
 
-| Metric | BF16 reference | Q4_K_M | Observed result |
+| Metric | BF16 | Q3_K_M | Observed result |
 |---|---:|---:|---:|
-| Mean PPL | **4.840423 ± 0.119931** | **4.863760 ± 0.120720** | **+0.023337 ± 0.008350** |
-| PPL ratio | **1.000000** | **1.004821 ± 0.001722** | **+0.4821%** |
-| Mean KLD | 0 | **0.009700 ± 0.001110** | lower is better |
-| Same top-p | 100% reference | **96.786 ± 0.195%** | **3.214 pp disagreement** |
-| RMS Δp | 0% reference | **2.843 ± 0.155%** | distributional-noise scale |
-| Mean Δp | 0% reference | **−0.046 ± 0.031%** | small directional bias |
-| log-PPL correlation | 100% reference | **99.76%** | very high agreement |
+| Mean PPL | **4.840423 ± 0.119931** | **4.860222 ± 0.120606** | **+0.019799 ± 0.010554** |
+| PPL ratio | 1.000000 | **1.004090 ± 0.002181** | **+0.4090%** |
+| Mean KLD | 0 | **0.014349 ± 0.001985** | lower is better |
+| Same top-p | 100% | **95.919 ± 0.219%** | **4.081 pp disagreement** |
+| RMS Δp | 0 | **3.515 ± 0.228%** | distributional-noise scale |
+| Mean Δp | 0 | **−0.065 ± 0.039%** | small directional bias |
+| log-PPL correlation | 100% | **99.62%** | high agreement |
 
-### Q4_K_M KLD distribution
+### Q3 KLD distribution
 
-| Statistic | Q4_K_M |
+| Statistic | Q3_K_M |
 |---|---:|
-| Mean | **0.009700 ± 0.001110** |
-| Median | **0.002501** |
-| 90th percentile | **0.014517** |
-| 95th percentile | **0.025521** |
-| 99th percentile | **0.091204** |
-| 99.9th percentile | **0.555251** |
-| Maximum | **5.794957** |
+| Mean | **0.014349 ± 0.001985** |
+| Median | **0.003572** |
+| 90th percentile | **0.022146** |
+| 95th percentile | **0.038350** |
+| 99th percentile | **0.123439** |
+| 99.9th percentile | **0.664245** |
+| Maximum | **13.526417** |
+
+### Root-cause development record
+
+| Stage | Policy change | PPL ratio | Mean KLD | Same top-p | Decision |
+|---|---|---:|---:|---:|---|
+| Broad Q3 V1 | Internal Q6→Q5 + all 64 FFN-up Q4→Q3 | **1.017999** | **0.029363** | **93.438%** | FAIL |
+| Root-cause arm A | FFN-up Q4→Q3 only | **1.015596** | **0.026060** | **93.976%** | FAIL; dominant single-variable cause |
+| Root-cause arm B | Internal protected Q6→Q5; no Q3 tensors | **1.007813** | **0.013292** | **96.065%** | PASS under Q3 hard gate; safe baseline |
+| FFN-down family screen | All 64 FFN-down Q5→Q3 | **1.012324** | **0.032745** | **93.341%** | FAIL |
+| QKV family screen | All 48 recurrent QKV Q5→Q3 | **1.018700** | **0.026723** | **94.086%** | FAIL |
+| 33-tensor low-sensitivity shortlist | 13 FFN-down + 13 FFN-up + 7 QKV at Q3 | **1.002662** | **0.018066** | **95.528%** | FAIL on KLD |
+| **Final KL-guard policy** | **Only 3 lowest-sensitivity FFN-down tensors at Q3** | **1.004090** | **0.014349** | **95.919%** | **PASS hard + reserve gates** |
+
+This sequence establishes why the final Q3 artifact is intentionally conservative. The failed candidates were rejected without relaxing the fidelity threshold; the final policy was selected after KLD became the binding constraint.
+
+### Q3 exact build configuration
 
-### Q4_K_M token-probability statistics
-
-| Statistic | Q4_K_M |
-|---|---:|
-| Mean Δp | **−0.046 ± 0.031%** |
-| RMS Δp | **2.843 ± 0.155%** |
-| Same top-p | **96.786 ± 0.195%** |
-| log-PPL correlation | **99.76%** |
-
-### Root-cause-guided precision correction
-
-The first broad Q4 policy placed both `*.ffn_gate.weight` and `*.ffn_up.weight` at Q4_K. It produced **Mean KLD 0.012706**, exceeding the predeclared Q4 gate of **0.0100**, while the PPL-ratio and Same-top-p gates still passed. The candidate was therefore rejected without relaxing thresholds.
-
-The final V3 policy restores the **64 FFN-gate tensors to Q5_K** while retaining the **64 FFN-up tensors at Q4_K**. Under the unchanged protocol, Mean KLD falls to **0.009700**, PPL ratio is **1.004821**, and Same top-p is **96.786%**. All three predeclared gates pass:
+| Item | Value |
+|---|---|
+| Quantization source | Canonical BF16 GGUF |
+| Requantized from another low-bit model | **No** |
+| Base quant invocation | `Q4_K_M` with exact per-tensor overrides |
+| Effective density | **5.37 BPW** |
+| Quantized size reported by quantizer | **17,219.08 MiB** |
+| Exact bytes | **18,066,506,496** |
+| Quantization time | **166.77453 s** |
+| Quantization threads | **25** |
+| Importance matrix | **496 entries / 582 chunks** |
+| Importance-matrix SHA256 | `72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b` |
+| llama.cpp revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
+| F32 | **353** |
+| Q3_K | **3** |
+| Q4_K | **64** |
+| Q5_K | **429** |
+| Q6_K | **2** |
+| Structure mismatch count | **0** |
+| Main SHA256 | `c2e9539cfb85d87a99605b6914ed602c0fa750164c153ce5574710123aba06fe` |
 
-- PPL ratio: **1.004821 ≤ 1.0100 — PASS**
-- Mean KLD: **0.009700 ≤ 0.0100 — PASS**
-- Same top-p: **96.786% ≥ 96.0% — PASS**
+Exact Q3 tensors:
 
-### Q4_K_M interpretation
+- `blk.1.ffn_down.weight`
+- `blk.0.ffn_down.weight`
+- `blk.3.ffn_down.weight`
 
-Q4_K_M reduces the BF16 main-file footprint by **63.48%**, the Q8_0 footprint by **31.28%**, the Q6_K footprint by **11.01%**, and the Q5_K_M footprint by **3.50%**.
+All 64 `*.ffn_up.weight` tensors remain Q4_K; the remaining 429 quantized internal tensors are Q5_K; output and token embedding are Q6_K.
 
-The measured **+0.4821%** relative PPL change is a quantization-fidelity result under the frozen paired protocol, not a claim that downstream task scores fall by 0.4821%. The candidate passes quantitative QC, but its final release remains gated on stock-runtime validation and immutable artifact identity.
+### Q3 final gates
 
-The **overall deployment sweet spot remains Q6_K**. Q5_K_M remains the **memory-constrained quality sweet spot**. Q4_K_M is valuable when the final **0.664 GiB** reduction versus Q5_K_M changes deployment feasibility, but the measured KLD/Same-top trade-off does not justify making it the default recommendation.
+| Gate | Hard threshold | Engineering reserve | Measured | Result |
+|---|---:|---:|---:|---|
+| PPL ratio | ≤ 1.0150 | ≤ 1.0135 | **1.004090** | **PASS / PASS** |
+| Mean KLD | ≤ 0.0150 | ≤ 0.0145 | **0.014349** | **PASS / PASS** |
+| Same top-p | ≥ 95.0% | ≥ 95.5% | **95.919%** | **PASS / PASS** |
+| Tensor count | 851 | 851 | **851** | **PASS** |
+| Tensor policy | exact | exact | **353/3/64/429/2** | **PASS** |
 
----
+Reserve headroom after measurement: PPL ratio **0.009410**, Mean KLD **0.000151**, Same top-p **0.419 pp**. The small KLD reserve is why no further Q3 expansion is claimed safe.
 
-## 8. Frozen Comparison Protocol
+## 10. Frozen comparison protocol
 
-The same BF16 logits and the same protocol are used across Q8_0, Q6_K, Q5_K_M, and Q4_K_M.
+The same BF16 logits and protocol are used across Q8_0, Q6_K, Q5_K_M, Q4_K_M, and Q3_K_M.
 
 | Item | Value |
 |---|---|
-| Reference model | Canonical VeriLoop E2 BF16 GGUF |
 | Corpus | WikiText-2 raw test |
 | Corpus SHA256 | `173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08` |
-| Context length | **2,048** |
+| Context | **2,048** |
 | Chunks | **8** |
 | Seed | **42** |
 | GPU layers | **40** |
 | Flash Attention | **Off** |
 | KV cache | **F16/F16** |
-| Batch size | **512** |
-| Micro-batch size | **512** |
-| Evaluation executable | `llama-perplexity` |
+| Batch / micro-batch | **512 / 512** |
+| Evaluator | `llama-perplexity` |
+| Reference method | `--kl-divergence-base` |
+| Candidate method | `--kl-divergence` |
 | BF16 logits reused | **Yes** |
-| Reference-logit method | `--kl-divergence-base` |
-| Candidate comparison | `--kl-divergence` |
-| llama.cpp revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
-
-The protocol is a **paired quantization-fidelity test**. PPL/KLD numbers from unrelated repositories are not treated as directly comparable unless corpus, tokenizer path, chunking, context, and runtime procedure match.
-
----
-
-## 9. Quantization Build Configuration
-
-### Q8_0
-
-| Item | Value |
-|---|---|
-| Quantization source | Canonical BF16 GGUF |
-| Requantized from another low-bit model | **No** |
-| Quant type | `Q8_0` |
-| Quantized size reported by quantizer | **27,260.56 MiB** |
-| Effective density | **8.50 BPW** |
-| Main tensors | **851** |
-| Q8_0 tensors | **498** |
-| Retained F32 tensors | **353** |
-| Quantization threads | **25** |
-| Quantization time | **38.315 s** |
-| Main build gate | **PASS** |
-| llama.cpp revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
-
-### Q6_K
-
-| Item | Value |
-|---|---|
-| Quantization source | Canonical BF16 GGUF |
-| Requantized from another low-bit model | **No — direct BF16 → Q6_K** |
-| Quant type | `Q6_K` |
-| Quantized size reported by quantizer | **21,049.06 MiB** |
-| Effective density | **6.57 BPW** |
-| Main tensor count | **851** |
-| Quantization threads | **25** |
-| Quantization time | **67.13647 s** |
-| Importance matrix | **496 entries, computed on 582 chunks** |
-| Importance-matrix bytes | **13,642,752** |
-| Importance-matrix SHA256 | `72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b` |
-| Calibration file | `Qwen3.8-27B-calibration-v6.txt` |
-| Main build gate | **PASS** |
-| Main structure gate | **PASS** |
 | llama.cpp revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
 
-### Q5_K_M
+## 11. Release gates
 
-| Item | Value |
-|---|---|
-| Quantization source | Canonical BF16 GGUF |
-| Requantized from another low-bit model | **No — direct BF16 → Q5_K_M** |
-| Base quant type | `Q5_K_M` |
-| Quantized size reported by quantizer | **19,409.82 MiB** |
-| Effective density | **6.05 BPW** |
-| Exact bytes | **20,363,666,176** |
-| Main tensor count | **851** |
-| Retained F32 tensors | **353** |
-| Q5_K tensors | **272** |
-| Protected Q6_K tensors | **226** |
-| Quantization threads | **25** |
-| Quantization time | **120.23577 s** |
-| Importance matrix | **496 entries, computed on 582 chunks** |
-| Importance-matrix bytes | **13,642,752** |
-| Importance-matrix SHA256 | `72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b` |
-| Calibration file | `Qwen3.8-27B-calibration-v6.txt` |
-| Quality policy | `q5_k_m_quality_v3_q6_attention_logits` |
-| Main build gate | **PASS** |
-| Main structure gate | **PASS** |
-| llama.cpp revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
+| Tier | PPL gate | KLD gate | Same-top gate | Structure | Stock runtime | Quantitative result |
+|---|---|---|---|---|---|---|
+| Q8_0 | ≤1.005 → **1.000643 PASS** | ≤0.005 → **0.002176 PASS** | ≥97.0 → **98.815 PASS** | PASS | PASS | PASS |
+| Q6_K | ≤1.005 → **0.999605 PASS** | ≤0.006 → **0.004409 PASS** | ≥97.0 → **98.204 PASS** | PASS | PASS | PASS |
+| Q5_K_M | ≤1.0075 → **1.004450 PASS** | ≤0.008 → **0.006919 PASS** | ≥96.5 → **97.251 PASS** | PASS | PASS | PASS |
+| Q4_K_M | ≤1.010 → **1.004821 PASS** | ≤0.010 → **0.009700 PASS** | ≥96.0 → **96.786 PASS** | PASS | Pending in supplied record | PASS quantitative |
+| **Q3_K_M** | ≤1.015 → **1.004090 PASS** | ≤0.015 → **0.014349 PASS** | ≥95.0 → **95.919 PASS** | **PASS** | **PASS** | **PASS** |
 
-Protected Q6_K tensor groups:
+These thresholds are **VeriLoop release-QA criteria**, not universal industry standards. For Q3, the final run also passed the stricter reserve thresholds shown above.
 
-| Tensor group | Count |
-|---|---:|
-| `output.weight` | **1** |
-| `token_embd.weight` | **1** |
-| `*.attn_v.weight` | **16** |
-| `*.attn_qkv.weight` | **48** |
-| `*.attn_output.weight` | **16** |
-| `*.ffn_down.weight` | **64** |
-| `*.ssm_out.weight` | **48** |
-| `*.attn_q.weight` | **16** |
-| `*.attn_k.weight` | **16** |
-| **Total Q6_K protected tensors** | **226** |
-
-### Q4_K_M
-
-| Item | Value |
-|---|---|
-| Quantization source | Canonical BF16 GGUF |
-| Requantized from another low-bit model | **No — direct BF16 → Q4_K_M** |
-| Base quant type | `Q4_K_M` |
-| Quantized size reported by quantizer | **18,729.82 MiB** |
-| Effective density | **5.84 BPW** |
-| Exact bytes | **19,650,634,496** |
-| Main tensor count | **851** |
-| Retained F32 tensors | **353** |
-| Q4_K tensors | **64** |
-| Q5_K tensors | **208** |
-| Q6_K tensors | **226** |
-| Quantization threads | **25** |
-| Quantization time | **119.16670 s** |
-| Importance matrix | **496 entries, computed on 582 chunks** |
-| Importance-matrix bytes | **13,642,752** |
-| Importance-matrix SHA256 | `72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b` |
-| Calibration file | `Qwen3.8-27B-calibration-v6.txt` |
-| Main build gate | **PASS** |
-| Main structure / precision-policy gate | **PASS** |
-| llama.cpp revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
-
-Precision assignment:
-
-| Tensor group | Precision | Count |
-|---|---:|---:|
-| `output.weight` | Q6_K | 1 |
-| `token_embd.weight` | Q6_K | 1 |
-| `*.attn_v.weight` | Q6_K | 16 |
-| `*.attn_qkv.weight` | Q6_K | 48 |
-| `*.attn_output.weight` | Q6_K | 16 |
-| `*.ffn_down.weight` | Q6_K | 64 |
-| `*.ssm_out.weight` | Q6_K | 48 |
-| `*.attn_q.weight` | Q6_K | 16 |
-| `*.attn_k.weight` | Q6_K | 16 |
-| `*.attn_gate.weight` | Q5_K | 48 |
-| `*.ssm_alpha.weight` | Q5_K | 48 |
-| `*.ssm_beta.weight` | Q5_K | 48 |
-| `*.ffn_gate.weight` | Q5_K | 64 |
-| `*.ffn_up.weight` | Q4_K | 64 |
-
----
-
-## 10. Release Gates
-
-### Q8_0
-
-| Gate | Criterion | Measured | Result |
-|---|---:|---:|---|
-| PPL ratio | ≤ 1.005 | **1.000643** | **PASS** |
-| Mean KLD | ≤ 0.005 | **0.002176** | **PASS** |
-| Same top-p | ≥ 97.0% | **98.815%** | **PASS** |
-| Main tensor count | 851 | **851** | **PASS** |
-| Stock llama.cpp load + generation | required | **PASS** | **PASS** |
-| MTP load + speculative draft smoke | required | **PASS** | **PASS** |
-| Overall release quality | required | **PASS** | **PASS** |
-
-### Q6_K
-
-| Gate | Criterion | Measured | Result |
-|---|---:|---:|---|
-| BF16 baseline identity | exact frozen reference | **Matched** | **PASS** |
-| PPL ratio | ≤ 1.005 | **0.999605** | **PASS** |
-| Mean KLD | ≤ 0.006 | **0.004409** | **PASS** |
-| Same top-p | ≥ 97.0% | **98.204%** | **PASS** |
-| Main tensor count | 851 | **851** | **PASS** |
-| Main structure | required | **PASS** | **PASS** |
-| MTP structure | required | **PASS** | **PASS** |
-| Quantitative quality | required | **PASS** | **PASS** |
-| Final stock llama.cpp main runtime | required | **PASS** | **PASS** |
-| Final stock llama.cpp MTP smoke | required | **PASS** | **PASS** |
-| Final SHA256 / immutable identity | required | **Frozen** | **PASS** |
-| Remote exact-size verification | required | **PASS** | **PASS** |
-| Overall release quality | required | **PASS** | **PASS** |
-
-### Q5_K_M
-
-| Gate | Criterion | Measured | Result |
-|---|---:|---:|---|
-| BF16 baseline identity | exact frozen reference | **Matched** | **PASS** |
-| PPL ratio | ≤ 1.0075 | **1.004450** | **PASS** |
-| Mean KLD | ≤ 0.008 | **0.006919** | **PASS** |
-| Same top-p | ≥ 96.5% | **97.251%** | **PASS** |
-| Main tensor count | 851 | **851** | **PASS** |
-| Mixed-precision protected-path policy | exact declared counts | **Matched** | **PASS** |
-| Main structure | required | **PASS** | **PASS** |
-| Quantitative quality | required | **PASS** | **PASS** |
-| Final stock llama.cpp main runtime | required | **PASS** | **PASS** |
-| Final stock llama.cpp MTP smoke | required | **PASS** | **PASS** |
-| Final SHA256 / immutable identity | required | **Frozen** | **PASS** |
-| Remote exact-size + LFS-SHA verification | required | **PASS** | **PASS** |
-| Overall release quality | required | **PASS** | **PASS** |
-
-### Q4_K_M
-
-| Gate | Criterion | Measured | Result |
-|---|---:|---:|---|
-| BF16 baseline identity | exact frozen reference | **Matched** | **PASS** |
-| PPL ratio | ≤ 1.0100 | **1.004821** | **PASS** |
-| Mean KLD | ≤ 0.0100 | **0.009700** | **PASS** |
-| Same top-p | ≥ 96.0% | **96.786%** | **PASS** |
-| Main tensor count | 851 | **851** | **PASS** |
-| Mixed-precision policy | exact declared counts | **Matched** | **PASS** |
-| Main structure | required | **PASS** | **PASS** |
-| Quantitative quality | required | **PASS** | **PASS** |
-| Final stock llama.cpp runtime | required before release | **Pending** | **PENDING** |
-| Final one-time SHA256 / immutable identity | required before release | **Pending** | **PENDING** |
-| Overall release quality | required | **Pending runtime + identity** | **PENDING** |
-
-These numerical thresholds are **VeriLoop release-QA gates**, not universal industry standards. They were not relaxed after observing Q5_K_M or Q4_K_M results.
-
----
-
-## 11. Runtime Validation
-
-### BF16
-
-| Item | Result |
-|---|---|
-| Main-model conversion | **PASS** |
-| Main tensor audit | **851 / PASS** |
-| Source-checkpoint immutability gate | **PASS** |
-| llama.cpp main-model load | **PASS** |
-| OpenAI-compatible endpoint validation | **PASS** |
-| Reference runtime context used in initial GGUF release gate | **32,768 tokens** |
-
-### Q8_0
+## 12. Q3 stock-runtime and MTP validation
 
 | Item | Result |
 |---|---|
-| Main Q8_0 load | **PASS** |
-| OpenAI-compatible generation | **PASS** |
-| Q8 MTP load | **PASS** |
-| Speculative-draft path | **PASS** |
-| MTP compatibility smoke context | **8,192 tokens** |
-| Draft tokens proposed / accepted | **3 / 3** |
-
-### Q6_K
-
-| Item | Result |
-|---|---|
-| Main build | **PASS** |
-| Main structure | **PASS** |
-| MTP build / structure | **PASS** |
-| BF16-paired quantitative fidelity | **PASS** |
-| Stock llama.cpp main-model generation | **PASS** |
-| Stock llama.cpp MTP draft smoke | **PASS** |
-| Runtime validation context | **8,192 tokens** |
-| Main GPU offload | **All model layers** |
-| Draft GPU layers | **0 in validated conservative smoke** |
-| Final release identity | **Frozen** |
-
-### Q5_K_M
-
-| Item | Result |
-|---|---|
-| Main build | **PASS** |
-| Main structure | **PASS** |
-| Protected-path precision audit | **PASS** |
-| BF16-paired quantitative fidelity | **PASS** |
-| Final stock llama.cpp main-model generation | **PASS** |
-| Final stock llama.cpp MTP draft smoke | **PASS** |
-| Runtime validation context | **8,192 tokens** |
-| Main GPU offload | **All model layers** |
-| Draft GPU layers | **0 in validated conservative smoke** |
-| Draft tokens accepted / proposed | **21 / 30 (0.70000)** |
-| Final release identity | **Frozen** |
-| Remote exact-size + LFS-SHA verification | **PASS** |
-
-### Q4_K_M
-
-| Item | Current state |
-|---|---|
-| Main build | **PASS** |
-| Main structure / precision-policy audit | **PASS** |
-| BF16-paired quantitative fidelity | **PASS** |
-| Final stock llama.cpp main-model generation | **Pending** |
-| Final release identity | **Pending** |
-
-No runtime PASS is claimed before the corresponding smoke is actually executed.
-
----
-
-## 12. Size Efficiency
-
-| Artifact | Exact size | Relative to BF16 main |
-|---|---:|---:|
-| BF16 main | **50.112870 GiB** | 100% reference |
-| Q8_0 main | **26.631882 GiB** | **53.14% of BF16 / 46.86% smaller** |
-| Q6_K main | **20.565961 GiB** | **41.04% of BF16 / 58.96% smaller** |
-| Q5_K_M main | **18.965142 GiB** | **37.84% of BF16 / 62.16% smaller** |
-| Q4_K_M main | **18.301080 GiB** | **36.52% of BF16 / 63.48% smaller** |
-| Q6_K vs Q8_0 | — | **22.78% smaller than Q8_0** |
-| Q5_K_M vs Q8_0 | — | **28.79% smaller than Q8_0** |
-| Q5_K_M vs Q6_K | — | **7.78% smaller than Q6_K** |
-| Q4_K_M vs Q8_0 | — | **31.28% smaller than Q8_0** |
-| Q4_K_M vs Q6_K | — | **11.01% smaller than Q6_K** |
-| Q4_K_M vs Q5_K_M | — | **3.50% smaller than Q5_K_M / 0.664 GiB saved** |
-
-## 13. Throughput Status
-
-A standalone Q8_0 `llama-bench` run produced:
+| Stock llama.cpp main-only load | **PASS** |
+| Main-only generation | **PASS** |
+| MTP interface detection (`--model-draft`) | **PASS** |
+| MTP GPU-offload interface (`--gpu-layers-draft`) | **PASS** |
+| Current speculative limit flag (`--spec-draft-n-max 8`) | **PASS** |
+| Main + MTP load | **PASS** |
+| Main + MTP generation | **PASS** |
+| MTP engagement log evidence | **PASS** |
+| Greedy target-output identity | **PASS** |
+| Main-only text SHA256 | `43944ea63376bab89382150a3e4fbcb3739901c82e791abfa7847c13df0391a9` |
+| Main+MTP text SHA256 | `43944ea63376bab89382150a3e4fbcb3739901c82e791abfa7847c13df0391a9` |
+| Runtime validation context | **2,048** |
+| Validated MTP bytes | **2,008,056,096** |
+| Validated MTP SHA256 | `0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671` |
+| Overall Q3 runtime/MTP gate | **PASS** |
+
+The validated MTP object is byte-identical to the existing Q5/Q6 18-tensor Q4_0 companion. This avoids publishing a duplicate binary solely to change its filename.
+
+## 13. Size efficiency
+
+| Artifact | GiB | Smaller than BF16 | Marginal size reduction |
+|---|---:|---:|---|
+| BF16 | **50.112870** | — | — |
+| Q8_0 | **26.631882** | **46.86%** | — |
+| Q6_K | **20.565961** | **58.96%** | **22.78% vs Q8** |
+| Q5_K_M | **18.965142** | **62.16%** | **7.78% vs Q6** |
+| Q4_K_M | **18.301080** | **63.48%** | **3.50% vs Q5 / 0.664 GiB** |
+| **Q3_K_M** | **16.825745** | **66.42%** | **8.06% vs Q4 / 11.28% vs Q5 / 18.19% vs Q6** |
 
-| Workload | Q8_0 |
-|---|---:|
-| Prompt processing (`pp512`) | **3102.722776 tok/s** |
-| Token generation (`tg128`) | **41.999677 tok/s** |
+### Pareto interpretation
 
-No paired BF16→Q8 speedup is claimed because the paired BF16 speed run did not complete under the original full-offload configuration.
+- **Q6_K** is the strongest overall balance because its large BF16 footprint reduction is achieved without detectable PPL degradation and with low distributional drift.
+- **Q5_K_M** is the best measured step down when memory matters: the extra **7.78%** saving versus Q6 costs a controlled amount of divergence.
+- **Q4_K_M** saves only **3.50%** beyond Q5 while KLD increases to **0.009700**; it is an intermediate candidate rather than a new sweet spot.
+- **Q3_K_M** materially reduces footprint again, but KLD reaches **0.014349** and the reserve margin becomes narrow. It is best understood as the **validated lower-footprint floor of the current ladder**, not the balance optimum.
 
-**Q6_K, Q5_K_M, and Q4_K_M do not have a frozen paired speed benchmark in this record, so no speedup percentage is claimed.** This avoids mixing quantization fidelity with hardware-specific throughput.
+## 14. Throughput status
 
----
+A standalone Q8_0 `llama-bench` run recorded **3102.722776 tok/s** prompt processing (`pp512`) and **41.999677 tok/s** token generation (`tg128`). There is no frozen paired BF16 throughput result. Q6/Q5/Q4/Q3 do not have a frozen paired speed benchmark in this record, so **no universal speedup percentage is claimed**.
 
-## 14. Model Capability vs Quantization Fidelity
+## 15. Model capability vs quantization fidelity
 
 ```text
 Parent-model benchmark evidence
-    → establishes VeriLoop E2 capability
+    -> establishes VeriLoop E2 capability
 
 Canonical BF16 GGUF
-    → establishes the deployment/reference baseline
+    -> establishes the deployment/reference baseline
 
-BF16 → Q8_0 paired PPL/KLD/token analysis
-    → establishes Q8_0 quantization fidelity
+BF16-paired PPL / KLD / token analysis
+    -> establishes quantization fidelity for each GGUF tier
 
-BF16 → Q6_K paired PPL/KLD/token analysis
-    → establishes Q6_K quantization fidelity
+Stock llama.cpp runtime + MTP validation
+    -> establishes runtime compatibility for the tested artifact
 
-BF16 → Q5_K_M paired PPL/KLD/token analysis
-    → establishes Q5_K_M quantization fidelity
-
-BF16 → Q4_K_M paired PPL/KLD/token analysis
-    → establishes Q4_K_M quantization fidelity
+No per-quant downstream task rerun
+    -> no quant-specific SWE/AIME/GPQA score-loss percentage is claimed
 ```
 
-The full task-benchmark campaign is not repeatedly relabeled as a quant-specific benchmark campaign.
-
----
-
-## 15. Precision-Tier Decision Matrix
-
-| Tier | State | Quality role | BF16-paired fidelity | Recommended use |
-|---|---|---|---|---|
-| **BF16** | Released | Reference fidelity | Baseline | Reproducibility / reference |
-| **Q8_0** | Released | High-fidelity local | **PASS** | Maximum local fidelity |
-| **Q6_K** | Released | **Overall quality / efficiency sweet spot** | **PASS** | **Default recommendation when ~20.6 GiB fits** |
-| **Q5_K_M** | Released | **Memory-constrained quality sweet spot** | **PASS** | Tighter memory budgets with controlled drift |
-| **Q4_K_M** | Release candidate | Validated mixed-precision 4-bit-core tier | **PASS** | Use when the additional **0.664 GiB** saving vs Q5_K_M matters after runtime/identity PASS |
-| `IQ4_XS` | Planned | Compact 4-bit class | Pending | Maximum footprint reduction |
-
-## 16. Current Decision
-
-```text
-BF16_REFERENCE_GATE=PASS
-
-Q8_STRUCTURE_GATE=PASS
-Q8_QUANTITATIVE_QUALITY_GATE=PASS
-Q8_STOCK_LLAMA_CPP_COMPATIBILITY=PASS
-Q8_RELEASE_QUALITY_GATE=PASS
-
-Q6_BUILD_STAGE=PASS
-Q6_STRUCTURE_GATE=PASS
-Q6_MTP_STRUCTURE_GATE=PASS
-Q6_BF16_BASELINE_IDENTITY_GATE=PASS
-Q6_PPL_RATIO_GATE=PASS
-Q6_MEAN_KLD_GATE=PASS
-Q6_SAME_TOP_P_GATE=PASS
-Q6_QUANTITATIVE_QUALITY_GATE=PASS
-Q6_FINAL_RUNTIME_GATE=PASS
-Q6_FINAL_RELEASE_IDENTITY=PASS
-Q6_REMOTE_RELEASE_GATE=PASS
-
-Q5_BUILD_STAGE=PASS
-Q5_STRUCTURE_GATE=PASS
-Q5_PROTECTED_PRECISION_POLICY_GATE=PASS
-Q5_BF16_BASELINE_IDENTITY_GATE=PASS
-Q5_PPL_RATIO_GATE=PASS
-Q5_MEAN_KLD_GATE=PASS
-Q5_SAME_TOP_P_GATE=PASS
-Q5_QUANTITATIVE_QUALITY_GATE=PASS
-Q5_FINAL_RUNTIME_GATE=PASS
-Q5_FINAL_RELEASE_IDENTITY=PASS
-Q5_REMOTE_RELEASE_GATE=PASS
-
-Q4_V3_BUILD_STAGE=PASS
-Q4_V3_STRUCTURE_POLICY_GATE=PASS
-Q4_V3_BF16_BASELINE_IDENTITY_GATE=PASS
-Q4_V3_PPL_RATIO_GATE=PASS
-Q4_V3_MEAN_KLD_GATE=PASS
-Q4_V3_SAME_TOP_P_GATE=PASS
-Q4_V3_QUANTITATIVE_QUALITY_GATE=PASS
-Q4_V3_FINAL_RUNTIME_GATE=PENDING
-Q4_V3_FINAL_RELEASE_IDENTITY=PENDING
-```
+## 16. Precision-tier decision matrix
 
-**Current evidence-supported conclusion:** **Q6_K is the overall VeriLoop E2 deployment sweet spot.** It is **58.96% smaller than BF16** while showing **no detectable PPL degradation** under the frozen paired protocol, with Mean KLD **0.004409 ± 0.000953** and Same top-p **98.204 ± 0.147%**.
+| Tier | Quality role | Main file | Recommendation |
+|---|---|---:|---|
+| BF16 | Reference | 50.113 GiB | Research / reproducibility |
+| Q8_0 | High fidelity | 26.632 GiB | Use when fidelity dominates footprint |
+| **Q6_K** | **Overall sweet spot** | **20.566 GiB** | **Default** |
+| **Q5_K_M** | **Memory-quality sweet spot** | **18.965 GiB** | **Use when Q6 footprint is too high** |
+| Q4_K_M | Intermediate candidate | 18.301 GiB | Lower drift than Q3, but its own final runtime/identity is pending in this supplied record |
+| **Q3_K_M** | **Minimum-footprint validated** | **16.826 GiB** | **Use when footprint is the dominant constraint and higher distributional drift is acceptable** |
 
-**Q5_K_M is the memory-constrained sweet spot.** It is **62.16% smaller than BF16** and **7.78% smaller than Q6_K**, with measured PPL ratio **1.004450 ± 0.001361 (+0.4450%)**, Mean KLD **0.006919 ± 0.000945**, and Same top-p **97.251 ± 0.181%**; its full runtime, immutable identity, and remote release gates are now PASS.
+## 17. Current evidence-supported conclusion
 
-**Q4_K_M is quantitatively validated but not the default sweet spot.** It is **63.48% smaller than BF16** and **3.50% smaller than Q5_K_M**, with PPL ratio **1.004821 ± 0.001722 (+0.4821%)**, Mean KLD **0.009700 ± 0.001110**, Same top-p **96.786 ± 0.195%**, and **99.76%** log-PPL correlation. It remains a release candidate until stock-runtime validation and immutable SHA256 identity are complete.
+**Q6_K remains the overall VeriLoop E2 GGUF sweet spot. Q5_K_M remains the memory-constrained quality sweet spot. Q3_K_M is the current minimum-footprint validated tier and the lower bound of the present release ladder, but it is not promoted to the default sweet spot because its KLD and top-token disagreement consume substantially more of the fidelity budget.**

README.md

@@ -14,6 +14,8 @@ tags:
   - q6_k
   - q5_k_m
   - q4_k_m
+  - q3_k
+  - mixed-precision
   - imatrix
   - mtp
   - speculative-decoding
@@ -35,7 +37,7 @@ tags:
 
 <p align="center">
   <strong>Official llama.cpp distribution of VeriLoop E2</strong><br>
-  <em>BF16 reference · Q8_0 high fidelity · Q6_K overall sweet spot · Q5_K_M memory-efficient quality · Q4_K_M validated 4-bit-core candidate</em><br><br>
+  <em>BF16 reference · Q8_0 high fidelity · Q6_K overall sweet spot · Q5_K_M memory-quality sweet spot · Q3_K_M minimum-footprint validated tier</em><br><br>
   <strong>27B post-trained model for code, mathematics, and physics · 262K native context · Apache License 2.0</strong><br>
   <strong>Developed by Tsinghua SIGS Robot Lab · Libo Wang</strong>
 </p>
@@ -43,12 +45,10 @@ tags:
 <p align="center">
   <a href="https://www.apache.org/licenses/LICENSE-2.0"><img src="https://img.shields.io/badge/License-Apache--2.0-2F80ED?style=flat-square" alt="License: Apache 2.0"></a>
   <img src="https://img.shields.io/badge/Format-GGUF-111827?style=flat-square" alt="Format: GGUF">
-  <img src="https://img.shields.io/badge/Reference-BF16-5B5BD6?style=flat-square" alt="Reference: BF16">
   <img src="https://img.shields.io/badge/Q8__0-High--Fidelity-7C3AED?style=flat-square" alt="Q8_0: High Fidelity">
-  <img src="https://img.shields.io/badge/Q6__K-Quality%2FEfficiency-0A7F6F?style=flat-square" alt="Q6_K: Quality/Efficiency">
-  <img src="https://img.shields.io/badge/Q5__K__M-Memory--Efficient%20Quality-0F766E?style=flat-square" alt="Q5_K_M: Memory-Efficient Quality">
-  <img src="https://img.shields.io/badge/Q4__K__M-4--bit--core%20Candidate-64748B?style=flat-square" alt="Q4_K_M: validated 4-bit-core release candidate">
-  <img src="https://img.shields.io/badge/Quant%20QC-Q8%2FQ6%2FQ5%2FQ4%20PASS-16A34A?style=flat-square" alt="Quantization QC: Q8 Q6 Q5 Q4 PASS">
+  <img src="https://img.shields.io/badge/Q6__K-Overall%20Sweet%20Spot-0A7F6F?style=flat-square" alt="Q6_K: Overall Sweet Spot">
+  <img src="https://img.shields.io/badge/Q5__K__M-Memory--Quality%20Sweet%20Spot-0F766E?style=flat-square" alt="Q5_K_M: Memory Quality Sweet Spot">
+  <img src="https://img.shields.io/badge/Q3__K__M-Validated%205.37%20BPW-334155?style=flat-square" alt="Q3_K_M: validated 5.37 BPW mixed precision">
   <img src="https://img.shields.io/badge/Runtime-llama.cpp-0A7F6F?style=flat-square" alt="Runtime: llama.cpp">
 </p>
 
@@ -61,384 +61,152 @@ tags:
 
 ---
 
-## Overview
+## Which file should I download?
 
-**VeriLoop E2 GGUF** is the official llama.cpp-oriented distribution of **VeriLoop E2**, a 27B post-trained model built on **Qwen3.8-27B** for code, software engineering, mathematics, and physics.
+| Priority | Recommended file | Main size | Measured retention | Use when |
+|---|---|---:|---|---|
+| **Default** | **`VeriLoop-E2-Q6_K.gguf`** | **20.566 GiB** | BF16-paired PPL parity within uncertainty; KLD **0.004409**; Same top-p **98.204%** | You want the strongest overall quality / footprint balance |
+| **Tighter memory** | **`VeriLoop-E2-Q5_K_M.gguf`** | **18.965 GiB** | PPL **+0.4450%**; KLD **0.006919**; Same top-p **97.251%** | You need lower memory while keeping distributional drift controlled |
+| **Smallest validated main file** | **`VeriLoop-E2-Q3_K_M.gguf`** | **16.826 GiB** | PPL **+0.4090%**; KLD **0.014349**; Same top-p **95.919%** | Footprint is the primary constraint and you accept higher distributional drift |
+| **High fidelity** | `VeriLoop-E2-Q8_0.gguf` | **26.632 GiB** | PPL **+0.0643%**; KLD **0.002176**; Same top-p **98.815%** | You prioritize low-bit fidelity over footprint |
+| **Reference** | `VeriLoop-E2-BF16.gguf` | **50.113 GiB** | Canonical reference | Reproducibility, attribution, quantization research |
 
-The release follows a **reference-first, measurement-driven precision ladder**. Every measured low-bit tier is built directly from the canonical BF16 GGUF rather than requantized from another low-bit artifact, then evaluated against the same frozen BF16 logit reference.
+**Current sweet spot.** **Q6_K remains the overall quality / efficiency sweet spot.** It is **58.96% smaller than BF16** while showing no detectable PPL degradation under the frozen paired protocol. **Q5_K_M remains the memory-constrained quality sweet spot**: it is **7.78% smaller than Q6_K** with controlled measured drift. **Q3_K_M is the current minimum-footprint validated tier, not the default sweet spot**: it saves another **11.28% vs Q5_K_M** but Mean KLD rises from **0.006919** to **0.014349** and Same top-p falls from **97.251%** to **95.919%**.
 
-### Precision ladder and deployment sweet spots
+> **Important Q3 naming note:** `VeriLoop-E2-Q3_K_M.gguf` is a **root-cause-guided mixed-precision artifact**, not a uniform stock 3-bit model. It contains **353 F32 + 3 Q3_K + 64 Q4_K + 429 Q5_K + 2 Q6_K tensors**, for **5.37 effective BPW**. The filename identifies the release tier / minimum active precision; the exact tensor policy is published below and in `RELEASE_MANIFEST.json`.
 
-| Tier | Role | Main size | BF16 footprint reduction | Measured BF16-paired fidelity | Release state |
-|---|---|---:|---:|---|---|
-| **BF16** | Reference fidelity | **53.808 GB / 50.113 GiB** | — | Canonical reference | **Released / verified** |
-| **Q8_0** | High-fidelity local | **28.596 GB / 26.632 GiB** | **46.86%** | PPL **+0.0643%**, Mean KLD **0.002176**, Same top-p **98.815%** | **Released / verified** |
-| **Q6_K** | **Overall quality / efficiency sweet spot** | **22.083 GB / 20.566 GiB** | **58.96%** | **No detectable PPL degradation**, Mean KLD **0.004409**, Same top-p **98.204%** | **Released / verified** |
-| **Q5_K_M** | **Memory-constrained quality sweet spot** | **20.364 GB / 18.965 GiB** | **62.16%** | PPL **+0.4450%**, Mean KLD **0.006919**, Same top-p **97.251%** | **Released / verified** |
-| **Q4_K_M** | Validated mixed-precision 4-bit-core candidate | **19.651 GB / 18.301 GiB** | **63.48%** | PPL **+0.4821%**, Mean KLD **0.009700**, Same top-p **96.786%** | **Quantitative QC PASS; runtime/identity pending** |
+## Precision ladder
 
-**Deployment recommendation.** On the measured Pareto frontier, **Q6_K is the overall sweet spot**: it removes **58.96%** of the BF16 main-file footprint while showing no detectable PPL degradation under the frozen paired protocol. **Q5_K_M is the memory-constrained sweet spot** when saving an additional **7.78%** versus Q6_K materially changes fit. Q4_K_M passes its frozen quantitative gates, but it saves only **3.50%** versus Q5_K_M while Mean KLD rises from **0.006919** to **0.009700** and Same top-p decreases from **97.251%** to **96.786%**; it is therefore positioned as a validated 4-bit-core candidate rather than the default recommendation.
+| Tier | Role | Main size | Reduction vs BF16 | Effective density | Release state |
+|---|---|---:|---:|---:|---|
+| **BF16** | Canonical reference | **53.808 GB / 50.113 GiB** | — | 16-bit-class | Verified reference |
+| **Q8_0** | High-fidelity local | **28.596 GB / 26.632 GiB** | **46.86%** | **8.50 BPW** | Verified / Release Quality PASS |
+| **Q6_K** | **Overall sweet spot** | **22.083 GB / 20.566 GiB** | **58.96%** | **6.57 BPW** | Verified / Release Quality PASS |
+| **Q5_K_M** | **Memory-quality sweet spot** | **20.364 GB / 18.965 GiB** | **62.16%** | **6.05 BPW** | Verified / Release Quality PASS |
+| **Q4_K_M** | Intermediate lower-footprint candidate | **19.651 GB / 18.301 GiB** | **63.48%** | **5.84 BPW** | Quantitative QC PASS; its own runtime/identity still pending in this record |
+| **Q3_K_M** | **Minimum-footprint validated tier** | **18.067 GB / 16.826 GiB** | **66.42%** | **5.37 BPW** | **Full release PASS; stock runtime + MTP + remote identity verified** |
 
-These measurements quantify **quantization fidelity under one fixed paired protocol**. They are not universal percentages of downstream code, mathematics, physics, or agent-benchmark capability loss.
+## Quantization-retention benchmark
 
-## Headline Quantization-Retention Benchmark
+All values below are measured against the **same frozen BF16 logits** using the same corpus, context, chunking, seed, cache dtype, and llama.cpp revision.
 
-| Tier | Mean PPL | PPL ratio vs BF16 | Observed PPL change | Mean KLD | Same top-p | log-PPL correlation |
+| Tier | Mean PPL | PPL ratio vs BF16 | Relative PPL change | Mean KLD | Same top-p | log-PPL correlation |
 |---|---:|---:|---:|---:|---:|---:|
-| **BF16** | **4.840423 ± 0.119931** | 1.000000 reference | — | 0 reference | 100% reference | 100% reference |
+| **BF16** | **4.840423 ± 0.119931** | 1.000000 | — | 0 reference | 100% reference | 100% reference |
 | **Q8_0** | **4.843536 ± 0.120062** | **1.000643 ± 0.000754** | **+0.0643%** | **0.002176 ± 0.000668** | **98.815 ± 0.120%** | **99.95%** |
 | **Q6_K** | **4.838514 ± 0.119694** | **0.999605 ± 0.001222** | **−0.0395% observed; parity within uncertainty** | **0.004409 ± 0.000953** | **98.204 ± 0.147%** | **99.88%** |
 | **Q5_K_M** | **4.861965 ± 0.120630** | **1.004450 ± 0.001361** | **+0.4450%** | **0.006919 ± 0.000945** | **97.251 ± 0.181%** | **99.85%** |
 | **Q4_K_M** | **4.863760 ± 0.120720** | **1.004821 ± 0.001722** | **+0.4821%** | **0.009700 ± 0.001110** | **96.786 ± 0.195%** | **99.76%** |
+| **Q3_K_M** | **4.860222 ± 0.120606** | **1.004090 ± 0.002181** | **+0.4090%** | **0.014349 ± 0.001985** | **95.919 ± 0.219%** | **99.62%** |
 
-> **Interpretation:** Q6_K's negative PPL point estimate is not treated as a quality gain because the reported uncertainty spans parity. Q5_K_M and Q4_K_M both show measurable but controlled fidelity cost. Q4_K_M clears its predeclared quantitative gates, but its incremental footprint reduction over Q5_K_M is only **3.50%**, so Q6_K remains the overall quality/efficiency sweet spot and Q5_K_M remains the tighter-memory sweet spot.
-
-## Choose the Right Precision
+**How to read the loss numbers.** The measured **+0.4090% Q3 PPL change is a quantization-fidelity result, not a claim that SWE-bench, AIME, GPQA, coding-agent, mathematics, or physics scores fall by 0.4090%.** Per-quant downstream task suites were **not rerun**, so no downstream benchmark-loss percentage is claimed. KLD and Same top-p measure distributional divergence directly and are essential here: Q3's PPL point estimate remains close to BF16 while its KLD is materially higher than Q5/Q4, which is why Q3 is positioned as a footprint-first tier rather than the default sweet spot.
 
-| Variant | Positioning | Main model size | Fidelity / efficiency role | Recommended for |
-|---|---|---:|---|---|
-| **BF16** | Reference fidelity | **53.808 GB / 50.113 GiB** | Canonical reference; no low-bit quantization | Research, reproducibility, attribution, quantization baselines |
-| **Q8_0** | High-fidelity local | **28.596 GB / 26.632 GiB** | Lowest measured distributional drift among released low-bit tiers | Quality-first local inference |
-| **Q6_K** | **Overall quality / efficiency sweet spot** | **22.083 GB / 20.566 GiB** | **58.96% smaller than BF16; no detectable PPL degradation** | Default recommendation when ~20.6 GiB model storage fits |
-| **Q5_K_M** | **Memory-constrained quality sweet spot** | **20.364 GB / 18.965 GiB** | **62.16% smaller than BF16; +0.4450% measured PPL drift** | Tighter memory budgets with controlled fidelity cost |
-| **Q4_K_M** | Validated 4-bit-core candidate | **19.651 GB / 18.301 GiB** | **63.48% smaller than BF16; +0.4821% PPL drift; quantitative QC PASS** | Deployments where the final ~0.664 GiB reduction versus Q5_K_M matters |
-| `IQ4_XS` | Compact 4-bit-class tier | Pending validation | Planned BF16-relative measurement | Maximum footprint reduction |
+### Downstream benchmark-loss status
 
-### BF16 — reference fidelity
+| Evidence layer | What is measured | What is **not** claimed |
+|---|---|---|
+| Parent VeriLoop E2 benchmark suite | Parent-model capability on nine public benchmark tracks | That every quant reproduces the parent score exactly |
+| BF16-paired GGUF benchmark | PPL ratio, KL divergence, token-probability drift, Same top-p, correlation | Universal task-score loss or gain |
+| Q3 stock runtime + MTP validation | Main load/generation, MTP load/generation, greedy target-output identity | Universal throughput speedup |
 
-Use **BF16** for reproducibility, attribution, or quantization research. It is the canonical GGUF against which every lower-bit tier is measured.
+## Q3_K_M — root-cause-guided low-footprint release
 
-### Q8_0 — high-fidelity local
+The Q3 release was **not** produced by applying a stock Q3 preset to the whole model. Broad Q3 policies failed the frozen fidelity gates, so the final artifact was derived by root-cause isolation and a KL-budgeted tensor policy.
 
-Use **Q8_0** when the closest measured low-bit behavior to BF16 matters more than minimizing footprint. The main file is **46.86% smaller than BF16** while retaining **98.815% same-top agreement** under the frozen test.
+### Final Q3 tensor policy
 
-### Q6_K — overall quality / efficiency sweet spot
+| Precision | Tensor assignment | Count |
+|---|---|---:|
+| F32 | Non-quantized tensors retained by the GGUF conversion | **353** |
+| Q6_K | `output.weight`, `token_embd.weight` | **2** |
+| Q5_K | Remaining attention / SSM / FFN-gate / FFN-down tensors | **429** |
+| Q4_K | all `*.ffn_up.weight` | **64** |
+| Q3_K | `blk.1.ffn_down.weight`, `blk.0.ffn_down.weight`, `blk.3.ffn_down.weight` | **3** |
+| **Total** |  | **851** |
 
-Use **Q6_K** as the default precision recommendation when its **20.566 GiB** main-file footprint fits. It is **58.96% smaller than BF16**, and its paired PPL result is statistically consistent with BF16 parity under the frozen protocol.
+### Why only three Q3 tensors?
 
-### Q5_K_M — memory-constrained quality sweet spot
+The development evidence showed that **FFN-up Q4→Q3 was the dominant failure source** in the first broad Q3 design; full-family FFN-down and QKV Q3 policies also produced excessive distributional drift. A 33-tensor low-sensitivity shortlist then passed PPL and Same top-p but still failed Mean KLD (**0.018066 > 0.015**). The final policy therefore made KLD the binding constraint and admitted only the three lowest-sensitivity FFN-down tensors. The resulting artifact passed both the hard Q3 gate and a stricter engineering reserve gate.
 
-Use **Q5_K_M** when fitting below Q6_K matters. Its main file is **18.965 GiB**, **62.16% smaller than BF16**, and **7.78% smaller than Q6_K**. The measured trade-off is **+0.4450% PPL**, **Mean KLD 0.006919**, and **97.251% Same top-p**.
+### Q3 final fidelity gates
 
-### Q4_K_M — validated 4-bit-core candidate
+| Gate | Hard threshold | Reserve threshold | Measured | Result |
+|---|---:|---:|---:|---|
+| PPL ratio | ≤ **1.0150** | ≤ **1.0135** | **1.004090** | **PASS / PASS** |
+| Mean KLD | ≤ **0.0150** | ≤ **0.0145** | **0.014349** | **PASS / PASS** |
+| Same top-p | ≥ **95.0%** | ≥ **95.5%** | **95.919%** | **PASS / PASS** |
+| Exact tensor structure | declared policy | declared policy | **851 tensors; 0 mismatch** | **PASS** |
+| Stock llama.cpp main runtime | required | required | **PASS** | **PASS** |
+| MTP load + generation | required | required | **PASS** | **PASS** |
+| Greedy target-output identity with/without MTP | exact match | exact match | **SHA256 identical** | **PASS** |
 
-Q4_K_M is a **mixed-precision Q4_K_M policy**, not a uniform 4-bit model. The 64 `*.ffn_up.weight` tensors remain Q4_K, `*.ffn_gate.weight` is retained at Q5_K after a controlled sensitivity correction, and high-sensitivity output/embedding/attention/FFN-down/SSM-output paths remain Q6_K. This yields **5.84 effective BPW** and an **18.301 GiB** main file. The tier passes all frozen quantitative gates, but final stock-runtime and immutable-identity validation are still pending.
+The Q3 Mean-KLD reserve margin is intentionally small (**0.000151**). The release therefore does **not** claim that further low-bit expansion is safe under the same policy.
 
-## Available Files
+## Available files
 
 ### Main models
 
-| File | Precision / quant | Exact size | Role | Status |
+| File | Quantization | Exact size | Role | Status |
 |---|---|---:|---|---|
-| `VeriLoop-E2-BF16.gguf` | BF16 / selected F32 tensors | **53,808,284,000 bytes / 50.113 GiB** | Canonical reference | **Verified** |
-| `VeriLoop-E2-Q8_0.gguf` | Q8_0 / retained F32 tensors | **28,595,765,600 bytes / 26.632 GiB** | High-fidelity local | **Verified / Release Quality PASS** |
-| `VeriLoop-E2-Q6_K.gguf` | Q6_K, imatrix-assisted | **22,082,532,096 bytes / 20.566 GiB** | **Overall quality / efficiency sweet spot** | **Verified / Release Quality PASS** |
-| `VeriLoop-E2-Q5_K_M.gguf` | Mixed Q5_K_M + protected Q6_K, imatrix-assisted | **20,363,666,176 bytes / 18.965 GiB** | **Memory-constrained quality sweet spot** | **Verified / Release Quality PASS** |
-| `VeriLoop-E2-Q4_K_M.gguf` | Mixed Q4/Q5/Q6, imatrix-assisted | **19,650,634,496 bytes / 18.301 GiB** | Validated 4-bit-core candidate | **Quantitative QC PASS; runtime/SHA pending; not yet public** |
-
-### Optional MTP draft files
-
-| File | Role | Exact size | Notes |
-|---|---|---:|---|
-| `mtp-VeriLoop-E2-BF16.gguf` | BF16 MTP companion | ~6.04 GB | Optional speculative-decoding artifact |
-| `mtp-VeriLoop-E2-Q8_0.gguf` | Q8 release MTP companion | **3,287,070,240 bytes / 3.061 GiB** | Stock llama.cpp load + speculative-draft smoke validated |
-| `mtp-VeriLoop-E2-Q6_K.gguf` | Q6 release MTP companion | **2,008,056,096 bytes / 1.870 GiB** | 18-tensor Q4_0 draft; stock-runtime smoke validated |
-| `mtp-VeriLoop-E2-Q5_K_M.gguf` | Q5 release MTP companion | **2,008,056,096 bytes / 1.870 GiB** | **Verified; stock-runtime MTP smoke PASS** |
-
-Artifact identity and validation states are tracked in [`RELEASE_MANIFEST.json`](./RELEASE_MANIFEST.json).
-
-## BF16-Paired Quantization Fidelity Benchmark
-
-### Q8_0 — high-fidelity tier
-
-| Metric | BF16 reference | Q8_0 | Observed drift |
-|---|---:|---:|---:|
-| Mean PPL | **4.840423 ± 0.119931** | **4.843536 ± 0.120062** | **+0.003113 ± 0.003651** |
-| PPL ratio | 1.000000 | **1.000643 ± 0.000754** | **+0.0643%** |
-| Mean KLD vs BF16 | 0 | **0.002176 ± 0.000668** | lower is better |
-| Same top-p | 100% reference | **98.815 ± 0.120%** | 1.185 pp disagreement |
-| RMS token-probability Δp | 0% reference | **1.251 ± 0.137%** | distributional-noise scale |
-| Mean token-probability Δp | 0% reference | **−0.007 ± 0.014%** | near-zero directional bias |
-| log-PPL correlation | 100% reference | **99.95%** | very high agreement |
-
-**Q8_0 quantitative gate: PASS.**
-
-### Q6_K — quality / efficiency tier
-
-| Metric | BF16 reference | Q6_K | Observed drift |
-|---|---:|---:|---:|
-| Mean PPL | **4.840423 ± 0.119931** | **4.838514 ± 0.119694** | **−0.001910 ± 0.005918** |
-| PPL ratio | 1.000000 | **0.999605 ± 0.001222** | **−0.0395% observed; no detectable degradation** |
-| Mean KLD vs BF16 | 0 | **0.004409 ± 0.000953** | lower is better |
-| Same top-p | 100% reference | **98.204 ± 0.147%** | 1.796 pp disagreement |
-| RMS token-probability Δp | 0% reference | **1.929 ± 0.201%** | distributional-noise scale |
-| Mean token-probability Δp | 0% reference | **−0.032 ± 0.021%** | small directional bias |
-| log-PPL correlation | 100% reference | **99.88%** | very high agreement |
+| `VeriLoop-E2-BF16.gguf` | BF16 / selected F32 tensors | **53,808,284,000 bytes / 50.113 GiB** | Canonical reference | Verified |
+| `VeriLoop-E2-Q8_0.gguf` | Q8_0 / retained F32 tensors | **28,595,765,600 bytes / 26.632 GiB** | High-fidelity local | Verified / Release Quality PASS |
+| `VeriLoop-E2-Q6_K.gguf` | Q6_K, imatrix-assisted | **22,082,532,096 bytes / 20.566 GiB** | **Overall sweet spot** | Verified / Release Quality PASS |
+| `VeriLoop-E2-Q5_K_M.gguf` | Mixed Q5/Q6, imatrix-assisted | **20,363,666,176 bytes / 18.965 GiB** | **Memory-quality sweet spot** | Verified / Release Quality PASS |
+| `VeriLoop-E2-Q4_K_M.gguf` | Mixed Q4/Q5/Q6 | **19,650,634,496 bytes / 18.301 GiB** | Intermediate lower-footprint candidate | Quantitative QC PASS; own runtime/identity pending |
+| `VeriLoop-E2-Q3_K_M.gguf` | **Custom mixed Q3/Q4/Q5/Q6** | **18,066,506,496 bytes / 16.826 GiB** | **Minimum-footprint validated tier** | **Full release PASS; runtime + MTP + remote identity verified** |
 
-**Q6_K quantitative gate: PASS.**
+### Q3 immutable identity
 
-### Q5_K_M — memory-constrained quality tier
-
-| Metric | BF16 reference | Q5_K_M | Observed drift |
-|---|---:|---:|---:|
-| Mean PPL | **4.840423 ± 0.119931** | **4.861965 ± 0.120630** | **+0.021542 ± 0.006609** |
-| PPL ratio | 1.000000 | **1.004450 ± 0.001361** | **+0.4450%** |
-| Mean KLD vs BF16 | 0 | **0.006919 ± 0.000945** | lower is better |
-| Same top-p | 100% reference | **97.251 ± 0.181%** | 2.749 pp disagreement |
-| RMS token-probability Δp | 0% reference | **2.490 ± 0.211%** | distributional-noise scale |
-| Mean token-probability Δp | 0% reference | **−0.127 ± 0.027%** | small negative directional bias |
-| log-PPL correlation | 100% reference | **99.85%** | very high agreement |
-
-**Q5_K_M quantitative gate: PASS.**
-
-### Q4_K_M — mixed-precision 4-bit-core candidate
-
-| Metric | BF16 reference | Q4_K_M | Observed drift |
-|---|---:|---:|---:|
-| Mean PPL | **4.840423 ± 0.119931** | **4.863760 ± 0.120720** | **+0.023337 ± 0.008350** |
-| PPL ratio | 1.000000 | **1.004821 ± 0.001722** | **+0.4821%** |
-| Mean KLD vs BF16 | 0 | **0.009700 ± 0.001110** | lower is better |
-| Same top-p | 100% reference | **96.786 ± 0.195%** | 3.214 pp disagreement |
-| RMS token-probability Δp | 0% reference | **2.843 ± 0.155%** | distributional-noise scale |
-| Mean token-probability Δp | 0% reference | **−0.046 ± 0.031%** | small directional bias |
-| log-PPL correlation | 100% reference | **99.76%** | very high agreement |
-
-**Q4_K_M quantitative gate: PASS.** The measured **+0.4821% PPL change is quantization-fidelity drift under this protocol, not a downstream benchmark-score loss**.
-
-### Quantitative release gates
+| Property | Value |
+|---|---|
+| Filename | `VeriLoop-E2-Q3_K_M.gguf` |
+| Exact bytes | **18,066,506,496** |
+| Binary size | **16.825745 GiB** |
+| SHA256 | `c2e9539cfb85d87a99605b6914ed602c0fa750164c153ce5574710123aba06fe` |
+| Effective density | **5.37 BPW** |
+| Tensor count | **851** |
+| llama.cpp QC/runtime revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
+| Hugging Face remote identity | **PASS — exact size + full SHA256 + remote readability** |
 
-The gates are **tier-specific frozen VeriLoop release-QA criteria**, not universal industry standards. Gate decisions use the measured point estimates; reported uncertainty remains visible alongside the measurements.
+### Optional MTP draft files
 
-| Tier | PPL-ratio gate | Mean-KLD gate | Same-top-p gate | Quantitative result |
-|---|---|---|---|---|
-| **Q8_0** | ≤ 1.005 → **1.000643 PASS** | ≤ 0.005 → **0.002176 PASS** | ≥ 97.0% → **98.815% PASS** | **PASS** |
-| **Q6_K** | ≤ 1.005 → **0.999605 PASS** | ≤ 0.006 → **0.004409 PASS** | ≥ 97.0% → **98.204% PASS** | **PASS** |
-| **Q5_K_M** | ≤ 1.0075 → **1.004450 PASS** | ≤ 0.008 → **0.006919 PASS** | ≥ 96.5% → **97.251% PASS** | **PASS** |
-| **Q4_K_M** | ≤ 1.0100 → **1.004821 PASS** | ≤ 0.0100 → **0.009700 PASS** | ≥ 96.0% → **96.786% PASS** | **PASS** |
+| File | Exact size | Role |
+|---|---:|---|
+| `mtp-VeriLoop-E2-Q8_0.gguf` | **3,287,070,240 bytes / 3.061 GiB** | Q8 MTP companion |
+| `mtp-VeriLoop-E2-Q6_K.gguf` | **2,008,056,096 bytes / 1.870 GiB** | 18-tensor Q4_0 MTP companion |
+| `mtp-VeriLoop-E2-Q5_K_M.gguf` | **2,008,056,096 bytes / 1.870 GiB** | 18-tensor Q4_0 MTP companion |
 
-For the full measurement record, see [`QUANTIZATION_QUALITY.md`](./QUANTIZATION_QUALITY.md).
+**Q3 MTP compatibility.** Q3 stock-runtime validation used an 18-tensor Q4_0 MTP object with SHA256 `0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671`. That object is byte-identical to the published Q5/Q6 MTP companion, so a duplicate Q3-specific MTP file is not required for the validated path.
 
-## Fixed BF16-Paired Comparison Protocol
+Artifact identity and validation states are tracked in [`RELEASE_MANIFEST.json`](./RELEASE_MANIFEST.json).
 
-Q8_0, Q6_K, Q5_K_M, and Q4_K_M use the **same BF16 reference logits and the same frozen evaluation protocol**, allowing within-repository trade-off analysis.
+## Frozen BF16-paired protocol
 
 | Item | Value |
 |---|---|
-| Baseline | Canonical VeriLoop E2 BF16 GGUF |
+| Reference model | Canonical VeriLoop E2 BF16 GGUF |
 | Corpus | WikiText-2 raw test |
 | Corpus SHA256 | `173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08` |
-| Context | **2,048 tokens** |
+| Context | **2,048** |
 | Chunks | **8** |
 | Seed | **42** |
 | GPU layers | **40** |
 | Flash Attention | **Off** |
-| KV cache | **F16 / F16** |
+| KV cache | **F16/F16** |
 | Batch / micro-batch | **512 / 512** |
-| Evaluation tool | `llama-perplexity` |
-| Reference-logit method | BF16 logits persisted with `--kl-divergence-base` |
-| Candidate method | Quantized model evaluated with `--kl-divergence` against the frozen BF16 logit file |
-| llama.cpp quantization / QC revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
-| BF16 reference SHA256 | `11bf5defde1a256b7582bc34fd2c4a85a61615ed88e7a422dcd24d814ea6d35d` |
-
-The BF16 logits are reused rather than recomputed for every quantization tier.
-
----
-
-## Quantization Configuration
-
-### Q8_0
-
-| Property | Value |
-|---|---|
-| Source | Canonical VeriLoop E2 BF16 GGUF |
-| Quantization | `Q8_0` |
-| Requantized from a lower-bit model | **No** |
-| Source size reported by quantizer | **51,305.09 MiB** |
-| Quantized size reported by quantizer | **27,260.56 MiB** |
-| Effective density | **8.50 BPW** |
-| Main tensor count | **851** |
-| Q8_0 tensors | **498** |
-| Retained F32 tensors | **353** |
-| Quantization threads | **25** |
-| Quantization time | **38.315 s** |
-| Build gate | **PASS** |
-
-### Q6_K
-
-| Property | Value |
-|---|---|
-| Source | Canonical VeriLoop E2 BF16 GGUF |
-| Quantization | `Q6_K` |
-| Requantized from another low-bit model | **No — direct BF16 → Q6_K** |
-| Quantized size reported by quantizer | **21,049.06 MiB** |
-| Effective density | **6.57 BPW** |
-| Exact main-file size | **22,082,532,096 bytes / 20.566 GiB** |
-| Main tensor count | **851** |
-| Quantization threads | **25** |
-| Quantization time | **67.13647 s** |
-| Importance matrix | **Model-specific; 496 entries; computed on 582 chunks** |
-| Importance-matrix size | **13,642,752 bytes** |
-| Importance-matrix SHA256 | `72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b` |
-| Calibration corpus | `Qwen3.8-27B-calibration-v6.txt` |
-| Main structure gate | **PASS** |
-| Quantitative quality gate | **PASS** |
-
-### Q5_K_M — mixed-precision quality-preserving build
-
-| Property | Value |
-|---|---|
-| Source | Canonical VeriLoop E2 BF16 GGUF |
-| Quantization | `Q5_K_M` base with protected Q6_K paths |
-| Requantized from another low-bit model | **No — direct BF16 → Q5_K_M** |
-| Quantized size reported by quantizer | **19,409.82 MiB** |
-| Effective density | **6.05 BPW** |
-| Exact main-file size | **20,363,666,176 bytes / 18.965 GiB** |
-| Main tensor count | **851** |
-| Retained F32 tensors | **353** |
-| Q5_K tensors | **272** |
-| Protected Q6_K tensors | **226** |
-| Quantization threads | **25** |
-| Quantization time | **120.23577 s** |
-| Importance matrix | **Model-specific; 496 entries; computed on 582 chunks** |
-| Importance-matrix size | **13,642,752 bytes** |
-| Importance-matrix SHA256 | `72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b` |
-| Calibration corpus | `Qwen3.8-27B-calibration-v6.txt` |
-| Quality policy | `q5_k_m_quality_v3_q6_attention_logits` |
-| Main structure gate | **PASS — 851 tensors** |
-| Quantitative quality gate | **PASS** |
-
-Protected Q6_K tensor groups:
-
-| Tensor group | Q6_K count |
-|---|---:|
-| `output.weight` | 1 |
-| `token_embd.weight` | 1 |
-| `*.attn_v.weight` | 16 |
-| `*.attn_qkv.weight` | 48 |
-| `*.attn_output.weight` | 16 |
-| `*.ffn_down.weight` | 64 |
-| `*.ssm_out.weight` | 48 |
-| `*.attn_q.weight` | 16 |
-| `*.attn_k.weight` | 16 |
-| **Total** | **226** |
-
-### Q4_K_M — root-cause-guided mixed-precision 4-bit-core build
-
-| Property | Value |
-|---|---|
-| Source | Canonical VeriLoop E2 BF16 GGUF |
-| Quantization | `Q4_K_M` base with Q5_K/Q6_K protected paths |
-| Requantized from another low-bit model | **No — direct BF16 → Q4_K_M** |
-| Quantized size reported by quantizer | **18,729.82 MiB** |
-| Effective density | **5.84 BPW** |
-| Exact main-file size | **19,650,634,496 bytes / 18.301 GiB** |
-| Main tensor count | **851** |
-| Retained F32 tensors | **353** |
-| Q4_K tensors | **64** |
-| Q5_K tensors | **208** |
-| Q6_K tensors | **226** |
-| Quantization threads | **25** |
-| Quantization time | **119.16670 s** |
-| Importance matrix | **Model-specific; 496 entries; computed on 582 chunks** |
-| Importance-matrix size | **13,642,752 bytes** |
-| Importance-matrix SHA256 | `72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b` |
-| Calibration corpus | `Qwen3.8-27B-calibration-v6.txt` |
-| Main structure / precision-policy gate | **PASS — 851 tensors** |
-| Quantitative quality gate | **PASS** |
-| Final stock runtime / immutable identity | **Pending** |
-
-The final Q4 policy is deliberately mixed precision. The **64 `*.ffn_up.weight` tensors are Q4_K**; **64 `*.ffn_gate.weight` tensors are retained at Q5_K**; attention-gate and SSM-control groups are also Q5_K; and **226 high-sensitivity output, embedding, attention, FFN-down, and SSM-output tensors remain Q6_K**. A broader development policy that also pushed FFN-gate tensors to Q4_K failed the frozen Mean-KLD gate (**0.012706 > 0.0100**). Restoring FFN-gate tensors to Q5_K reduced Mean KLD to **0.009700** while preserving a true Q4 core, which is the measured basis for the final candidate policy.
-
----
-
-## Release Provenance
-
-### BF16 reference
-
-| Check | Result |
-|---|---|
-| Parent checkpoint | `tsinghua-sigs-robot-lab/VeriLoop-E2` |
-| Base model family | Qwen3.8-27B |
-| HF architecture | `Qwen3_5ForConditionalGeneration` |
-| Native model context | 262,144 tokens |
-| Conversion tool | `convert_hf_to_gguf.py` |
-| BF16 conversion llama.cpp revision | `bddf8263c31c3dce3212263b00ebd2d98c1a752b` |
-| BF16 target tensor audit | **851 tensors / PASS** |
-| Non-zero tensor audit | **PASS** |
-| Source-checkpoint immutability gate | **PASS** |
-| BF16 main-model load gate | **PASS** |
-
-### Q8_0
-
-| Check | Result |
-|---|---|
-| Quantization source | Canonical BF16 GGUF |
-| Requantization from lower precision | **No** |
-| Quantization type | **Q8_0** |
-| llama.cpp quantization / QC revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
-| Structure gate | **PASS** |
-| Quantitative fidelity gate | **PASS** |
-| Main-model runtime generation | **PASS** |
-| MTP compatibility smoke | **PASS** |
-| Release identity | **Frozen** |
-
-### Q6_K
-
-| Check | Result |
-|---|---|
-| Quantization source | Canonical BF16 GGUF |
-| Requantization from lower precision | **No** |
-| Importance-matrix-assisted | **Yes** |
-| Structure gate | **PASS** |
-| Quantitative fidelity gate | **PASS** |
-| Stock llama.cpp main runtime | **PASS** |
-| MTP stock-runtime smoke | **PASS** |
-| Main SHA256 | `15d8f856471c4853f6bf0036b2a517426c6cb30a0313cef58a7f9577fd26fe9e` |
-| MTP SHA256 | `0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671` |
-| Release identity | **Frozen / remote exact-size verification PASS** |
-
-### Q5_K_M
-
-| Check | Result |
-|---|---|
-| Quantization source | Canonical BF16 GGUF |
-| Requantization from lower precision | **No** |
-| Importance-matrix-assisted | **Yes** |
-| Mixed-precision protected-path policy | **PASS** |
-| Main tensor count | **851 / PASS** |
-| Quantitative fidelity gate | **PASS** |
-| Final stock llama.cpp main runtime | **PASS** |
-| Final MTP speculative-runtime smoke | **PASS** |
-| Main SHA256 | `f90ec14d7ec8f084a292413ae0e06483c87ab5e1422ddce51f2e3409d8853b61` |
-| MTP SHA256 | `0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671` |
-| Remote exact-size + LFS-SHA verification | **PASS** |
-| Release identity | **Frozen / public release verified** |
-
-### Q4_K_M
-
-| Check | Result |
-|---|---|
-| Quantization source | Canonical BF16 GGUF |
-| Requantization from lower precision | **No** |
-| Importance-matrix-assisted | **Yes** |
-| Mixed Q4/Q5/Q6 precision policy | **PASS** |
-| Main tensor count | **851 / PASS** |
-| Quantitative fidelity gate | **PASS** |
-| Final stock llama.cpp runtime | **Pending** |
-| Final one-time SHA256 / immutable identity | **Pending** |
-| Public upload authorization | **Pending runtime + identity freeze** |
+| Evaluator | `llama-perplexity` |
+| Reference logits | `--kl-divergence-base` |
+| Candidate comparison | `--kl-divergence` |
+| BF16 logits reused across tiers | **Yes** |
+| llama.cpp revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
 
----
-
-## Parent-Model Benchmark Record
+The protocol supports **within-repository paired comparison**. PPL/KLD values from unrelated repositories are not treated as directly comparable unless the corpus, tokenizer path, chunking, context, cache, and runtime procedure are matched.
 
-The following scores identify the **parent VeriLoop E2 release**. They are not relabeled as Q8_0-, Q6_K-, Q5_K_M-, or Q4_K_M-specific task scores; quantization fidelity is measured directly against BF16 above.
+## Parent-model benchmark record
 
-<p align="center">
-  <img src="https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2/resolve/main/.eval_results/veriloop_e2_benchmark_result.png" width="100%" alt="VeriLoop E2 benchmark comparison across nine public benchmarks">
-</p>
+These scores identify the **parent VeriLoop E2 release**. They are included for model capability context and are **not relabeled as quant-specific results**.
 
-| Benchmark | VeriLoop E2 release score | Public evidence |
+| Benchmark | VeriLoop E2 parent score | Public evidence |
 |---|---:|---|
 | SWE-bench Pro | **76.2** | [Evidence](https://huggingface.co/datasets/tsinghua-sigs-robot-lab/VeriLoop-E2-Evaluation-Evidence/tree/main/swe-bench-pro) |
 | Terminal-Bench 2.1 | **88.8** | [Evidence](https://huggingface.co/datasets/tsinghua-sigs-robot-lab/VeriLoop-E2-Evaluation-Evidence/tree/main/terminal-bench-2.1) |
@@ -450,146 +218,95 @@ The following scores identify the **parent VeriLoop E2 release**. They are not r
 | GPQA Diamond | **93.9** | [Evidence](https://huggingface.co/datasets/tsinghua-sigs-robot-lab/VeriLoop-E2-Evaluation-Evidence/tree/main/gpqa-diamond) |
 | Apex 2025 | **89.6** | [Evidence](https://huggingface.co/datasets/tsinghua-sigs-robot-lab/VeriLoop-E2-Evaluation-Evidence/tree/main/apex-2025) |
 
-For exact benchmark protocols, task-level outputs, evaluator receipts, and provenance, use the **[VeriLoop E2 Evaluation Evidence](https://huggingface.co/datasets/tsinghua-sigs-robot-lab/VeriLoop-E2-Evaluation-Evidence)** repository.
-
----
-
-## Quick Start with llama.cpp
+For exact protocols and evaluator receipts, use the **[VeriLoop E2 Evaluation Evidence](https://huggingface.co/datasets/tsinghua-sigs-robot-lab/VeriLoop-E2-Evaluation-Evidence)** repository.
 
-Use a recent llama.cpp build with Qwen3.8 support.
+## Quick start with llama.cpp
 
-### Install
-
-```bash
-curl -LsSf https://llama.app/install.sh | sh
-```
-
-Or build from source:
-
-```bash
-git clone https://github.com/ggml-org/llama.cpp.git
-cd llama.cpp
-cmake -B build -DGGML_CUDA=ON
-cmake --build build -j --target llama-server llama-cli
-```
+Use a recent llama.cpp build with Qwen3.8 support. The frozen Q3 runtime validation used revision `42916d83f4a225e56709f873aa8050ac11f5b6a4`.
 
 ### Download
 
 ```bash
-# BF16 reference
-hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-BF16.gguf --local-dir .
-
-# Q8_0 high-fidelity local tier
-hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-Q8_0.gguf --local-dir .
-
-# Q6_K overall quality / efficiency sweet spot
+# Recommended default
 hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-Q6_K.gguf --local-dir .
 
-# Q5_K_M memory-constrained quality sweet spot
+# Memory-quality sweet spot
 hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-Q5_K_M.gguf --local-dir .
-```
 
-`Q4_K_M` is not added to the public download block until its final stock-runtime and immutable-identity gates are frozen.
+# Minimum-footprint validated tier
+hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-Q3_K_M.gguf --local-dir .
+
+# High-fidelity low-bit tier
+hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-Q8_0.gguf --local-dir .
+```
 
-### Run
+### Run Q3
 
 ```bash
 llama-server \
-  -m ./VeriLoop-E2-Q6_K.gguf \
-  -ngl all \
+  -m ./VeriLoop-E2-Q3_K_M.gguf \
+  -ngl 99 \
   -c 32768 \
   --host 127.0.0.1 \
   --port 8080
 ```
 
-The OpenAI-compatible endpoint is available at:
+The OpenAI-compatible endpoint is available at `http://127.0.0.1:8080/v1`.
 
-```text
-http://127.0.0.1:8080/v1
-```
-
-`-ngl all` is appropriate only when model weights, KV cache, and runtime workspace fit accelerator memory. Partial GPU offload is supported when full offload is not practical.
-
----
+`-ngl 99` requests full model-layer GPU offload. It is appropriate only when model weights, KV cache, compute buffers, and runtime workspace fit accelerator memory. Partial offload is supported.
 
-## Optional MTP Speculative Decoding
+### Q3 + MTP speculative decoding
 
-The Q8_0, Q6_K, and Q5_K_M MTP companions have completed stock llama.cpp compatibility validation.
+The validated Q3 path can reuse the existing byte-identical 18-tensor Q4_0 MTP companion:
 
 ```bash
-hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF mtp-VeriLoop-E2-Q8_0.gguf --local-dir .
-hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF mtp-VeriLoop-E2-Q6_K.gguf --local-dir .
 hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF mtp-VeriLoop-E2-Q5_K_M.gguf --local-dir .
-```
-
-Validated resource-conservative configuration:
 
-```bash
 llama-server \
-  -m ./VeriLoop-E2-Q6_K.gguf \
-  --model-draft ./mtp-VeriLoop-E2-Q6_K.gguf \
-  --spec-type draft-mtp \
-  -ngl all \
-  --n-gpu-layers-draft 0 \
-  -c 8192 \
+  -m ./VeriLoop-E2-Q3_K_M.gguf \
+  --model-draft ./mtp-VeriLoop-E2-Q5_K_M.gguf \
+  --gpu-layers-draft 99 \
+  --spec-draft-n-max 8 \
+  -ngl 99 \
+  -c 2048 \
   --host 127.0.0.1 \
   --port 8080
 ```
 
-MTP acceleration is workload- and runtime-dependent; no universal speedup is claimed.
+For the frozen Q3 runtime gate, **main-only and main+MTP greedy outputs were byte-identical**; the generated-text SHA256 was `43944ea63376bab89382150a3e4fbcb3739901c82e791abfa7847c13df0391a9` in both runs. MTP acceleration remains workload- and runtime-dependent; no universal speedup is claimed.
 
----
-
-## Memory Guidance
+## Memory guidance
 
-| Variant | Main model file | Practical implication |
+| Variant | Main file | Practical implication |
 |---|---:|---|
-| BF16 | **50.113 GiB** | Reference-quality deployment; substantial VRAM/RAM required |
-| Q8_0 | **26.632 GiB** | Near-reference fidelity at materially lower footprint |
-| **Q6_K** | **20.566 GiB** | **Overall quality / efficiency sweet spot** |
-| **Q5_K_M** | **18.965 GiB** | **Released memory-constrained quality sweet spot** |
-| **Q4_K_M** | **18.301 GiB** | Validated 4-bit-core candidate; only **0.664 GiB** smaller than Q5_K_M, with higher measured distributional drift |
-| Q8_0 MTP companion | **3.061 GiB** | Additional memory only when speculative decoding is enabled |
-| Q6/Q5 MTP companion | **1.870 GiB** | Additional draft-model footprint |
-| KV cache / runtime workspace | Workload-dependent | Increases with context length, batch size, cache dtype, and parallelism |
-
-Model-file size is not total runtime-memory consumption. Leave headroom for KV cache, compute buffers, runtime metadata, and optional draft-model execution.
-
-## Quantization Release Policy
-
-Every lower-bit VeriLoop E2 release is compared against the **same canonical BF16 reference**. A quant is not labeled “recommended” solely because the format is common.
-
-| Quant | Role | Current state | Evidence status |
-|---|---|---|---|
-| **BF16** | Reference fidelity | Released | Canonical baseline |
-| **Q8_0** | High-fidelity local | Released | **Full release quality PASS** |
-| **Q6_K** | **Overall quality / efficiency sweet spot** | **Released** | **Full release quality PASS** |
-| **Q5_K_M** | **Memory-constrained quality sweet spot** | **Released** | **Full release quality PASS; remote identity verified** |
-| **Q4_K_M** | Validated mixed-precision 4-bit-core tier | **Release candidate** | **Build + structure + quantitative fidelity PASS; runtime/identity pending** |
-| `IQ4_XS` | Compact 4-bit-class tier | Planned | Pending |
-
-The comparison protocol prioritizes exact artifact identity, direct-from-BF16 provenance, effective BPW, PPL ratio, KL divergence and distribution tails, token-probability drift, Same top-p agreement, structural integrity, stock-runtime compatibility, and immutable release identity.
+| BF16 | **50.113 GiB** | Reference deployment; substantial RAM/VRAM required |
+| Q8_0 | **26.632 GiB** | High fidelity; larger local footprint |
+| **Q6_K** | **20.566 GiB** | **Default quality / efficiency choice** |
+| **Q5_K_M** | **18.965 GiB** | **Best memory-quality balance below Q6** |
+| Q4_K_M | **18.301 GiB** | Intermediate footprint; its own final runtime/identity still pending in this record |
+| **Q3_K_M** | **16.826 GiB** | **Smallest validated main file; higher distributional drift** |
+| Q5/Q6-compatible MTP | **1.870 GiB** | Optional additional draft-model footprint |
 
-**Overall sweet spot:** **Q6_K**. It is the point on the currently measured ladder with the strongest balance of footprint reduction and BF16 retention. **Memory-constrained sweet spot:** **Q5_K_M**. Q4_K_M is retained as a validated lower-footprint candidate, not promoted to the default sweet spot merely because it is lower-bit.
+**A 16.826 GiB model file does not mean a 16 GiB GPU can fully offload it.** The Q3 weights alone exceed 16 GiB, and runtime memory also includes KV cache, compute buffers, allocator overhead, and optional MTP weights. Long-context memory can dominate well before the model's native 262K context limit.
 
-**No cross-repository PPL or KLD comparison is claimed unless the corpus, tokenizer path, chunking, context, and runtime protocol are identical.**
+## Throughput status
 
-## Throughput Status
+A standalone Q8_0 `llama-bench` record exists (**pp512 3102.722776 tok/s; tg128 41.999677 tok/s**), but there is no frozen paired BF16 throughput run. **No Q6/Q5/Q4/Q3 universal speedup percentage is claimed.** This deliberately separates quantization fidelity from hardware-specific throughput.
 
-A standalone Q8_0 `llama-bench` run produced:
+## Release policy and current positioning
 
-| Workload | Q8_0 |
-|---|---:|
-| Prompt processing (`pp512`) | **3102.722776 tok/s** |
-| Token generation (`tg128`) | **41.999677 tok/s** |
+| Tier | Positioning | Evidence status |
+|---|---|---|
+| BF16 | Canonical reference | Verified |
+| Q8_0 | High-fidelity local | Full release quality PASS |
+| **Q6_K** | **Overall sweet spot** | **Full release quality PASS** |
+| **Q5_K_M** | **Memory-constrained quality sweet spot** | **Full release quality PASS** |
+| Q4_K_M | Intermediate mixed-precision candidate | Quantitative QC PASS; own runtime/identity pending |
+| **Q3_K_M** | **Minimum-footprint validated tier** | **Full release PASS; stock runtime + MTP + remote identity verified** |
 
-No paired BF16→Q8 speedup is claimed because the paired BF16 speed run did not complete under the original full-offload configuration.
+The selection policy prioritizes exact artifact identity, direct-from-BF16 provenance, effective BPW, paired PPL/KLD/token drift, Same top-p, structural integrity, stock-runtime compatibility, and immutable release identity. **Lower bit-count alone is not treated as evidence of a better deployment point.**
 
-**No Q6_K, Q5_K_M, or Q4_K_M speedup claim is made without a frozen paired throughput protocol.** This deliberately separates fidelity retention from hardware-specific throughput.
-
----
-
-## Model Identity
+## Model identity
 
 | Property | Value |
 |---|---|
@@ -605,61 +322,39 @@ No paired BF16→Q8 speedup is claimed because the paired BF16 speed run did not
 | Languages | English, Chinese |
 | GGUF runtime | llama.cpp |
 | Reference precision | BF16 |
-| Released quantization | Q8_0, Q6_K, Q5_K_M |
-| Validated release candidate | Q4_K_M |
+| Validated low-bit ladder | Q8_0, Q6_K, Q5_K_M, Q3_K_M; Q4_K_M quantitative candidate |
 | License | Apache License 2.0 |
 
----
+## Measurement boundaries
 
-## Prompt and Chat Protocol
+- PPL/KLD/token metrics are **quantization-retention measurements**, not universal percentages of downstream capability loss.
+- Q6_K's negative PPL point estimate is not treated as a capability gain; its uncertainty spans BF16 parity.
+- Q3_K_M's **+0.4090% PPL** does not mean a 0.4090% drop on SWE-bench, AIME, GPQA, or any other task benchmark.
+- Parent-model leaderboard scores are not represented as quant-specific re-runs.
+- The Q3 filename does not imply uniform 3-bit storage; the exact mixed-precision tensor policy and 5.37 BPW effective density are disclosed above.
+- Native 262K context comes from the parent configuration; practical context depends on memory and runtime settings.
+- MTP speedup is workload-dependent.
+- The model can still produce incorrect code, mathematics, scientific reasoning, or commands.
 
-Use the chat template embedded in the GGUF / inherited from VeriLoop E2. Prompt-template changes, reasoning delimiters, stop rules, samplers, cache settings, and tool-call parsers can materially change observed behavior even when model weights are identical.
-
-For the full model protocol and system-level behavior, see the **[VeriLoop E2 parent model card](https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2)**.
-
----
-
-## Measurement Boundaries
-
-- **Q8_0 +0.0643%** is the measured relative PPL increase under the fixed paired protocol; it is not a universal capability-loss percentage.
-- **Q6_K −0.0395% observed mean PPL change is not claimed as a capability gain**; the reported uncertainty spans parity.
-- **Q5_K_M +0.4450%** is a protocol-specific PPL increase, not a claim that downstream task scores fall by 0.4450%.
-- **Q4_K_M +0.4821%** is likewise a protocol-specific PPL increase; it is not a measured percentage loss on SWE-bench, AIME, GPQA, or other downstream tasks.
-- Parent-model leaderboard results are included for model identity and provenance, not represented as quant-specific task re-runs.
-- PPL/KLD values are meaningful for within-repository BF16-paired comparisons because the reference and candidates use the same frozen protocol.
-- The native 262K context comes from the parent-model configuration. Practical local context depends on available memory and runtime settings.
-- MTP acceleration is runtime- and workload-dependent.
-- The model can still produce incorrect code, mathematical reasoning, scientific arguments, or commands.
-
----
-
-## Source Model and Evidence
+## Source model and evidence
 
 | Resource | Link |
 |---|---|
 | Parent model | [VeriLoop E2](https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2) |
 | GGUF repository | [VeriLoop E2 GGUF](https://huggingface.co/tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF) |
 | Quantization quality record | [Quantization Quality](./QUANTIZATION_QUALITY.md) |
-| Unified artifact identity / release state | [Release Manifest](./RELEASE_MANIFEST.json) |
+| Artifact identity / release state | [Release Manifest](./RELEASE_MANIFEST.json) |
 | Technical report | [OpenReview](https://openreview.net/forum?id=P6FIQILHwX&noteId=P6FIQILHwX) |
 | Evaluation evidence | [VeriLoop E2 Evaluation Evidence](https://huggingface.co/datasets/tsinghua-sigs-robot-lab/VeriLoop-E2-Evaluation-Evidence) |
 | Riemann ζ artifact | [Public artifact](https://github.com/brucewang123456789/GeniusTrail/tree/VeriLoop-E2/riemann-hypothesis) |
 | llama.cpp | [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp) |
 
----
-
 ## License
 
-The VeriLoop E2 model weights and this GGUF distribution are released under the **Apache License 2.0**, consistent with the parent model release.
-
-Third-party software, benchmarks, and separately published scientific artifacts retain their own terms.
-
----
+The VeriLoop E2 model weights and this GGUF distribution are released under the **Apache License 2.0**, consistent with the parent model release. Third-party software, benchmarks, and separately published scientific artifacts retain their own terms.
 
 ## Citation
 
-If you use VeriLoop E2 or this GGUF distribution in research, please cite the parent model release:
-
 ```bibtex
 @misc{wang2026veriloope2,
   title        = {VeriLoop E2: A 27B Post-Trained Model for Code, Mathematics, and Scientific Reasoning},
@@ -671,12 +366,10 @@ If you use VeriLoop E2 or this GGUF distribution in research, please cite the pa
 }
 ```
 
-For quantization-specific comparisons, identify the exact GGUF filename and release manifest.
-
----
+For quantization-specific comparisons, identify the exact GGUF filename and the corresponding release-manifest identity.
 
 ## Acknowledgements
 
 VeriLoop E2 builds on **Qwen3.8-27B** and the open-source model-serving and evaluation ecosystem. We thank the communities behind Qwen, llama.cpp, Transformers, Hugging Face, software-engineering benchmarks, mathematical evaluation suites, and reproducible scientific computing.
 
-This repository treats **reference fidelity, quantization quality, immutable artifact identity, and reproducible runtime validation as first-class release properties**.
+This repository treats **reference fidelity, transparent mixed-precision policy, immutable artifact identity, and reproducible stock-runtime validation as first-class release properties**.

RELEASE_MANIFEST.json

@@ -1,5 +1,5 @@
 {
-  "schema": "veriloop.e2.gguf.release_manifest.v5",
+  "schema": "veriloop.e2.gguf.release_manifest.v6",
   "repository": "tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF",
   "model": {
     "name": "VeriLoop E2",
@@ -17,13 +17,16 @@
     "quality_method": "paired_BF16_vs_quant_fixed_protocol",
     "parent_benchmark_rerun_per_quant": false,
     "performance_claim_boundary": "PPL_KLD_token_drift_are_quantization_fidelity_metrics_not_universal_task_score_loss",
-    "immutable_identity_required_before_public_release": true
+    "immutable_identity_required_before_public_release": true,
+    "q3_filename_semantics": "custom_mixed_precision_release_tier_minimum_active_precision_Q3_K_not_uniform_3bit",
+    "lower_bit_alone_is_not_a_recommendation_criterion": true
   },
   "sweet_spots": {
     "BF16": "reference_fidelity_sweet_spot",
     "Q8_0": "high_fidelity_local_sweet_spot",
     "Q6_K": "overall_quality_efficiency_sweet_spot",
-    "Q5_K_M": "memory_constrained_quality_sweet_spot"
+    "Q5_K_M": "memory_constrained_quality_sweet_spot",
+    "Q3_K_M": "minimum_footprint_validated_tier_not_default_sweet_spot"
   },
   "toolchains": {
     "bf16_conversion": {
@@ -139,6 +142,114 @@
         "final_quantitative_gate": "PASS"
       },
       "validation_run": "q4_k_m_v3_release_candidate_20260925_005537"
+    },
+    "q3_quantization_and_qc": {
+      "runtime": "llama.cpp",
+      "llama_cpp_revision": "42916d83f4a225e56709f873aa8050ac11f5b6a4",
+      "quantization_source": "VeriLoop-E2-BF16.gguf",
+      "quantization": "custom_mixed_Q3_Q4_Q5_Q6",
+      "filename": "VeriLoop-E2-Q3_K_M.gguf",
+      "filename_semantics": "release_tier_and_minimum_active_precision_not_uniform_stock_Q3_K_M",
+      "direct_from_bf16": true,
+      "low_bit_requantization": false,
+      "importance_matrix_assisted": true,
+      "importance_matrix_sha256": "72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b",
+      "quality_policy_id": "q3_kl_guard_lowest_sensitivity_ffn_down_v1",
+      "quantization_threads": 25,
+      "quantization_time_seconds": 166.77453,
+      "quantized_size_mib_reported_by_quantizer": 17219.08,
+      "effective_bpw": 5.37,
+      "tensor_type_counts": {
+        "F32": 353,
+        "Q3_K": 3,
+        "Q4_K": 64,
+        "Q5_K": 429,
+        "Q6_K": 2
+      },
+      "precision_policy": {
+        "Q3_K": {
+          "exact_tensors": [
+            "blk.1.ffn_down.weight",
+            "blk.0.ffn_down.weight",
+            "blk.3.ffn_down.weight"
+          ],
+          "total": 3
+        },
+        "Q4_K": {
+          "*.ffn_up.weight": 64,
+          "total": 64
+        },
+        "Q5_K": {
+          "remaining_*.ffn_down.weight": 61,
+          "*.ffn_gate.weight": 64,
+          "*.attn_qkv.weight": 48,
+          "*.attn_q.weight": 16,
+          "*.attn_k.weight": 16,
+          "*.attn_v.weight": 16,
+          "*.attn_output.weight": 16,
+          "*.attn_gate.weight": 48,
+          "*.ssm_alpha.weight": 48,
+          "*.ssm_beta.weight": 48,
+          "*.ssm_out.weight": 48,
+          "total": 429
+        },
+        "Q6_K": {
+          "output.weight": 1,
+          "token_embd.weight": 1,
+          "total": 2
+        }
+      },
+      "root_cause_development": {
+        "broad_q3_v1": {
+          "ppl_ratio": 1.017999,
+          "mean_kld": 0.029363,
+          "same_top_p_percent": 93.438,
+          "result": "FAIL"
+        },
+        "ffn_up_q3_ablation": {
+          "ppl_ratio": 1.015596,
+          "mean_kld": 0.02606,
+          "same_top_p_percent": 93.976,
+          "result": "FAIL",
+          "interpretation": "DOMINANT_SINGLE_VARIABLE_ROOT_CAUSE"
+        },
+        "internal_q6_to_q5_baseline": {
+          "ppl_ratio": 1.007813,
+          "mean_kld": 0.013292,
+          "same_top_p_percent": 96.065,
+          "result": "PASS_Q3_HARD_GATE_NO_Q3_TENSORS"
+        },
+        "ffn_down_full_family_q3": {
+          "ppl_ratio": 1.012324,
+          "mean_kld": 0.032745,
+          "same_top_p_percent": 93.341,
+          "result": "FAIL"
+        },
+        "qkv_full_family_q3": {
+          "ppl_ratio": 1.0187,
+          "mean_kld": 0.026723,
+          "same_top_p_percent": 94.086,
+          "result": "FAIL"
+        },
+        "imatrix_33_tensor_shortlist": {
+          "ppl_ratio": 1.002662,
+          "mean_kld": 0.018066,
+          "same_top_p_percent": 95.528,
+          "result": "FAIL_KLD"
+        },
+        "final_kl_guard": {
+          "q3_tensor_count": 3,
+          "ppl_ratio": 1.00409,
+          "mean_kld": 0.014349,
+          "same_top_p_percent": 95.919,
+          "hard_gate": "PASS",
+          "engineering_reserve_gate": "PASS"
+        }
+      },
+      "importance_matrix_entries": 496,
+      "importance_matrix_chunks": 582,
+      "importance_matrix_bytes": 13642752,
+      "validation_run": "q3_kl_guard_release_candidate_20260925_024427"
     }
   },
   "artifacts": {
@@ -300,6 +411,43 @@
       "runtime_gate": "PENDING",
       "status": "RELEASE_CANDIDATE",
       "validation_run": "q4_k_m_v3_release_candidate_20260925_005537"
+    },
+    "VeriLoop-E2-Q3_K_M.gguf": {
+      "tier": "Q3_K_M",
+      "role": "minimum_footprint_validated_main",
+      "quantization": "custom_mixed_Q3_Q4_Q5_Q6",
+      "filename_semantics": "not_uniform_stock_Q3_K_M",
+      "source": "VeriLoop-E2-BF16.gguf",
+      "direct_from_bf16": true,
+      "low_bit_requantization": false,
+      "bytes": 18066506496,
+      "decimal_gb": 18.066506496,
+      "binary_gib": 16.825745344161987,
+      "sha256": "c2e9539cfb85d87a99605b6914ed602c0fa750164c153ce5574710123aba06fe",
+      "sha256_state": "FROZEN",
+      "tensor_count": 851,
+      "f32_tensor_count": 353,
+      "q3_k_tensor_count": 3,
+      "q4_k_tensor_count": 64,
+      "q5_k_tensor_count": 429,
+      "q6_k_tensor_count": 2,
+      "effective_bpw": 5.37,
+      "quantized_size_mib_reported_by_quantizer": 17219.08,
+      "quantization_time_seconds": 166.77453,
+      "importance_matrix_sha256": "72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b",
+      "quality_policy_id": "q3_kl_guard_lowest_sensitivity_ffn_down_v1",
+      "structure_gate": "PASS",
+      "precision_policy_gate": "PASS",
+      "quantitative_quality_gate": "PASS",
+      "engineering_reserve_gate": "PASS",
+      "runtime_gate": "PASS",
+      "mtp_runtime_gate": "PASS",
+      "speculative_output_identity_gate": "PASS",
+      "remote_release_gate": "PASS",
+      "status": "VERIFIED",
+      "remote_exact_size_gate": "PASS",
+      "remote_lfs_sha256_gate": "PASS",
+      "remote_readability_gate": "PASS"
     }
   },
   "artifact_scope": {
@@ -310,7 +458,8 @@
       "VeriLoop-E2-Q6_K.gguf",
       "mtp-VeriLoop-E2-Q6_K.gguf",
       "VeriLoop-E2-Q5_K_M.gguf",
-      "mtp-VeriLoop-E2-Q5_K_M.gguf"
+      "mtp-VeriLoop-E2-Q5_K_M.gguf",
+      "VeriLoop-E2-Q3_K_M.gguf"
     ],
     "bf16_mtp_companion": {
       "filename": "mtp-VeriLoop-E2-BF16.gguf",
@@ -321,7 +470,14 @@
     "release_candidate_set": [
       "VeriLoop-E2-Q4_K_M.gguf"
     ],
-    "q4_release_candidate_note": "Q4_K_M has completed direct-from-BF16 build, exact mixed-precision structural audit, and BF16-paired quantitative QC. Public release remains pending final stock-runtime validation and one-time SHA256 identity freeze."
+    "q4_release_candidate_note": "Q4_K_M has completed direct-from-BF16 build, exact mixed-precision structural audit, and BF16-paired quantitative QC. Public release remains pending final stock-runtime validation and one-time SHA256 identity freeze.",
+    "q3_release_note": "Q3_K_M is a root-cause-guided mixed-precision artifact with 3 Q3_K tensors and 5.37 effective BPW. Structure, BF16-paired quality, engineering reserve, stock llama.cpp runtime, MTP runtime, immutable SHA256 identity, and Hugging Face remote release verification are PASS.",
+    "q3_mtp_compatibility": {
+      "dedicated_duplicate_file_required": false,
+      "validated_mtp_bytes": 2008056096,
+      "validated_mtp_sha256": "0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671",
+      "note": "The validated Q3 MTP object is byte-identical to the published 18-tensor Q4_0 Q5/Q6 companion."
+    }
   },
   "quality_protocol": {
     "name": "paired_BF16_vs_quant",
@@ -464,6 +620,39 @@
       "size_reduction_vs_q6_percent": 11.01276606064805,
       "size_reduction_vs_q5_percent": 3.501489730961893,
       "quantitative_quality_gate": "PASS"
+    },
+    "Q3_K_M": {
+      "mean_ppl": 4.860222,
+      "ppl_uncertainty": 0.120606,
+      "absolute_ppl_delta": 0.019799,
+      "absolute_ppl_delta_uncertainty": 0.010554,
+      "ppl_ratio": 1.00409,
+      "ppl_ratio_uncertainty": 0.002181,
+      "relative_ppl_change_percent_observed": 0.409,
+      "ln_ppl_ratio": 0.004082,
+      "ln_ppl_ratio_uncertainty": 0.002172,
+      "log_ppl_correlation_percent": 99.62,
+      "mean_kld": 0.014349,
+      "mean_kld_uncertainty": 0.001985,
+      "median_kld": 0.003572,
+      "kld_p90": 0.022146,
+      "kld_p95": 0.03835,
+      "kld_p99": 0.123439,
+      "kld_p999": 0.664245,
+      "max_kld": 13.526417,
+      "mean_delta_p_percent": -0.065,
+      "mean_delta_p_uncertainty_percent": 0.039,
+      "rms_delta_p_percent": 3.515,
+      "rms_delta_p_uncertainty_percent": 0.228,
+      "same_top_p_percent": 95.919,
+      "same_top_p_uncertainty_percent": 0.219,
+      "size_reduction_vs_bf16_percent": 66.42430281552929,
+      "size_reduction_vs_q8_percent": 36.821042846987105,
+      "size_reduction_vs_q6_percent": 18.186436150261308,
+      "size_reduction_vs_q5_percent": 11.280678342229766,
+      "size_reduction_vs_q4_percent": 8.06145979826991,
+      "quantitative_quality_gate": "PASS",
+      "engineering_reserve_gate": "PASS"
     }
   },
   "release_gates": {
@@ -568,6 +757,60 @@
       "main_runtime_generation": "PENDING",
       "final_release_identity": "PENDING",
       "overall_release_quality_gate": "PENDING_RUNTIME_AND_IDENTITY"
+    },
+    "Q3_K_M": {
+      "bf16_baseline_identity": "PASS",
+      "ppl_ratio": {
+        "hard_threshold_max": 1.015,
+        "reserve_threshold_max": 1.0135,
+        "measured": 1.00409,
+        "hard_result": "PASS",
+        "reserve_result": "PASS"
+      },
+      "mean_kld": {
+        "hard_threshold_max": 0.015,
+        "reserve_threshold_max": 0.0145,
+        "measured": 0.014349,
+        "hard_result": "PASS",
+        "reserve_result": "PASS"
+      },
+      "same_top_p_percent": {
+        "hard_threshold_min": 95.0,
+        "reserve_threshold_min": 95.5,
+        "measured": 95.919,
+        "hard_result": "PASS",
+        "reserve_result": "PASS"
+      },
+      "main_tensor_count": {
+        "expected": 851,
+        "measured": 851,
+        "result": "PASS"
+      },
+      "mixed_precision_policy": {
+        "expected_counts": {
+          "F32": 353,
+          "Q3_K": 3,
+          "Q4_K": 64,
+          "Q5_K": 429,
+          "Q6_K": 2
+        },
+        "mismatch_count": 0,
+        "result": "PASS"
+      },
+      "quantitative_quality_gate": "PASS",
+      "engineering_reserve_gate": "PASS",
+      "main_runtime_generation": "PASS",
+      "mtp_runtime_load": "PASS",
+      "mtp_runtime_generation": "PASS",
+      "mtp_log_evidence": "PASS",
+      "speculative_output_identity": "PASS",
+      "final_release_identity": "FROZEN",
+      "remote_release_verification": "PASS",
+      "overall_local_release_quality_gate": "PASS",
+      "remote_exact_size_verification": "PASS",
+      "remote_full_sha256_verification": "PASS",
+      "remote_readability_verification": "PASS",
+      "overall_release_quality_gate": "PASS"
     }
   },
   "runtime_validation": {
@@ -612,6 +855,24 @@
       "main_model_generation": "PENDING",
       "final_release_identity": "PENDING",
       "note": "Direct-from-BF16 build, exact mixed-precision structure/policy audit, and paired quantitative fidelity have passed. Final stock llama.cpp runtime validation remains pending."
+    },
+    "Q3_K_M": {
+      "stock_llama_cpp_only": true,
+      "veriloop_harness_used": false,
+      "llama_cpp_revision": "42916d83f4a225e56709f873aa8050ac11f5b6a4",
+      "validated_context_tokens": 2048,
+      "main_model_load": "PASS",
+      "main_model_generation": "PASS",
+      "mtp_interface": "PASS",
+      "mtp_load": "PASS",
+      "mtp_generation": "PASS",
+      "mtp_log_evidence": "PASS",
+      "speculative_output_identity": "PASS",
+      "main_only_text_sha256": "43944ea63376bab89382150a3e4fbcb3739901c82e791abfa7847c13df0391a9",
+      "main_plus_mtp_text_sha256": "43944ea63376bab89382150a3e4fbcb3739901c82e791abfa7847c13df0391a9",
+      "mtp_bytes": 2008056096,
+      "mtp_sha256": "0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671",
+      "runtime_mtp_gate": "PASS"
     }
   },
   "remote_repository_verification": {
@@ -632,6 +893,18 @@
       "mtp_remote_lfs_sha256": "0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671",
       "remote_release_gate": "PASS",
       "public_release": "PASS"
+    },
+    "Q3_K_M": {
+      "state": "PASS",
+      "remote_path": "VeriLoop-E2-Q3_K_M.gguf",
+      "main_remote_bytes": 18066506496,
+      "main_remote_lfs_sha256": "c2e9539cfb85d87a99605b6914ed602c0fa750164c153ce5574710123aba06fe",
+      "remote_exact_size_gate": "PASS",
+      "remote_full_sha256_gate": "PASS",
+      "remote_readability_gate": "PASS",
+      "remote_head_identity_gate": "PASS",
+      "remote_tail_identity_gate": "PASS",
+      "remote_release_gate": "PASS"
     }
   },
   "throughput": {
@@ -655,6 +928,11 @@
       "paired_speed_protocol_run": false,
       "speedup_claimed": false,
       "note": "No Q4 throughput speedup claim is made without a frozen paired speed protocol."
+    },
+    "Q3_K_M": {
+      "paired_speed_protocol_run": false,
+      "speedup_claimed": false,
+      "note": "Stock runtime and MTP correctness are validated. No universal speedup claim is made without a frozen paired throughput protocol."
     }
   },
   "release_matrix": {
@@ -694,6 +972,17 @@
       "state": "planned",
       "role": "compact_4bit_class",
       "quality_record": "pending"
+    },
+    "Q3_K_M": {
+      "state": "verified_release",
+      "role": "minimum_footprint_validated_tier",
+      "quality_record": "PASS",
+      "engineering_reserve_record": "PASS",
+      "runtime_record": "PASS",
+      "mtp_record": "PASS",
+      "release_identity": "FROZEN",
+      "remote_release": "PASS",
+      "overall_release": "PASS"
     }
   },
   "current_release_state": {
@@ -701,12 +990,15 @@
     "Q8_0": "RELEASED_VERIFIED",
     "Q6_K": "RELEASED_VERIFIED_OVERALL_SWEET_SPOT",
     "Q5_K_M": "RELEASED_VERIFIED_MEMORY_CONSTRAINED_SWEET_SPOT",
-    "Q4_K_M": "QUANTITATIVE_QC_PASS_RUNTIME_AND_SHA_PENDING"
+    "Q4_K_M": "QUANTITATIVE_QC_PASS_RUNTIME_AND_SHA_PENDING",
+    "Q3_K_M": "VERIFIED_RELEASE"
   },
   "deployment_recommendation": {
     "overall_sweet_spot": "Q6_K",
     "memory_constrained_sweet_spot": "Q5_K_M",
-    "q4_positioning": "validated_mixed_precision_4bit_core_release_candidate",
-    "basis": "Q6_K preserves BF16 PPL parity within reported uncertainty at 58.96% smaller main-file footprint; Q5_K_M saves a further 7.78% with controlled measured drift; Q4_K_M saves only 3.50% beyond Q5_K_M while Mean KLD rises to 0.009700 and Same top-p falls to 96.786%."
+    "minimum_footprint_validated_tier": "Q3_K_M",
+    "q4_positioning": "validated_intermediate_mixed_precision_candidate",
+    "q3_positioning": "footprint_first_not_default_sweet_spot",
+    "basis": "Q6_K preserves BF16 PPL parity within reported uncertainty at 58.96% smaller main-file footprint. Q5_K_M saves a further 7.78% with controlled measured drift. Q3_K_M is 66.42% smaller than BF16 and 11.28% smaller than Q5_K_M, but Mean KLD rises to 0.014349 and Same top-p falls to 95.919%, so Q3 is the minimum-footprint validated tier rather than the balance optimum."
   }
 }