Upload 3 files

QUANTIZATION_QUALITY.md

@@ -7,36 +7,34 @@
 **Runtime:** llama.cpp  
 **Reference precision:** BF16  
 **Measured ladder:** Q8_0, Q6_K, Q5_K_M, Q4_K_M, Q3_K_M, IQ2_S, IQ1_M  
-**Parent-model benchmark re-run per quant:** No
+**Parent-model benchmark rerun per quantization tier:** No
 
-> This document separates **model capability** from **quantization fidelity**. Parent-model benchmark scores establish the parent VeriLoop E2 capability. The measurements below quantify how closely each GGUF tier preserves the canonical BF16 reference under one frozen paired protocol.
+> This document separates **parent-model capability** from **quantization fidelity**. Parent-model benchmark scores establish the VeriLoop E2 capability record; the measurements below quantify how closely each GGUF tier preserves the canonical BF16 reference under one frozen paired protocol.
 
 ## 1. Executive decision
 
-**Overall sweet spot — Q6_K.** **58.96%** smaller than BF16 with PPL parity within reported uncertainty, Mean KLD **0.004409 ± 0.000953**, Same top-p **98.204 ± 0.147%**.
+**Overall sweet spot — Q6_K.** **58.96%** smaller than BF16 with PPL parity within reported uncertainty, Mean KLD **0.004409 ± 0.000953**, and Same top-p **98.204 ± 0.147%**.
 
-**Memory-quality sweet spot — Q5_K_M.** **62.16%** smaller than BF16 and **7.78%** smaller than Q6_K, with **+0.4450%** PPL drift, Mean KLD **0.006919 ± 0.000945**, Same top-p **97.251 ± 0.181%**.
+**Memory-quality sweet spot — Q5_K_M.** **62.16%** smaller than BF16 and **7.78%** smaller than Q6_K, with **+0.4450%** PPL drift, Mean KLD **0.006919 ± 0.000945**, and Same top-p **97.251 ± 0.181%**.
 
-**Current released low-footprint sweet spot — IQ2_S.** **16.798508 GiB**, full release verification PASS, PPL ratio **1.003457 ± 0.002100**, Mean KLD **0.014023 ± 0.001408**, Same top-p **95.516 ± 0.229%**.
+**Minimum-footprint sweet spot — IQ1_M.** **16.790078 GiB**, **66.4955% smaller than BF16**, PPL ratio **1.003191 ± 0.002175**, Mean KLD **0.014357 ± 0.001317**, and Same top-p **95.870 ± 0.220%**. Structure, hard fidelity, engineering reserve, stock llama.cpp runtime, and real MTP engagement all pass.
 
-**Ultra-low-footprint frontier candidate — IQ1_M.** The measured V2 artifact is **16.790078 GiB**, **66.4955% smaller than BF16**, and passes the unchanged hard and engineering-reserve fidelity gates with PPL ratio **1.003191 ± 0.002175**, Mean KLD **0.014357 ± 0.001317**, Same top-p **95.870 ± 0.220%**. It is smaller than IQ2_S and has better PPL/Same-top point estimates, but slightly worse KLD/RMS. The accepted rebuild reproduced the same policy and froze the local SHA256. Stock runtime, MTP, and remote verification still remain before IQ1_M is promoted to a released sweet spot.
-
-**Objective low-footprint conclusion.** IQ1_M and IQ2_S are neighboring Pareto points rather than a clean first/second ranking. IQ2_S currently wins on release maturity and Mean KLD; IQ1_M wins on measured footprint, PPL ratio, and Same top-p. Their uncertainty intervals substantially overlap.
+**Lower-KLD low-footprint alternative — IQ2_S.** **16.798508 GiB**, Mean KLD **0.014023 ± 0.001408**, and Same top-p **95.516 ± 0.229%**.
 
 ## 2. Precision matrix
 
-| 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%** | PASS |
-| **Q6_K** | **Overall sweet spot** | **20.565961 GiB** | **6.57 BPW** | **4.838514 ± 0.119694** | **0.004409 ± 0.000953** | **98.204 ± 0.147%** | 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%** | PASS |
-| Q4_K_M | Intermediate candidate | **18.301080 GiB** | **5.84 BPW** | **4.863760 ± 0.120720** | **0.009700 ± 0.001110** | **96.786 ± 0.195%** | Quantitative PASS |
-| Q3_K_M | Adjacent low-footprint release | **16.825745 GiB** | **5.37 BPW** | **4.860222 ± 0.120606** | **0.014349 ± 0.001985** | **95.919 ± 0.219%** | Full release PASS |
-| **IQ2_S** | **Released low-footprint sweet spot** | **16.798508 GiB** | **5.36 BPW** | **4.857159 ± 0.120482** | **0.014023 ± 0.001408** | **95.516 ± 0.229%** | **Full release PASS** |
-| **IQ1_M** | **Ultra-low-footprint frontier** | **16.790078 GiB** | **5.36 BPW** | **4.855870 ± 0.120480** | **0.014357 ± 0.001317** | **95.870 ± 0.220%** | **Hard + reserve quantitative PASS; final release gates pending** |
+| Tier | Role | Main size | Effective density | Mean PPL | Mean KLD | Same top-p |
+|---|---|---:|---:|---:|---:|---:|
+| BF16 | Reference | **50.112870 GiB** | 16-bit-class | **4.840423 ± 0.119931** | 0 | 100% |
+| Q8_0 | High fidelity | **26.631882 GiB** | **8.50 BPW** | **4.843536 ± 0.120062** | **0.002176 ± 0.000668** | **98.815 ± 0.120%** |
+| **Q6_K** | **Overall sweet spot** | **20.565961 GiB** | **6.57 BPW** | **4.838514 ± 0.119694** | **0.004409 ± 0.000953** | **98.204 ± 0.147%** |
+| **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%** |
+| Q4_K_M | Balanced compact | **18.301080 GiB** | **5.84 BPW** | **4.863760 ± 0.120720** | **0.009700 ± 0.001110** | **96.786 ± 0.195%** |
+| Q3_K_M | Low-footprint alternative | **16.825745 GiB** | **5.37 BPW** | **4.860222 ± 0.120606** | **0.014349 ± 0.001985** | **95.919 ± 0.219%** |
+| IQ2_S | Lower-KLD low-footprint alternative | **16.798508 GiB** | **5.36 BPW** | **4.857159 ± 0.120482** | **0.014023 ± 0.001408** | **95.516 ± 0.229%** |
+| **IQ1_M** | **Minimum-footprint sweet spot** | **16.790078 GiB** | **5.36 BPW** | **4.855870 ± 0.120480** | **0.014357 ± 0.001317** | **95.870 ± 0.220%** |
 
-## 3. Performance-loss benchmark
+## 3. Quantization-retention benchmark
 
 | Tier | PPL ratio | Relative PPL change | Mean KLD | Same top-p | Top-token disagreement | RMS Δp | log-PPL corr. |
 |---|---:|---:|---:|---:|---:|---:|---:|
@@ -48,7 +46,7 @@
 | IQ2_S | **1.003457 ± 0.002100** | **+0.3457%** | **0.014023 ± 0.001408** | **95.516 ± 0.229%** | **4.484 pp** | **3.679 ± 0.231%** | **99.64%** |
 | **IQ1_M** | **1.003191 ± 0.002175** | **+0.3191%** | **0.014357 ± 0.001317** | **95.870 ± 0.220%** | **4.130 pp** | **3.691 ± 0.205%** | **99.62%** |
 
-**Claim boundary.** These are **quantization-retention measurements**, not downstream task-score loss. The nine parent-model benchmarks were not rerun independently per quant. Therefore **+0.3191% PPL does not mean a 0.3191% loss on SWE-bench, Terminal-Bench, DeepSWE, AIME, GPQA, Apex, coding-agent, mathematics, or physics performance.**
+**Claim boundary.** These are quantization-retention measurements, not downstream task-score loss. The nine parent-model benchmarks were not independently rerun for every quantization tier.
 
 ## 4. Frozen BF16-paired protocol
 
@@ -65,7 +63,7 @@
 | Batch / micro-batch | 512 / 512 |
 | Evaluator | `llama-perplexity` |
 | Reference method | `--kl-divergence-base` |
-| Candidate method | `--kl-divergence` |
+| Quantized comparison method | `--kl-divergence` |
 | BF16 logits reused | Yes |
 | llama.cpp revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
 
@@ -81,7 +79,7 @@
 | Mean Δp | 0 | **−0.042 ± 0.041%** | small directional bias |
 | log-PPL correlation | 100% | **99.62%** | high agreement |
 
-### IQ1_M KLD distribution
+### KLD distribution
 
 | Statistic | Value |
 |---|---:|
@@ -95,11 +93,11 @@
 
 ## 6. IQ1_M exact tensor policy
 
-The measured Q1 V2 policy differs from the successful IQ2_S release by exactly **one tensor**:
+IQ1_M differs from IQ2_S by exactly one tensor:
 
 - `blk.1.ffn_down.weight`: **IQ2_S → IQ1_M**
-- `blk.0.ffn_down.weight`: stays **IQ2_S**
-- `blk.3.ffn_down.weight`: stays **IQ2_S**
+- `blk.0.ffn_down.weight`: **IQ2_S**
+- `blk.3.ffn_down.weight`: **IQ2_S**
 
 | Precision | Count |
 |---|---:|
@@ -111,7 +109,7 @@ The measured Q1 V2 policy differs from the successful IQ2_S release by exactly *
 | Q6_K | **2** |
 | **Total** | **851** |
 
-| Build item | Measured V2 value |
+| Build item | Value |
 |---|---|
 | Source | Canonical BF16 |
 | Low-bit requantization | **No** |
@@ -123,20 +121,21 @@ The measured Q1 V2 policy differs from the successful IQ2_S release by exactly *
 | End-to-end quantization timer | **168.26090 s** |
 | SHA256 | `e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b` |
 
-## 7. IQ1_M quantitative gates
-
-| Gate | Hard | Reserve | Measured V2 | Result |
-|---|---:|---:|---:|---|
-| PPL ratio | ≤1.0150 | ≤1.0135 | **1.003191** | **PASS / PASS** |
-| Mean KLD | ≤0.0150 | ≤0.0145 | **0.014357** | **PASS / PASS** |
-| Same top-p | ≥95.0% | ≥95.5% | **95.870%** | **PASS / PASS** |
-| Tensor structure | exact | exact | **851; 0 mismatch** | **PASS** |
+## 7. IQ1_M final validation
 
-Reserve headroom is **0.010309** on PPL ratio, **0.000143** on Mean KLD, and **0.370 pp** on Same top-p. Mean KLD is the binding Q1 reserve metric.
+| Validation | Requirement | Measured | Result |
+|---|---|---|---|
+| PPL ratio hard / reserve | ≤1.0150 / ≤1.0135 | **1.003191** | **PASS / PASS** |
+| Mean KLD hard / reserve | ≤0.0150 / ≤0.0145 | **0.014357** | **PASS / PASS** |
+| Same top-p hard / reserve | ≥95.0% / ≥95.5% | **95.870%** | **PASS / PASS** |
+| Tensor structure | exact policy | **851; 0 mismatch** | **PASS** |
+| Stock llama.cpp runtime | real HTTP generation | **200; non-empty output** | **PASS** |
+| MTP runtime | real HTTP generation | **200; non-empty output** | **PASS** |
+| MTP engagement | generated > 0; accepted > 0 | **76 / 104 accepted (73.0769%)** | **PASS** |
+| Acceptance-rate consistency | exact record | **0.73076923** | **PASS** |
+| Main/MTP output audit | advisory | **IDENTICAL** | **PASS** |
 
-The V2 artifact was rejected by an additional experimental rule requiring Q1 to beat IQ2_S on **every selected point estimate**. That rule was stricter than the frozen release-quality gate and has been removed from V3. The **hard and reserve thresholds themselves are unchanged**.
-
-## 8. IQ1_M vs IQ2_S vs Q3_K_M frontier
+## 8. Low-footprint comparison
 
 | Metric | Q3_K_M | IQ2_S | IQ1_M |
 |---|---:|---:|---:|
@@ -146,47 +145,28 @@ The V2 artifact was rejected by an additional experimental rule requiring Q1 to
 | Same top-p | **95.919%** | **95.516%** | **95.870%** |
 | RMS Δp | **3.515%** | **3.679%** | **3.691%** |
 
-IQ1_M is **8.633 MiB** smaller than IQ2_S and **36.523 MiB** smaller than Q3_K_M. Versus IQ2_S it has a better PPL point estimate and **+0.354 pp** Same top-p, but Mean KLD is **+0.000334 / +2.38%** higher. The differences are smaller than the reported uncertainty scale, so this record treats IQ1_M and IQ2_S as **neighboring low-footprint Pareto points**, not as proof of categorical fidelity superiority.
-
-## 9. Release-readiness distinction
-
-| Evidence layer | IQ2_S | IQ1_M |
-|---|---|---|
-| Exact tensor structure | PASS | PASS |
-| Frozen BF16-paired hard gate | PASS | PASS |
-| Engineering reserve gate | PASS | PASS |
-| Immutable final SHA256 | PASS | **PASS — local identity frozen** |
-| Stock llama.cpp runtime | PASS | **Pending** |
-| MTP load/generation/engagement | PASS | **Pending** |
-| Remote exact-size + full-SHA + readability | PASS | **Pending** |
-| Current public positioning | **Released low-footprint sweet spot** | **Ultra-low-footprint frontier candidate** |
-
-This distinction is why IQ2_S remains the current released sweet spot: IQ1_M now has quantitative and local-identity PASS, but stock runtime, MTP, and remote verification are not yet closed.
-
-## 10. IQ2_S retained release record
-
-IQ2_S remains fully verified at **18,037,261,056 bytes / 16.798508 GiB**, SHA256 `0dd44d41319efa9f383b9a867b5fd2114335ccd0a653d1c3e7605d535edc601b`, with PPL ratio **1.003457**, Mean KLD **0.014023**, Same top-p **95.516%**, stock runtime PASS, MTP PASS, and remote exact-size/full-SHA/readability PASS.
+IQ1_M is **8.633 MiB** smaller than IQ2_S and **36.523 MiB** smaller than Q3_K_M. IQ2_S has the lowest Mean KLD, while IQ1_M has the smallest footprint, the lowest PPL ratio, and higher Same top-p than IQ2_S. The differences remain small relative to the reported uncertainty scale.
 
-## 11. Artifact identity
+## 9. Frozen artifact identities
 
-| Artifact | Bytes | GiB | SHA256 | State |
-|---|---:|---:|---|---|
-| BF16 | **53,808,284,000** | **50.112870** | `11bf5defde1a256b7582bc34fd2c4a85a61615ed88e7a422dcd24d814ea6d35d` | Verified |
-| Q8_0 | **28,595,765,600** | **26.631882** | `6204a47274cfbc0c69c39877fb06615ce842bbab264eea77e2e0a5e3ae2fb8e8` | Verified |
-| Q6_K | **22,082,532,096** | **20.565961** | `15d8f856471c4853f6bf0036b2a517426c6cb30a0313cef58a7f9577fd26fe9e` | Verified |
-| Q5_K_M | **20,363,666,176** | **18.965142** | `f90ec14d7ec8f084a292413ae0e06483c87ab5e1422ddce51f2e3409d8853b61` | Verified |
-| Q4_K_M | **19,650,634,496** | **18.301080** | not frozen in supplied Q4 record | Quantitative candidate |
-| Q3_K_M | **18,066,506,496** | **16.825745** | `c2e9539cfb85d87a99605b6914ed602c0fa750164c153ce5574710123aba06fe` | Verified release |
-| IQ2_S | **18,037,261,056** | **16.798508** | `0dd44d41319efa9f383b9a867b5fd2114335ccd0a653d1c3e7605d535edc601b` | Verified release |
-| **IQ1_M** | **18,028,208,896** | **16.790078** | `e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b` | **Quantitative + local identity PASS; runtime/MTP/remote pending** |
+| Artifact | Bytes | GiB | SHA256 |
+|---|---:|---:|---|
+| BF16 | **53,808,284,000** | **50.112870** | `11bf5defde1a256b7582bc34fd2c4a85a61615ed88e7a422dcd24d814ea6d35d` |
+| Q8_0 | **28,595,765,600** | **26.631882** | `6204a47274cfbc0c69c39877fb06615ce842bbab264eea77e2e0a5e3ae2fb8e8` |
+| Q6_K | **22,082,532,096** | **20.565961** | `15d8f856471c4853f6bf0036b2a517426c6cb30a0313cef58a7f9577fd26fe9e` |
+| Q5_K_M | **20,363,666,176** | **18.965142** | `f90ec14d7ec8f084a292413ae0e06483c87ab5e1422ddce51f2e3409d8853b61` |
+| Q3_K_M | **18,066,506,496** | **16.825745** | `c2e9539cfb85d87a99605b6914ed602c0fa750164c153ce5574710123aba06fe` |
+| IQ2_S | **18,037,261,056** | **16.798508** | `0dd44d41319efa9f383b9a867b5fd2114335ccd0a653d1c3e7605d535edc601b` |
+| **IQ1_M** | **18,028,208,896** | **16.790078** | `e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b` |
 
-## 12. Evidence-supported conclusion
+Q4_K_M remains part of the measured retention ladder; this identity table intentionally lists only artifacts with a frozen SHA256 in the supplied release record.
 
-The deployment ladder currently has three confirmed objective-specific sweet spots plus one frontier candidate:
+## 10. Conclusion
 
 - **Q6_K — overall quality / efficiency sweet spot.**
 - **Q5_K_M — memory-quality sweet spot.**
-- **IQ2_S — current released low-footprint sweet spot.**
-- **IQ1_M — ultra-low-footprint frontier candidate with frozen local identity; promote to minimum-footprint sweet spot only after stock runtime/MTP/remote identity PASS.**
+- **IQ1_M — minimum-footprint sweet spot.**
+- **IQ2_S — lower-KLD low-footprint alternative.**
+- **Q3_K_M — adjacent low-footprint alternative.**
 
-If IQ1_M reproduces the V2 fidelity record and closes those final release gates, the professionally accurate positioning becomes: **Q6_K overall**, **Q5_K_M memory-quality**, **IQ1_M minimum-footprint**, and **IQ2_S lower-KLD low-footprint alternative**. That is a deployment trade-off classification, not a claim that lower nominal bit count is universally better.
+This classification is deployment-oriented. It does not imply that lower nominal bit count is universally superior.

README.md

@@ -39,7 +39,7 @@ 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-quality sweet spot · IQ2_S released low-footprint sweet spot · IQ1_M ultra-low-footprint frontier</em><br><br>
+  <em>BF16 reference · Q8_0 high fidelity · Q6_K overall sweet spot · Q5_K_M memory-quality sweet spot · IQ1_M minimum-footprint sweet spot</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>
@@ -49,8 +49,7 @@ tags:
   <img src="https://img.shields.io/badge/Format-GGUF-111827?style=flat-square" alt="Format: GGUF">
   <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/IQ2__S-Low--Footprint%20Sweet%20Spot-475569?style=flat-square" alt="IQ2_S: Low-Footprint Sweet Spot">
-  <img src="https://img.shields.io/badge/IQ1__M-Ultra--Low--Footprint%20Frontier-334155?style=flat-square" alt="IQ1_M: Ultra-Low-Footprint Frontier">
+  <img src="https://img.shields.io/badge/IQ1__M-Minimum%20Footprint-334155?style=flat-square" alt="IQ1_M: Minimum Footprint Sweet Spot">
   <img src="https://img.shields.io/badge/Runtime-llama.cpp-0A7F6F?style=flat-square" alt="Runtime: llama.cpp">
 </p>
 
@@ -63,48 +62,45 @@ tags:
 
 ---
 
-## Which file should I download?
+## Model variants
 
-| 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%** | 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%** | Smaller deployment with a wider fidelity margin |
-| **Released low-footprint sweet spot** | **`VeriLoop-E2-IQ2_S.gguf`** | **16.799 GiB** | PPL **+0.3457%**; KLD **0.014023**; Same top-p **95.516%** | Smallest fully release-verified low-footprint tier |
-| **Ultra-low-footprint frontier candidate** | **`VeriLoop-E2-IQ1_M.gguf`** | **16.790 GiB** | PPL **+0.3191%**; KLD **0.014357**; Same top-p **95.870%** | Smallest measured candidate; quantitative fidelity and local identity verified |
-| **Low-footprint alternative** | `VeriLoop-E2-Q3_K_M.gguf` | **16.826 GiB** | PPL **+0.4090%**; KLD **0.014349**; Same top-p **95.919%** | Adjacent fully validated option |
-| **High fidelity** | `VeriLoop-E2-Q8_0.gguf` | **26.632 GiB** | PPL **+0.0643%**; KLD **0.002176**; Same top-p **98.815%** | Distributional fidelity matters more than footprint |
-| **Reference** | `VeriLoop-E2-BF16.gguf` | **50.113 GiB** | Canonical reference | Reproducibility, attribution, quantization research |
+| Use case | File | Main size | BF16-paired retention |
+|---|---|---:|---|
+| **Default / overall balance** | **`VeriLoop-E2-Q6_K.gguf`** | **20.566 GiB** | PPL parity within uncertainty; KLD **0.004409**; Same top-p **98.204%** |
+| **Memory-quality balance** | **`VeriLoop-E2-Q5_K_M.gguf`** | **18.965 GiB** | PPL **+0.4450%**; KLD **0.006919**; Same top-p **97.251%** |
+| **Minimum footprint** | **`VeriLoop-E2-IQ1_M.gguf`** | **16.790 GiB** | PPL **+0.3191%**; KLD **0.014357**; Same top-p **95.870%** |
+| Lower-KLD low-footprint alternative | `VeriLoop-E2-IQ2_S.gguf` | **16.799 GiB** | PPL **+0.3457%**; KLD **0.014023**; Same top-p **95.516%** |
+| Low-footprint alternative | `VeriLoop-E2-Q3_K_M.gguf` | **16.826 GiB** | PPL **+0.4090%**; KLD **0.014349**; Same top-p **95.919%** |
+| Balanced compact | `VeriLoop-E2-Q4_K_M.gguf` | **18.301 GiB** | PPL **+0.4821%**; KLD **0.009700**; Same top-p **96.786%** |
+| High fidelity | `VeriLoop-E2-Q8_0.gguf` | **26.632 GiB** | PPL **+0.0643%**; KLD **0.002176**; Same top-p **98.815%** |
+| Reference | `VeriLoop-E2-BF16.gguf` | **50.113 GiB** | Canonical BF16 reference |
 
-### Current sweet spots and frontier
+### Recommended deployment points
 
-**Q6_K remains the overall quality / efficiency sweet spot.** It is **58.96% smaller than BF16** while its PPL point estimate remains statistically consistent with BF16 parity under the frozen paired protocol.
+**Q6_K — overall quality / efficiency sweet spot.** It is **58.96% smaller than BF16** while remaining statistically consistent with BF16 PPL parity under the frozen paired protocol.
 
-**Q5_K_M remains the memory-quality sweet spot.** It offers a materially smaller footprint than Q6_K while keeping substantially wider KLD and Same-top margins than the sub-17 GiB frontier.
+**Q5_K_M — memory-quality sweet spot.** It reduces the main-file footprint to **18.965 GiB** while preserving wider KLD and Same-top margins than the sub-17 GiB variants.
 
-**IQ2_S remains the current released low-footprint sweet spot.** It is fully verified through structure, BF16-paired fidelity, engineering reserve, stock llama.cpp runtime, MTP engagement, immutable SHA256, and remote readability.
+**IQ1_M — minimum-footprint sweet spot.** It is **16.790078 GiB**, **66.4955% smaller than BF16**, and passes the frozen hard gate, engineering-reserve gate, stock llama.cpp runtime validation, and real MTP engagement validation. IQ2_S remains the lower-KLD low-footprint alternative.
 
-**IQ1_M is the current ultra-low-footprint frontier candidate.** Its measured V2 artifact is **16.790078 GiB**, **66.4955% smaller than BF16**, **8.633 MiB / 0.0502% smaller than IQ2_S**, and **36.523 MiB / 0.2120% smaller than Q3_K_M**. Against IQ2_S, IQ1_M has a slightly better PPL ratio (**1.003191 vs 1.003457**) and higher Same top-p (**95.870% vs 95.516%**), but a slightly higher Mean KLD (**0.014357 vs 0.014023**). The uncertainty bands substantially overlap, so this release does **not** claim that either low-footprint tier is universally superior in fidelity.
-
-**Positioning.** IQ1_M defines the current minimum-footprint frontier, while IQ2_S remains the released low-footprint balance point / lower-KLD alternative.
-
-> **Naming note:** `VeriLoop-E2-IQ1_M.gguf` is not a uniform 1-bit model. Its measured policy is **353 F32 + 1 IQ1_M + 2 IQ2_S + 64 Q4_K + 429 Q5_K + 2 Q6_K = 851 tensors**, with **5.36 effective BPW**. `VeriLoop-E2-IQ2_S.gguf` is likewise a mixed-precision release rather than a uniform 2-bit model.
+> **Naming note:** `VeriLoop-E2-IQ1_M.gguf` is a mixed-precision artifact, not a uniform 1-bit model. Its measured tensor policy is **353 F32 + 1 IQ1_M + 2 IQ2_S + 64 Q4_K + 429 Q5_K + 2 Q6_K = 851 tensors**, with **5.36 effective BPW**. `VeriLoop-E2-IQ2_S.gguf` is likewise mixed precision rather than uniform 2-bit quantization.
 
 ## Precision ladder
 
-| 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 |
-| **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 mixed-precision candidate | **19.651 GB / 18.301 GiB** | **63.48%** | **5.84 BPW** | Quantitatively validated candidate |
-| **Q3_K_M** | Adjacent low-footprint alternative | **18.067 GB / 16.826 GiB** | **66.42%** | **5.37 BPW** | Full release PASS |
-| **IQ2_S** | **Released low-footprint sweet spot** | **18.037 GB / 16.799 GiB** | **66.48%** | **5.36 BPW** | **Full release PASS** |
-| **IQ1_M** | **Ultra-low-footprint frontier candidate** | **18.028 GB / 16.790 GiB** | **66.50%** | **5.36 BPW** | **Quantitative fidelity PASS; local identity verified** |
+| Tier | Role | Main size | Reduction vs BF16 | Effective density |
+|---|---|---:|---:|---:|
+| **BF16** | Canonical reference | **53.808 GB / 50.113 GiB** | — | 16-bit-class |
+| **Q8_0** | High fidelity | **28.596 GB / 26.632 GiB** | **46.86%** | **8.50 BPW** |
+| **Q6_K** | **Overall sweet spot** | **22.083 GB / 20.566 GiB** | **58.96%** | **6.57 BPW** |
+| **Q5_K_M** | **Memory-quality sweet spot** | **20.364 GB / 18.965 GiB** | **62.16%** | **6.05 BPW** |
+| **Q4_K_M** | Balanced compact | **19.651 GB / 18.301 GiB** | **63.48%** | **5.84 BPW** |
+| **Q3_K_M** | Low-footprint alternative | **18.067 GB / 16.826 GiB** | **66.42%** | **5.37 BPW** |
+| **IQ2_S** | Lower-KLD low-footprint alternative | **18.037 GB / 16.799 GiB** | **66.48%** | **5.36 BPW** |
+| **IQ1_M** | **Minimum-footprint sweet spot** | **18.028 GB / 16.790 GiB** | **66.50%** | **5.36 BPW** |
 
 ## Quantization-retention benchmark
 
-All tiers below use the same frozen BF16 logits and the same paired protocol.
+All measured tiers use the same frozen BF16 logits and the same paired protocol.
 
 | Tier | Mean PPL | PPL ratio vs BF16 | Relative PPL change | Mean KLD | Same top-p | log-PPL correlation |
 |---|---:|---:|---:|---:|---:|---:|
@@ -117,15 +113,11 @@ All tiers below use the same frozen BF16 logits and the same paired protocol.
 | **IQ2_S** | **4.857159 ± 0.120482** | **1.003457 ± 0.002100** | **+0.3457%** | **0.014023 ± 0.001408** | **95.516 ± 0.229%** | **99.64%** |
 | **IQ1_M** | **4.855870 ± 0.120480** | **1.003191 ± 0.002175** | **+0.3191%** | **0.014357 ± 0.001317** | **95.870 ± 0.220%** | **99.62%** |
 
-### What “performance loss” means here
-
-The **+0.3191% IQ1_M** and **+0.3457% IQ2_S** figures are **PPL drift under the frozen BF16-paired quantization benchmark**. They are not claims that SWE-bench, Terminal-Bench, DeepSWE, AIME, GPQA, Apex, coding-agent capability, mathematics, or physics scores fall by those percentages. The parent nine-benchmark campaign has **not** been rerun independently for each quant tier, so no quant-specific downstream score-loss percentage is claimed.
+These figures measure **quantization retention against the BF16 reference**. They are not downstream benchmark-score loss percentages. The nine parent-model benchmarks were not independently rerun for every quantization tier.
 
-For the low-footprint frontier, PPL must be read together with KLD, Same top-p, RMS Δp, and uncertainty. IQ1_M improves the PPL point estimate and Same top-p versus IQ2_S, while IQ2_S has **2.38% lower Mean KLD** and slightly lower RMS Δp. These differences are small relative to their reported uncertainties; the professional interpretation is **trade-off parity**, not a claim of categorical quality superiority.
+## IQ1_M validation record
 
-## IQ1_M — ultra-low-footprint frontier candidate
-
-The measured IQ1_M policy is a **single-factor step from the successful IQ2_S release**: only `blk.1.ffn_down.weight` changes from IQ2_S to IQ1_M. Every other tensor assignment remains on the Q2 policy.
+IQ1_M changes exactly one tensor relative to the IQ2_S precision policy: `blk.1.ffn_down.weight` moves from IQ2_S to IQ1_M. All other tensor assignments remain unchanged.
 
 | Precision | Tensor assignment | Count |
 |---|---|---:|
@@ -137,112 +129,44 @@ The measured IQ1_M policy is a **single-factor step from the successful IQ2_S re
 | IQ1_M | `blk.1.ffn_down.weight` | **1** |
 | **Total** |  | **851** |
 
-### IQ1_M measured retention
-
-| Metric | BF16 | IQ1_M | Observed drift |
-|---|---:|---:|---:|
-| Mean PPL | **4.840423 ± 0.119931** | **4.855870 ± 0.120480** | **+0.015446 ± 0.010522** |
-| PPL ratio | 1.000000 | **1.003191 ± 0.002175** | **+0.3191%** |
-| Mean KLD | 0 | **0.014357 ± 0.001317** | lower is better |
-| Same top-p | 100% | **95.870 ± 0.220%** | **4.130 pp disagreement** |
-| RMS Δp | 0 | **3.691 ± 0.205%** | distributional-noise scale |
-| Mean Δp | 0 | **−0.042 ± 0.041%** | small directional bias |
-| log-PPL correlation | 100% | **99.62%** | high agreement |
-
-### IQ1_M final quantitative gates
-
-| Gate | Hard threshold | Reserve threshold | Measured V2 | Quantitative result |
-|---|---:|---:|---:|---|
-| PPL ratio | ≤ **1.0150** | ≤ **1.0135** | **1.003191** | **PASS / PASS** |
-| Mean KLD | ≤ **0.0150** | ≤ **0.0145** | **0.014357** | **PASS / PASS** |
-| Same top-p | ≥ **95.0%** | ≥ **95.5%** | **95.870%** | **PASS / PASS** |
-| Exact tensor structure | declared policy | declared policy | **851 tensors; 0 mismatch** | **PASS** |
-
-The original V2 script rejected this artifact only because an additional experimental promotion rule demanded that **every Q1 point estimate also dominate IQ2_S**, including Mean KLD. That requirement has been removed from V3; the underlying hard and engineering-reserve thresholds are unchanged. The V2 artifact itself was deleted after that rejection, so its SHA256 is not represented as a frozen release identity. The accepted rebuild has frozen the identical tensor policy and local SHA256; runtime/MTP/remote identity must still close before public release.
-
-### IQ1_M vs IQ2_S — low-footprint frontier
-
-| Metric | IQ2_S | IQ1_M | IQ1_M delta |
-|---|---:|---:|---:|
-| Main size | **16.798508 GiB** | **16.790078 GiB** | **−8.633 MiB / −0.0502%** |
-| PPL ratio | **1.003457** | **1.003191** | **−0.000266** |
-| Mean KLD | **0.014023** | **0.014357** | **+0.000334 / +2.38%** |
-| Same top-p | **95.516%** | **95.870%** | **+0.354 pp** |
-| RMS Δp | **3.679%** | **3.691%** | **+0.012 pp** |
-| log-PPL correlation | **99.64%** | **99.62%** | **−0.02 pp** |
-
-**Interpretation:** IQ1_M is not merely “worse Q2.” It is a neighboring Pareto point: **smaller, slightly better on PPL and top-token agreement, slightly worse on KLD/RMS**, with overlapping uncertainty. Its major deployment advantage is the **absolute 16.790 GiB footprint / 66.4955% reduction from BF16**; the incremental saving over IQ2_S itself is only **8.633 MiB**, so that incremental difference should not be exaggerated.
-
-## IQ2_S — root-cause-guided low-footprint sweet spot
-
-IQ2_S is derived **directly from the canonical BF16 GGUF**. It preserves the successful Q3 precision spine and changes only the three root-cause-selected FFN-down tensors from Q3_K to IQ2_S.
-
-| Precision | Tensor assignment | Count |
-|---|---|---:|
-| F32 | Non-quantized tensors retained by GGUF conversion | **353** |
-| Q6_K | `output.weight`, `token_embd.weight` | **2** |
-| Q5_K | Remaining quantized internal tensors | **429** |
-| Q4_K | all `*.ffn_up.weight` | **64** |
-| IQ2_S | `blk.0.ffn_down.weight`, `blk.1.ffn_down.weight`, `blk.3.ffn_down.weight` | **3** |
-| **Total** |  | **851** |
-
-### Q2 development path
-
-| Stage | Targeted change | Main size | PPL ratio | Mean KLD | Same top-p | Decision |
-|---|---|---:|---:|---:|---:|---|
-| IQ2_S V1 | Four low-sensitivity `ffn_up` tensors → IQ2_S | **16.809533 GiB** | **1.008995** | **0.014951** | **95.809%** | **FAIL reserve KLD** |
-| **IQ2_S V2 final** | **Q3 spine; three selected `ffn_down` tensors → IQ2_S** | **16.798508 GiB** | **1.003457** | **0.014023** | **95.516%** | **PASS hard + reserve** |
-
-V1 was rejected rather than promoted: it saved only **0.0964%** versus Q3 and exceeded the frozen **0.0145** Mean-KLD reserve. V2 changes the tensor family instead of relaxing the threshold.
+### Fidelity and runtime validation
 
-### IQ2_S final gates
+| Validation | Requirement | Measured | Result |
+|---|---|---|---|
+| PPL ratio hard / reserve | ≤ **1.0150** / ≤ **1.0135** | **1.003191** | **PASS / PASS** |
+| Mean KLD hard / reserve | ≤ **0.0150** / ≤ **0.0145** | **0.014357** | **PASS / PASS** |
+| Same top-p hard / reserve | ≥ **95.0%** / ≥ **95.5%** | **95.870%** | **PASS / PASS** |
+| Exact tensor structure | 851 tensors; declared policy | **851 tensors; 0 mismatch** | **PASS** |
+| Stock llama.cpp main runtime | Successful real generation | HTTP **200**, non-empty generation | **PASS** |
+| MTP runtime | Successful real generation | HTTP **200**, non-empty generation | **PASS** |
+| Real MTP engagement | generated > 0; accepted > 0 | **76 accepted / 104 generated (73.0769%)** | **PASS** |
+| Main vs MTP deterministic audit | Advisory | Identical output SHA256 | **IDENTICAL** |
 
-| Gate | Hard threshold | Reserve threshold | Measured | Result |
-|---|---:|---:|---:|---|
-| PPL ratio | ≤ **1.0150** | ≤ **1.0135** | **1.003457** | **PASS / PASS** |
-| Mean KLD | ≤ **0.0150** | ≤ **0.0145** | **0.014023** | **PASS / PASS** |
-| Same top-p | ≥ **95.0%** | ≥ **95.5%** | **95.516%** | **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** |
-| Real MTP engagement | >0 accepted | >0 accepted | **116 / 331 accepted (35.045%)** | **PASS** |
-
-Reserve headroom is **0.010043** on PPL ratio, **0.000477** on Mean KLD, and **0.016 pp** on Same top-p. The final value is intentionally exposed: **this is a validated frontier, not evidence that further low-bit expansion is safe.**
-
-## IQ2_S immutable identity
+**IQ1_M artifact identity**
 
 | Property | Value |
 |---|---|
-| Filename | `VeriLoop-E2-IQ2_S.gguf` |
-| Exact bytes | **18,037,261,056** |
-| Binary size | **16.798508 GiB** |
-| SHA256 | `0dd44d41319efa9f383b9a867b5fd2114335ccd0a653d1c3e7605d535edc601b` |
+| Filename | `VeriLoop-E2-IQ1_M.gguf` |
+| Exact bytes | **18,028,208,896** |
+| Binary size | **16.790078 GiB** |
+| SHA256 | `e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b` |
 | Effective density | **5.36 BPW** |
 | Tensor count | **851** |
-| Tensor distribution | **353 F32 + 3 IQ2_S + 64 Q4_K + 429 Q5_K + 2 Q6_K** |
-| Quantized size reported by quantizer | **17,191.19 MiB** |
-| Quantization time | **168.08618 s** |
-| llama.cpp QC/runtime revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
-| Hugging Face remote identity | **PASS — exact size + full SHA256 + remote readability** |
-
-## Q3_K_M — adjacent low-footprint alternative
+| Quantizer-reported size | **17,182.55 MiB** |
+| Quantizer time | **167.77645 s** |
+| llama.cpp validation revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
 
-Q3 remains a valid release. It has **0.403 pp higher Same top-p** than IQ2_S, while IQ2_S is **27.891 MiB / 0.162% smaller**, with a lower PPL ratio and **2.27% lower Mean KLD**. For minimum footprint, IQ2_S is the preferred frontier point; Q3 remains the adjacent option when top-token agreement is prioritized.
+### Low-footprint comparison
 
-## Optional MTP draft
-
-IQ2_S and Q3 were validated with the same 18-tensor Q4_0 MTP object:
-
-| Property | Value |
-|---|---|
-| MTP bytes | **2,008,056,096 / 1.870 GiB** |
-| MTP SHA256 | `0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671` |
-| IQ2_S draft tokens generated | **331** |
-| IQ2_S draft tokens accepted | **116** |
-| IQ2_S acceptance rate | **0.35045 / 35.045%** |
-| Mean draft length | **3.76** |
+| Metric | Q3_K_M | IQ2_S | IQ1_M |
+|---|---:|---:|---:|
+| Main size | **16.825745 GiB** | **16.798508 GiB** | **16.790078 GiB** |
+| PPL ratio | **1.004090** | **1.003457** | **1.003191** |
+| Mean KLD | **0.014349** | **0.014023** | **0.014357** |
+| Same top-p | **95.919%** | **95.516%** | **95.870%** |
+| RMS Δp | **3.515%** | **3.679%** | **3.691%** |
 
-A duplicate Q2-specific MTP file is not required for this validated path. Main-only and main+MTP continuations diverged in the deterministic audit; that result is retained as an advisory. MTP engagement is established by explicit draft-model loading plus non-zero generated and accepted draft tokens.
+IQ1_M is **8.633 MiB** smaller than IQ2_S and **36.523 MiB** smaller than Q3_K_M. IQ2_S retains the lowest Mean KLD of the three; IQ1_M has the smallest footprint, the lowest PPL ratio, and higher Same top-p than IQ2_S. The point-estimate differences remain small relative to the reported uncertainty scale.
 
 ## Frozen BF16-paired protocol
 
@@ -260,13 +184,13 @@ A duplicate Q2-specific MTP file is not required for this validated path. Main-o
 | Batch / micro-batch | **512 / 512** |
 | Evaluator | `llama-perplexity` |
 | Reference logits | `--kl-divergence-base` |
-| Candidate comparison | `--kl-divergence` |
+| Quantized comparison | `--kl-divergence` |
 | BF16 logits reused across tiers | **Yes** |
 | llama.cpp revision | `42916d83f4a225e56709f873aa8050ac11f5b6a4` |
 
 ## Parent-model benchmark record
 
-These scores identify the **parent VeriLoop E2 release**. They provide capability context and are **not relabeled as quant-specific re-runs**.
+The scores below describe the **parent VeriLoop E2 release** and are not relabeled as quantization-specific reruns.
 
 | Benchmark | VeriLoop E2 parent score | Public evidence |
 |---|---:|---|
@@ -286,66 +210,41 @@ These scores identify the **parent VeriLoop E2 release**. They provide capabilit
 # Overall sweet spot
 hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-Q6_K.gguf --local-dir .
 
-# Released low-footprint sweet spot
-hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-IQ2_S.gguf --local-dir .
-```
+# Minimum-footprint sweet spot
+hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF VeriLoop-E2-IQ1_M.gguf --local-dir .
 
-> **IQ1_M is staged as the ultra-low-footprint frontier artifact.** Its quantitative evidence is PASS and its local SHA256 is frozen as `e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b`.
-
-Run IQ2_S:
-
-```bash
-llama-server   -m ./VeriLoop-E2-IQ2_S.gguf   -ngl 99   -c 32768   --host 127.0.0.1   --port 8080
+# Run
+llama-server \
+  -m ./VeriLoop-E2-IQ1_M.gguf \
+  -ngl 99 \
+  -c 32768 \
+  --host 127.0.0.1 \
+  --port 8080
 ```
 
-IQ2_S + validated MTP companion:
+## Memory guidance
 
-```bash
-hf download tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF mtp-VeriLoop-E2-Q5_K_M.gguf --local-dir .
-
-llama-server   -m ./VeriLoop-E2-IQ2_S.gguf   --model-draft ./mtp-VeriLoop-E2-Q5_K_M.gguf   --spec-type draft-mtp   --spec-draft-n-max 8   --gpu-layers-draft 99   -ngl 99   -c 2048   --host 127.0.0.1   --port 8080
-```
-
-## Memory and throughput guidance
-
-| Variant | Main file | Position |
+| Variant | Main file | Positioning |
 |---|---:|---|
 | BF16 | **50.113 GiB** | Reference |
 | Q8_0 | **26.632 GiB** | High fidelity |
 | **Q6_K** | **20.566 GiB** | **Overall sweet spot** |
 | **Q5_K_M** | **18.965 GiB** | **Memory-quality sweet spot** |
-| Q4_K_M | **18.301 GiB** | Intermediate quantitative candidate |
-| Q3_K_M | **16.826 GiB** | Adjacent low-footprint release |
-| **IQ2_S** | **16.799 GiB** | **Released low-footprint sweet spot** |
-| **IQ1_M** | **16.790 GiB** | **Ultra-low-footprint frontier candidate** |
-
-A **16.79 GiB model file does not imply full offload on a 16 GiB GPU**. Runtime memory also includes KV cache, compute buffers, allocator overhead, and optional MTP weights.
+| Q4_K_M | **18.301 GiB** | Balanced compact |
+| Q3_K_M | **16.826 GiB** | Low-footprint alternative |
+| IQ2_S | **16.799 GiB** | Lower-KLD low-footprint alternative |
+| **IQ1_M** | **16.790 GiB** | **Minimum-footprint sweet spot** |
 
-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 campaign. No universal speedup percentage is claimed for Q6/Q5/Q4/Q3/IQ2_S/IQ1_M.
+A **16.79 GiB model file does not imply full offload on a 16 GiB GPU**. Runtime memory also includes KV cache, compute buffers, allocator overhead, and optional speculative-decoding weights.
 
-## Release positioning
-
-| Tier | Positioning |
-|---|---|
-| BF16 | Canonical reference |
-| Q8_0 | High-fidelity local |
-| **Q6_K** | **Overall quality / efficiency sweet spot** |
-| **Q5_K_M** | **Memory-quality sweet spot** |
-| Q4_K_M | Intermediate mixed-precision candidate |
-| Q3_K_M | Adjacent low-footprint release |
-| **IQ2_S** | **Current released low-footprint sweet spot / lower-KLD frontier point** |
-| **IQ1_M** | **Ultra-low-footprint frontier candidate** |
+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 campaign. No universal speedup percentage is claimed for the quantization ladder.
 
 ## Measurement boundaries
 
-- PPL/KLD/token metrics are **quantization-retention measurements**, not universal percentages of downstream capability loss.
-- IQ1_M's **+0.3191% PPL** does not mean a 0.3191% drop on SWE-bench, Terminal-Bench, AIME, GPQA, or another downstream benchmark.
-- IQ2_S's **+0.3457% PPL** likewise does not imply a 0.3457% downstream-score loss.
-- IQ1_M and IQ2_S are both mixed-precision artifacts; the nominal tier name is not the model-wide effective bit width.
-- IQ1_M has a frozen local SHA256 (`e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b`) and quantitative PASS evidence.
-- The IQ1_M vs IQ2_S point-estimate differences are small relative to the reported uncertainty; this repository does not claim categorical fidelity superiority for either frontier point.
+- PPL, KLD, Same top-p, and token-probability statistics are **quantization-retention measurements**, not universal downstream capability-loss percentages.
+- IQ1_M and IQ2_S are mixed-precision artifacts; the tier name does not equal the model-wide effective bit width.
 - Native 262K context comes from the parent configuration; practical context depends on runtime memory.
-- MTP acceptance is prompt/workload-dependent.
+- MTP acceptance is prompt- and workload-dependent.
 - The model can still produce incorrect code, mathematics, scientific reasoning, or commands.
 
 ## Source model and evidence
@@ -355,7 +254,7 @@ A standalone Q8_0 `llama-bench` record exists (**pp512 3102.722776 tok/s; tg128
 | 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) |
-| Artifact identity / release state | [Release Manifest](./RELEASE_MANIFEST.json) |
+| Artifact manifest | [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) |
@@ -378,4 +277,4 @@ The VeriLoop E2 model weights and this GGUF distribution are released under the
 }
 ```
 
-For quantization-specific comparisons, identify the exact GGUF filename and corresponding release-manifest identity.
+For quantization-specific comparisons, identify the exact GGUF filename and corresponding manifest identity.

RELEASE_MANIFEST.json

@@ -1,5 +1,5 @@
 {
-  "schema": "veriloop.e2.gguf.release_manifest.v8_q1_local_identity_frozen",
+  "schema": "veriloop.e2.gguf.release_manifest.v9",
   "repository": "tsinghua-sigs-robot-lab/VeriLoop-E2-GGUF",
   "model": {
     "name": "VeriLoop E2",
@@ -16,23 +16,18 @@
     "quantization_methodology": "direct_from_BF16_no_low_bit_requantization",
     "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,
-    "lower_bit_alone_is_not_a_recommendation_criterion": true,
-    "q3_filename_semantics": "custom_mixed_precision_release_tier_minimum_active_precision_Q3_K_not_uniform_3bit",
-    "q2_filename_semantics": "custom_mixed_precision_release_tier_minimum_active_precision_IQ2_S_not_uniform_2bit",
-    "q1_filename_semantics": "custom_mixed_precision_release_tier_minimum_active_precision_IQ1_M_not_uniform_1bit",
-    "q1_prerelease_status": "QUANTITATIVE_PASS_LOCAL_IDENTITY_FROZEN_RUNTIME_MTP_REMOTE_PENDING",
-    "sweet_spot_assignment_requires_full_release_validation": true
+    "performance_claim_boundary": "PPL_KLD_and_token_drift_are_quantization_fidelity_metrics_not_universal_task_score_loss",
+    "q3_filename_semantics": "mixed_precision_tier_minimum_active_precision_Q3_K",
+    "q2_filename_semantics": "mixed_precision_tier_minimum_active_precision_IQ2_S",
+    "q1_filename_semantics": "mixed_precision_tier_minimum_active_precision_IQ1_M",
+    "internal_release_readiness": "PASS"
   },
-  "sweet_spots": {
-    "BF16": "reference_fidelity",
-    "Q8_0": "high_fidelity_local",
-    "Q6_K": "overall_quality_efficiency_sweet_spot",
-    "Q5_K_M": "memory_quality_sweet_spot",
-    "Q3_K_M": "adjacent_low_footprint_alternative",
-    "IQ2_S": "current_released_low_footprint_sweet_spot",
-    "IQ1_M": "ultra_low_footprint_frontier_candidate_pending_full_release_validation"
+  "recommended_variants": {
+    "overall_sweet_spot": "Q6_K",
+    "memory_quality_sweet_spot": "Q5_K_M",
+    "minimum_footprint_sweet_spot": "IQ1_M",
+    "lower_kld_low_footprint_alternative": "IQ2_S",
+    "adjacent_low_footprint_alternative": "Q3_K_M"
   },
   "quality_protocol": {
     "name": "paired_BF16_vs_quant",
@@ -48,78 +43,59 @@
     "micro_batch_size": 512,
     "evaluation_executable": "llama-perplexity",
     "reference_logit_method": "--kl-divergence-base",
-    "candidate_comparison_method": "--kl-divergence",
+    "quantized_comparison_method": "--kl-divergence",
     "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
     "bf16_logits_reused_across_quant_tiers": true,
     "llama_cpp_revision": "42916d83f4a225e56709f873aa8050ac11f5b6a4"
   },
-  "artifacts": {
+  "frozen_artifacts": {
     "VeriLoop-E2-BF16.gguf": {
       "tier": "BF16",
-      "role": "canonical_reference_main",
       "bytes": 53808284000,
       "binary_gib": 50.11286959052086,
       "sha256": "11bf5defde1a256b7582bc34fd2c4a85a61615ed88e7a422dcd24d814ea6d35d",
       "tensor_count": 851,
-      "status": "VERIFIED"
+      "identity_gate": "PASS"
     },
     "VeriLoop-E2-Q8_0.gguf": {
       "tier": "Q8_0",
-      "role": "quantized_main",
       "bytes": 28595765600,
       "binary_gib": 26.631882041692734,
       "sha256": "6204a47274cfbc0c69c39877fb06615ce842bbab264eea77e2e0a5e3ae2fb8e8",
       "effective_bpw": 8.5,
-      "status": "VERIFIED"
+      "identity_gate": "PASS"
     },
     "VeriLoop-E2-Q6_K.gguf": {
       "tier": "Q6_K",
-      "role": "quantized_main",
       "bytes": 22082532096,
       "binary_gib": 20.56596064567566,
       "sha256": "15d8f856471c4853f6bf0036b2a517426c6cb30a0313cef58a7f9577fd26fe9e",
       "effective_bpw": 6.57,
-      "status": "VERIFIED"
+      "identity_gate": "PASS"
     },
     "VeriLoop-E2-Q5_K_M.gguf": {
       "tier": "Q5_K_M",
-      "role": "quantized_main",
       "bytes": 20363666176,
       "binary_gib": 18.965142011642456,
       "sha256": "f90ec14d7ec8f084a292413ae0e06483c87ab5e1422ddce51f2e3409d8853b61",
       "effective_bpw": 6.05,
-      "status": "VERIFIED"
-    },
-    "VeriLoop-E2-Q4_K_M.gguf": {
-      "tier": "Q4_K_M",
-      "role": "quantized_main",
-      "bytes": 19650634496,
-      "binary_gib": 18.301079511642456,
-      "sha256": null,
-      "effective_bpw": 5.84,
-      "status": "QUANTITATIVE_QC_PASS_RUNTIME_IDENTITY_PENDING"
+      "identity_gate": "PASS"
     },
     "VeriLoop-E2-Q3_K_M.gguf": {
       "tier": "Q3_K_M",
-      "role": "quantized_main",
       "bytes": 18066506496,
       "binary_gib": 16.825745344161987,
       "sha256": "c2e9539cfb85d87a99605b6914ed602c0fa750164c153ce5574710123aba06fe",
       "effective_bpw": 5.37,
-      "status": "VERIFIED"
+      "identity_gate": "PASS"
     },
     "VeriLoop-E2-IQ2_S.gguf": {
       "tier": "IQ2_S",
-      "role": "quantized_main_minimum_footprint_sweet_spot",
-      "quantization": "custom_mixed_IQ2_S_Q4_Q5_Q6",
-      "source": "VeriLoop-E2-BF16.gguf",
-      "direct_from_bf16": true,
-      "low_bit_requantization": false,
       "bytes": 18037261056,
       "decimal_gb": 18.037261056,
       "binary_gib": 16.798508405685425,
       "sha256": "0dd44d41319efa9f383b9a867b5fd2114335ccd0a653d1c3e7605d535edc601b",
-      "sha256_state": "FROZEN",
+      "effective_bpw": 5.36,
       "tensor_count": 851,
       "tensor_type_counts": {
         "F32": 353,
@@ -128,33 +104,19 @@
         "Q5_K": 429,
         "Q6_K": 2
       },
-      "effective_bpw": 5.36,
-      "quantized_size_mib_reported_by_quantizer": 17191.19,
-      "quantization_time_seconds": 168.08618,
-      "importance_matrix_sha256": "72d7c3fb0ee839cb591deaa10201fb8e826f1f6b58a6a889f897b02f6381748b",
-      "structure_gate": "PASS",
-      "precision_policy_gate": "PASS",
-      "quantitative_quality_gate": "PASS",
+      "identity_gate": "PASS",
+      "quality_gate": "PASS",
       "engineering_reserve_gate": "PASS",
       "runtime_gate": "PASS",
-      "mtp_gate": "PASS",
-      "remote_exact_size_gate": "PASS",
-      "remote_full_sha256_gate": "PASS",
-      "remote_readability_gate": "PASS",
-      "status": "VERIFIED"
+      "mtp_gate": "PASS"
     },
     "VeriLoop-E2-IQ1_M.gguf": {
       "tier": "IQ1_M",
-      "role": "ultra_low_footprint_frontier_candidate",
-      "quantization": "custom_mixed_IQ1_M_IQ2_S_Q4_Q5_Q6",
-      "source": "VeriLoop-E2-BF16.gguf",
-      "direct_from_bf16": true,
-      "low_bit_requantization": false,
       "bytes": 18028208896,
       "decimal_gb": 18.028208896,
       "binary_gib": 16.790077924728394,
       "sha256": "e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b",
-      "sha256_state": "FROZEN_LOCAL",
+      "effective_bpw": 5.36,
       "tensor_count": 851,
       "tensor_type_counts": {
         "F32": 353,
@@ -164,23 +126,28 @@
         "Q5_K": 429,
         "Q6_K": 2
       },
-      "effective_bpw": 5.36,
-      "quantized_size_mib_reported_by_quantizer_v2": 17182.55,
-      "structure_gate_v2": "PASS",
-      "quantitative_quality_gate_v2": "PASS",
-      "engineering_reserve_gate_v2": "PASS",
-      "runtime_gate": "PENDING",
-      "mtp_gate": "PENDING",
-      "remote_release_gate": "PENDING",
-      "status": "QUANTITATIVE_PASS_LOCAL_IDENTITY_FROZEN_RUNTIME_MTP_REMOTE_PENDING",
+      "quantized_size_mib_reported_by_quantizer": 17182.55,
       "quantization_time_seconds": 167.77645,
       "end_to_end_quantization_seconds": 168.2609,
-      "local_identity_gate": "PASS"
+      "identity_gate": "PASS",
+      "structure_gate": "PASS",
+      "quality_gate": "PASS",
+      "engineering_reserve_gate": "PASS",
+      "runtime_gate": "PASS",
+      "mtp_gate": "PASS",
+      "release_readiness_gate": "PASS"
+    }
+  },
+  "measured_tiers": {
+    "Q4_K_M": {
+      "bytes": 19650634496,
+      "binary_gib": 18.301079511642456,
+      "effective_bpw": 5.84,
+      "quantitative_quality_gate": "PASS"
     }
   },
   "quality_results": {
     "BF16": {
-      "role": "reference",
       "mean_ppl": 4.840423,
       "ppl_uncertainty": 0.119931
     },
@@ -189,15 +156,12 @@
       "ppl_uncertainty": 0.120062,
       "ppl_ratio": 1.000643,
       "ppl_ratio_uncertainty": 0.000754,
-      "relative_ppl_change_percent_observed": 0.0643,
       "mean_kld": 0.002176,
       "mean_kld_uncertainty": 0.000668,
       "same_top_p_percent": 98.815,
       "same_top_p_uncertainty_percent": 0.12,
       "rms_delta_p_percent": 1.251,
-      "rms_delta_p_uncertainty_percent": 0.137,
       "log_ppl_correlation_percent": 99.95,
-      "size_reduction_vs_bf16_percent": 46.856202290338786,
       "quantitative_quality_gate": "PASS"
     },
     "Q6_K": {
@@ -205,15 +169,12 @@
       "ppl_uncertainty": 0.119694,
       "ppl_ratio": 0.999605,
       "ppl_ratio_uncertainty": 0.001222,
-      "relative_ppl_change_percent_observed": -0.0395,
       "mean_kld": 0.004409,
       "mean_kld_uncertainty": 0.000953,
       "same_top_p_percent": 98.204,
       "same_top_p_uncertainty_percent": 0.147,
       "rms_delta_p_percent": 1.929,
-      "rms_delta_p_uncertainty_percent": 0.201,
       "log_ppl_correlation_percent": 99.88,
-      "size_reduction_vs_bf16_percent": 58.960720442227824,
       "quantitative_quality_gate": "PASS"
     },
     "Q5_K_M": {
@@ -221,15 +182,12 @@
       "ppl_uncertainty": 0.12063,
       "ppl_ratio": 1.00445,
       "ppl_ratio_uncertainty": 0.001361,
-      "relative_ppl_change_percent_observed": 0.445,
       "mean_kld": 0.006919,
       "mean_kld_uncertainty": 0.000945,
       "same_top_p_percent": 97.251,
       "same_top_p_uncertainty_percent": 0.181,
       "rms_delta_p_percent": 2.49,
-      "rms_delta_p_uncertainty_percent": 0.211,
       "log_ppl_correlation_percent": 99.85,
-      "size_reduction_vs_bf16_percent": 62.1551466387592,
       "quantitative_quality_gate": "PASS"
     },
     "Q4_K_M": {
@@ -237,15 +195,12 @@
       "ppl_uncertainty": 0.12072,
       "ppl_ratio": 1.004821,
       "ppl_ratio_uncertainty": 0.001722,
-      "relative_ppl_change_percent_observed": 0.4821,
       "mean_kld": 0.0097,
       "mean_kld_uncertainty": 0.00111,
       "same_top_p_percent": 96.786,
       "same_top_p_uncertainty_percent": 0.195,
       "rms_delta_p_percent": 2.843,
-      "rms_delta_p_uncertainty_percent": 0.155,
       "log_ppl_correlation_percent": 99.76,
-      "size_reduction_vs_bf16_percent": 63.48028029290062,
       "quantitative_quality_gate": "PASS"
     },
     "Q3_K_M": {
@@ -253,15 +208,12 @@
       "ppl_uncertainty": 0.120606,
       "ppl_ratio": 1.00409,
       "ppl_ratio_uncertainty": 0.002181,
-      "relative_ppl_change_percent_observed": 0.409,
       "mean_kld": 0.014349,
       "mean_kld_uncertainty": 0.001985,
       "same_top_p_percent": 95.919,
       "same_top_p_uncertainty_percent": 0.219,
       "rms_delta_p_percent": 3.515,
-      "rms_delta_p_uncertainty_percent": 0.228,
       "log_ppl_correlation_percent": 99.62,
-      "size_reduction_vs_bf16_percent": 66.4243028155293,
       "quantitative_quality_gate": "PASS"
     },
     "IQ2_S": {
@@ -269,39 +221,22 @@
       "ppl_uncertainty": 0.120482,
       "ppl_ratio": 1.003457,
       "ppl_ratio_uncertainty": 0.0021,
-      "relative_ppl_change_percent_observed": 0.3457,
       "mean_kld": 0.014023,
       "mean_kld_uncertainty": 0.001408,
       "same_top_p_percent": 95.516,
       "same_top_p_uncertainty_percent": 0.229,
       "rms_delta_p_percent": 3.679,
-      "rms_delta_p_uncertainty_percent": 0.231,
       "log_ppl_correlation_percent": 99.64,
-      "size_reduction_vs_bf16_percent": 66.47865400056244,
       "quantitative_quality_gate": "PASS",
-      "absolute_ppl_delta": 0.016736,
-      "absolute_ppl_delta_uncertainty": 0.01016,
-      "median_kld": 0.003619,
-      "kld_p90": 0.022931,
-      "kld_p95": 0.03865,
-      "kld_p99": 0.129293,
-      "kld_p999": 0.971304,
-      "max_kld": 8.040164,
-      "mean_delta_p_percent": -0.067,
-      "mean_delta_p_uncertainty_percent": 0.041,
-      "size_reduction_vs_q3_percent": 0.16187656427364672,
-      "size_reduction_vs_q5_percent": 11.424294131976248,
       "engineering_reserve_gate": "PASS"
     },
     "IQ1_M": {
-      "measurement_source": "V2_same_tensor_policy_as_V3",
       "mean_ppl": 4.85587,
       "ppl_uncertainty": 0.12048,
       "absolute_ppl_delta": 0.015446,
       "absolute_ppl_delta_uncertainty": 0.010522,
       "ppl_ratio": 1.003191,
       "ppl_ratio_uncertainty": 0.002175,
-      "relative_ppl_change_percent_observed": 0.3191,
       "mean_kld": 0.014357,
       "mean_kld_uncertainty": 0.001317,
       "median_kld": 0.003778,
@@ -324,318 +259,47 @@
       "engineering_reserve_gate": "PASS"
     }
   },
-  "release_gates": {
-    "IQ2_S": {
-      "ppl_ratio": {
-        "hard_threshold_max": 1.015,
-        "reserve_threshold_max": 1.0135,
-        "measured": 1.003457,
-        "hard_result": "PASS",
-        "reserve_result": "PASS"
-      },
-      "mean_kld": {
-        "hard_threshold_max": 0.015,
-        "reserve_threshold_max": 0.0145,
-        "measured": 0.014023,
-        "hard_result": "PASS",
-        "reserve_result": "PASS"
-      },
-      "same_top_p_percent": {
-        "hard_threshold_min": 95.0,
-        "reserve_threshold_min": 95.5,
-        "measured": 95.516,
-        "hard_result": "PASS",
-        "reserve_result": "PASS"
-      },
-      "mixed_precision_policy": {
-        "expected_counts": {
-          "F32": 353,
-          "IQ2_S": 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_draft_model_load_evidence": "PASS",
-      "mtp_draft_tokens_generated_gt_zero": "PASS",
-      "mtp_draft_tokens_accepted_gt_zero": "PASS",
-      "output_identity_audit": "DIVERGED_ADVISORY_NON_BLOCKING",
-      "final_release_identity": "FROZEN",
-      "remote_release_verification": "PASS",
-      "overall_release_quality_gate": "PASS"
-    },
-    "IQ1_M": {
-      "measurement_source": "V2_same_tensor_policy_as_V3",
-      "ppl_ratio": {
-        "hard_threshold_max": 1.015,
-        "reserve_threshold_max": 1.0135,
-        "measured": 1.003191,
-        "hard_result": "PASS",
-        "reserve_result": "PASS"
-      },
-      "mean_kld": {
-        "hard_threshold_max": 0.015,
-        "reserve_threshold_max": 0.0145,
-        "measured": 0.014357,
-        "hard_result": "PASS",
-        "reserve_result": "PASS"
-      },
-      "same_top_p_percent": {
-        "hard_threshold_min": 95.0,
-        "reserve_threshold_min": 95.5,
-        "measured": 95.87,
-        "hard_result": "PASS",
-        "reserve_result": "PASS"
-      },
-      "mixed_precision_policy": {
-        "expected_counts": {
-          "F32": 353,
-          "IQ1_M": 1,
-          "IQ2_S": 2,
-          "Q4_K": 64,
-          "Q5_K": 429,
-          "Q6_K": 2
-        },
-        "mismatch_count": 0,
-        "result": "PASS"
-      },
-      "quantitative_quality_gate": "PASS",
-      "engineering_reserve_gate": "PASS",
-      "obsolete_v2_point_estimate_dominance_gate": "REMOVED_IN_V3",
-      "stock_runtime_gate": "PENDING",
-      "mtp_gate": "PENDING",
-      "remote_release_gate": "PENDING",
-      "overall_release_quality_gate": "PENDING_RUNTIME_MTP_REMOTE",
-      "final_local_identity_gate": "PASS"
-    }
-  },
   "runtime_validation": {
     "IQ2_S": {
       "stock_llama_cpp_only": true,
-      "veriloop_harness_used": false,
       "llama_cpp_revision": "42916d83f4a225e56709f873aa8050ac11f5b6a4",
-      "main_model_load": "PASS",
-      "main_model_generation": "PASS",
-      "mtp_interface": "PASS",
-      "mtp_load": "PASS",
-      "mtp_generation": "PASS",
-      "mtp_draft_model_load_evidence": "PASS",
+      "main_runtime": "PASS",
+      "mtp_runtime": "PASS",
+      "mtp_engagement": "PASS",
       "draft_tokens_generated": 331,
       "draft_tokens_accepted": 116,
-      "draft_acceptance_rate": 0.35045,
-      "mean_draft_length": 3.76,
-      "output_identity_audit": "DIVERGED",
-      "output_identity_is_release_blocker": false,
-      "mtp_bytes": 2008056096,
-      "mtp_sha256": "0e7f2dfe254f3a5d195d105832411acc1007fd7bb9ac980fe9aae4a1f290d671",
-      "runtime_mtp_gate": "PASS"
+      "draft_acceptance_rate": 0.35045
     },
     "IQ1_M": {
-      "state": "PENDING_V3",
       "stock_llama_cpp_only": true,
-      "required": [
-        "main_model_load",
-        "main_model_generation",
-        "mtp_load",
-        "mtp_generation",
-        "mtp_draft_engagement"
-      ],
-      "claim": "NO_RUNTIME_OR_MTP_PASS_CLAIM_BEFORE_V3_EVIDENCE"
-    }
-  },
-  "remote_repository_verification": {
-    "IQ2_S": {
-      "state": "PASS",
-      "remote_path": "VeriLoop-E2-IQ2_S.gguf",
-      "main_remote_bytes": 18037261056,
-      "main_remote_lfs_sha256": "0dd44d41319efa9f383b9a867b5fd2114335ccd0a653d1c3e7605d535edc601b",
-      "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"
-    },
-    "IQ1_M": {
-      "state": "PENDING_V3",
-      "remote_path": "VeriLoop-E2-IQ1_M.gguf",
-      "required": [
-        "remote_exact_size",
-        "remote_full_sha256",
-        "remote_readability",
-        "remote_head_identity",
-        "remote_tail_identity"
-      ]
-    }
-  },
-  "throughput": {
-    "IQ2_S": {
-      "paired_speed_protocol_run": false,
-      "speedup_claimed": false,
-      "mtp_draft_tokens_generated": 331,
-      "mtp_draft_tokens_accepted": 116,
-      "mtp_draft_acceptance_rate": 0.35045,
-      "note": "Draft acceptance is reported descriptively; no universal speedup is claimed without a frozen paired throughput protocol."
-    },
-    "IQ1_M": {
-      "paired_speed_protocol_run": false,
-      "speedup_claimed": false,
-      "mtp_acceptance_measured": false,
-      "note": "No IQ1_M throughput or MTP acceptance claim before final V3 runtime validation."
+      "llama_cpp_revision": "42916d83f4a225e56709f873aa8050ac11f5b6a4",
+      "main_http_status": 200,
+      "main_runtime": "PASS",
+      "mtp_http_status": 200,
+      "mtp_runtime": "PASS",
+      "mtp_engagement": "PASS",
+      "draft_tokens_generated": 104,
+      "draft_tokens_accepted": 76,
+      "draft_acceptance_rate": 0.73076923,
+      "acceptance_rate_consistency_gate": "PASS",
+      "output_identity_audit": "IDENTICAL",
+      "runtime_mtp_gate": "PASS"
     }
   },
-  "release_matrix": {
-    "BF16": {
-      "state": "released",
-      "role": "canonical_reference"
-    },
-    "Q8_0": {
-      "state": "released",
-      "role": "high_fidelity_local"
-    },
-    "Q6_K": {
-      "state": "released",
-      "role": "overall_quality_efficiency_sweet_spot",
-      "quality_record": "PASS"
-    },
-    "Q5_K_M": {
-      "state": "released",
-      "role": "memory_quality_sweet_spot",
-      "quality_record": "PASS"
-    },
-    "Q4_K_M": {
-      "state": "release_candidate",
-      "role": "intermediate_mixed_precision_candidate",
-      "quality_record": "PASS",
-      "runtime_record": "PENDING"
-    },
-    "Q3_K_M": {
-      "state": "verified_release",
-      "role": "adjacent_low_footprint_alternative",
-      "quality_record": "PASS",
-      "runtime_record": "PASS",
-      "remote_release": "PASS"
-    },
-    "IQ2_S": {
-      "state": "verified_release",
-      "role": "minimum_footprint_low_footprint_sweet_spot",
-      "quality_record": "PASS",
-      "engineering_reserve_record": "PASS",
-      "runtime_record": "PASS",
-      "mtp_record": "PASS",
-      "release_identity": "FROZEN",
-      "remote_release": "PASS",
-      "overall_release": "PASS"
-    },
+  "release_validation": {
     "IQ1_M": {
-      "state": "prerelease_quantitative_pass_local_identity_frozen",
-      "role": "ultra_low_footprint_frontier_candidate",
-      "quality_record": "PASS_V2_SAME_POLICY",
-      "engineering_reserve_record": "PASS_V2_SAME_POLICY",
-      "runtime_record": "PENDING",
-      "mtp_record": "PENDING",
-      "release_identity": "FROZEN_LOCAL",
-      "remote_release": "PENDING",
-      "overall_release": "PENDING_RUNTIME_MTP_REMOTE"
-    }
-  },
-  "current_release_state": {
-    "BF16": "RELEASED_REFERENCE",
-    "Q8_0": "RELEASED_VERIFIED",
-    "Q6_K": "RELEASED_VERIFIED_OVERALL_SWEET_SPOT",
-    "Q5_K_M": "RELEASED_VERIFIED_MEMORY_QUALITY_SWEET_SPOT",
-    "Q4_K_M": "QUANTITATIVE_QC_PASS_RUNTIME_AND_SHA_PENDING",
-    "Q3_K_M": "VERIFIED_RELEASE_LOW_FOOTPRINT_ALTERNATIVE",
-    "IQ2_S": "VERIFIED_RELEASE_CURRENT_LOW_FOOTPRINT_SWEET_SPOT",
-    "IQ1_M": "PRERELEASE_QUANTITATIVE_PASS_LOCAL_IDENTITY_FROZEN_RUNTIME_MTP_REMOTE_PENDING"
-  },
-  "deployment_recommendation": {
-    "overall_sweet_spot": "Q6_K",
-    "memory_quality_sweet_spot": "Q5_K_M",
-    "current_released_low_footprint_sweet_spot": "IQ2_S",
-    "ultra_low_footprint_frontier_candidate": "IQ1_M",
-    "q1_post_validation_promotion_rule": "If the identical V3 policy reproduces the V2 hard/reserve fidelity record and passes immutable identity, stock runtime, MTP, and remote verification, promote IQ1_M to minimum-footprint sweet spot; retain IQ2_S as the lower-KLD low-footprint alternative.",
-    "basis": "IQ1_M is 16.790078 GiB versus IQ2_S at 16.798508 GiB. IQ1_M has a lower PPL ratio (1.003191 vs 1.003457) and higher Same top-p (95.870% vs 95.516%), while IQ2_S has lower Mean KLD (0.014023 vs 0.014357) and slightly lower RMS delta-p. The point-estimate differences are small relative to reported uncertainty, so the two are treated as neighboring Pareto points rather than categorical quality winners. IQ2_S retains the current released low-footprint sweet-spot designation because its immutable identity, stock runtime, MTP, and remote release verification are already closed."
-  },
-  "q2_development": {
-    "v1_rejected": {
-      "target": "four_low_sensitivity_ffn_up_tensors_to_IQ2_S",
-      "bytes": 18049098496,
-      "binary_gib": 16.809533,
-      "ppl_ratio": 1.008995,
-      "mean_kld": 0.014951,
-      "same_top_p_percent": 95.809,
-      "hard_gate": "PASS",
-      "engineering_reserve_gate": "FAIL",
-      "failure_axis": "MEAN_KLD",
-      "large_candidate_retention": "DELETE_FAILED_ARTIFACT_PRESERVE_REPORTS"
-    },
-    "v2_final": {
-      "target": "three_q3_spine_ffn_down_tensors_to_IQ2_S",
-      "exact_tensors": [
-        "blk.0.ffn_down.weight",
-        "blk.1.ffn_down.weight",
-        "blk.3.ffn_down.weight"
-      ],
-      "bytes": 18037261056,
-      "binary_gib": 16.798508405685425,
-      "ppl_ratio": 1.003457,
-      "mean_kld": 0.014023,
-      "same_top_p_percent": 95.516,
-      "hard_gate": "PASS",
-      "engineering_reserve_gate": "PASS"
-    }
-  },
-  "q1_development": {
-    "v1_rejected": {
-      "policy": "blk.1_and_blk.0_ffn_down_to_IQ1_M_blk.3_restored_Q3_K",
-      "bytes": 18028905216,
-      "binary_gib": 16.790726,
-      "ppl_ratio": 1.006006,
-      "mean_kld": 0.016349,
-      "same_top_p_percent": 95.736,
-      "hard_quality_gate": "FAIL",
-      "engineering_reserve_gate": "FAIL",
-      "candidate_file_deleted": true
-    },
-    "v2_measured": {
-      "policy": "single_factor_blk.1_ffn_down_IQ2_S_to_IQ1_M_everything_else_identical_to_Q2",
-      "bytes": 18028208896,
-      "binary_gib": 16.790077924728394,
-      "ppl_ratio": 1.003191,
-      "mean_kld": 0.014357,
-      "same_top_p_percent": 95.87,
-      "hard_quality_gate": "PASS",
-      "engineering_reserve_gate": "PASS",
-      "obsolete_point_estimate_dominance_gate": "FAIL_MEAN_KLD_VS_Q2",
-      "candidate_file_deleted": true,
-      "interpretation": "Quantitative release quality passed; file deletion was caused only by an experimental cross-tier dominance rule removed in V3."
-    },
-    "accepted_rebuild": {
-      "quantization_policy_change_vs_v2": false,
-      "hard_gate_change_vs_v2": false,
-      "reserve_gate_change_vs_v2": false,
-      "bytes": 18028208896,
-      "binary_gib": 16.790078,
-      "ppl_ratio": 1.003191,
-      "mean_kld": 0.014357,
-      "same_top_p_percent": 95.87,
+      "identity_gate": "PASS",
+      "structure_gate": "PASS",
       "hard_quality_gate": "PASS",
       "engineering_reserve_gate": "PASS",
-      "sweetspot_promotion_gate": "PASS",
-      "candidate_acceptance_gate": "PASS",
-      "sha256": "e4d395806994cdbcfecf7711e22456af7a663b3ea37c89ef15d5bab4571cf68b",
-      "local_identity": "FROZEN",
-      "stock_runtime": "PENDING",
-      "mtp": "PENDING",
-      "remote_release": "PENDING"
+      "stock_runtime_gate": "PASS",
+      "mtp_runtime_gate": "PASS",
+      "mtp_engagement_gate": "PASS",
+      "release_readiness_gate": "PASS"
     }
+  },
+  "throughput_claims": {
+    "paired_bf16_speedup_claimed": false,
+    "note": "MTP acceptance is prompt-dependent; no universal throughput speedup is claimed without a frozen paired throughput protocol."
   }
 }