Add bilingual 4L and MALINT benchmark evidence

.gitattributes

@@ -1,3 +1,6 @@
 *.safetensors filter=lfs diff=lfs merge=lfs -text
 benchmark/figures/02_per_label_transfer_metrics.png filter=lfs diff=lfs merge=lfs -text
 benchmark/figures/04_compression_ladder.png filter=lfs diff=lfs merge=lfs -text
+benchmark/figures/cross_task/05_why_4l_edge_tradeoff.png filter=lfs diff=lfs merge=lfs -text
+benchmark/figures/cross_task/06_malint_4l_quality.png filter=lfs diff=lfs merge=lfs -text
+benchmark/figures/cross_task/07_malint_quality_resource.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -10,9 +10,11 @@ tags:
   - qwen3.5
   - classification-backbone
   - text-classification
+  - multilabel-classification
   - knowledge-distillation
   - model-compression
   - edge-ai
+  - malint
 ---
 
 # Qwen3.5 Standalone 4L Classification Base
@@ -31,6 +33,26 @@ The repository follows one integrated 1-1-1 layout:
    classifier preserved as a separate task-specific lineage.
 3. **`distillation/`:** reusable 24L→8L→6L→4L platform for user-owned data.
 
+## 한국어 요약
+
+이 저장소의 루트 모델은 Qwen3.5-0.8B의 text backbone을 **4개 layer로 축소한
+분류 전용 기반 모델**이다. LM head와 task label, 분류 head가 없으므로 채팅 모델이나
+즉시 사용할 수 있는 범용 분류기가 아니다. 사용자가 자신의 label과 데이터로 새 head를
+붙여 학습하는 출발점이다.
+
+4L을 선택한 근거는 한 숫자가 아니라 세 범주의 측정이다. 첫째, full-vocabulary
+task-free backbone은 24L의 752.39M parameter/1,435.11 MiB에서 4L의
+334.10M/637.24 MiB로 줄었다. 둘째, 동일 V128k/F16 UNO Q 깊이 비교에서 역사적
+4L proxy는 8L 대비 artifact `-27.74%`, PP128 latency `-49.84%`, peak RSS
+`-21.81%`, core-seconds `-48.96%`를 기록했다. 셋째, 별도 task 적응 결과에서
+4L endpoint가 propaganda와 MALINT 모두에서 실제 분류 가능성을 보였다.
+
+다만 이 근거는 “4L이 항상 8L/24L보다 정확하다”는 뜻이 아니다. UNO Q의 4L은
+historical early-path proxy이고, propaganda 24L/8L/4L endpoint는 계보와 vocabulary가
+달라 순수 depth ablation이 아니다. MALINT의 Qwen student 세 모델은 모두 4L이므로,
+MALINT는 **4L 내부 pooling/KD interface 개선**을 검증할 뿐 4L-vs-8L/24L 우월성을
+검증하지 않는다. 아래 표와 그림은 이 경계를 수치와 함께 명시한다.
+
 ## Model at a glance
 
 | Item | Root model |
@@ -44,7 +66,7 @@ The repository follows one integrated 1-1-1 layout:
 | Root BF16 weights | 668,198,976 bytes (637.24 MiB) |
 | Output | hidden states; no LM head, task head, labels, or thresholds |
 | Task-free KD data | 4,096 unlabeled WikiText-103 raw rows |
-| Downstream evidence | one English 14-label SemEval-derived transfer probe |
+| Downstream evidence | English 14-label propaganda transfer and 5-label MALINT adaptation studies under separate protocols |
 
 ## Load the headless base
 
@@ -107,6 +129,126 @@ The chart reports the actual unsharded safetensors parameter counts and BF16
 file sizes. It does not imply that downstream quality scales monotonically with
 layer count.
 
+## Why four layers? / 왜 4L인가?
+
+The choice is an **edge-oriented operating point**, not a universal optimum.
+The task-free full-vocabulary root has 334,096,704 parameters and a 637.24 MiB
+BF16 weight file, versus 752,393,024 parameters and 1,435.11 MiB for the 24L
+text teacher: a `55.59%` reduction in both parameter count and raw BF16 bytes.
+
+A separate controlled Arduino UNO Q chain held vocabulary (`128k`), precision
+(`F16`), runtime, and benchmark protocol fixed while moving 8L→6L→4L. Values
+below normalize the 8L arm to 100; lower is better.
+
+| Controlled UNO Q arm | Artifact | PP128 latency | Peak RSS | Core-seconds |
+|---|---:|---:|---:|---:|
+| 8L | 100.00 | 100.00 | 100.00 | 100.00 |
+| 6L | 85.59 | 76.37 | 87.51 | 76.17 |
+| Historical 4L proxy | **72.26** | **50.16** | **78.19** | **51.04** |
+
+Thus the matched 8L→4L route reduced artifact size by `27.74%`, PP128 latency
+by `49.84%`, peak RSS by `21.81%`, and core-seconds by `48.96%`. CPU occupancy
+remained near 3.6–3.7 saturated cores; the benefit was finishing sooner, not
+lower active occupancy. The 4L point is a historical early-path proxy, not the
+final common-head quality model.
+
+The earlier English 14-label propaganda benchmark supplies task evidence under
+one common five-epoch protocol. Its endpoints are operationally relevant but
+not a pure depth study because lineage and vocabulary differ.
+
+| Propaganda endpoint | Macro-F1 | Micro-F1 | Subset exact match | Test s/article | Peak CUDA allocated | Parameters |
+|---|---:|---:|---:|---:|---:|---:|
+| Qwen3.5 24L | 0.56426 ± 0.00626 | 0.62648 ± 0.01399 | 0.03030 ± 0.02777 | 0.58636 ± 0.00598 | 7,635,339,264 B | 752.41M |
+| Qwen3.5 C8Q 8L | 0.53669 ± 0.01726 | 0.62388 ± 0.01362 | 0.00606 ± 0.01050 | 0.20666 ± 0.00207 | 4,257,894,400 B | 420.33M |
+| Qwen3.5 final 4L | **0.58757 ± 0.00717** | **0.64798 ± 0.00187** | 0.00606 ± 0.01050 | **0.08544 ± 0.00060** | **2,143,655,936 B** | **210.90M** |
+
+These results show that a compact four-layer endpoint can be useful; they do
+not prove that reducing depth caused the quality difference. The headless root
+also differs from the specialized 4L endpoint: the root keeps the full 248,320
+token vocabulary and has 334.10M parameters.
+
+![Why four layers: controlled edge and propaganda evidence](benchmark/figures/cross_task/05_why_4l_edge_tradeoff.png)
+
+한국어로 요약하면, 4L 선택은 동일 조건 edge chain에서 확인된 latency·memory·artifact
+감소와 별도 task에서 확인된 분류 가능성을 함께 본 결과다. 그러나 propaganda endpoint의
+계보/vocabulary 차이 때문에 4L의 품질 우위를 depth 하나의 인과 효과로 해석하면 안 된다.
+
+## MALINT adaptation evidence / MALINT 적응 근거
+
+MALINT is an English five-label malicious-intent multilabel task covering
+changing political views (CPV), promoting social stereotypes/antagonisms
+(PSSA), undermining international organizations/alliances (UIOA), promoting
+anti-scientific views (PASV), and undermining public institutions (UCPI). No
+MALINT row is redistributed here.
+
+The strongest changed-model evidence is a frozen **3 repeats × 3 folds** matched
+study over the combined 1,100-row development corpus. Each repeat reconstructs
+exact out-of-fold predictions for all 1,100 rows at threshold `0.30`. There was
+no fold checkpoint selection, threshold tuning, or official held-out access.
+
+| Arm | Weighted-F1 | Macro-F1 | Micro-F1 | Subset exact match | Label-wise accuracy |
+|---|---:|---:|---:|---:|---:|
+| Rightmost token + hard labels (4L) | 0.52218 ± 0.01497 | 0.51096 ± 0.01628 | 0.53203 ± 0.01076 | 0.64030 ± 0.00367 | 0.87467 ± 0.00121 |
+| Mean pooling + hard labels (4L) | 0.61180 ± 0.00886 | 0.60551 ± 0.00753 | 0.61769 ± 0.00746 | **0.66970 ± 0.00517** | **0.88764 ± 0.00175** |
+| **Mean pooling + KD (4L)** | **0.63465 ± 0.01197** | **0.62818 ± 0.01423** | **0.63896 ± 0.00985** | 0.65576 ± 0.00319 | 0.87576 ± 0.00204 |
+| Fresh RoBERTa teacher | 0.60468 ± 0.01580 | 0.59557 ± 0.01637 | 0.60462 ± 0.01777 | 0.62000 ± 0.00596 | 0.85079 ± 0.01087 |
+
+Values are mean ± sample SD across three repeated OOF reconstructions. Subset
+exact match requires all five labels for a row to match. Label-wise accuracy is
+the mean over all row-label decisions and is imbalance-sensitive.
+
+Mean+KD improved weighted F1 over rightmost-hard by `+0.11247`, with positive
+deltas in `9/9` folds and corrected interval `[+0.07160, +0.15925]`. Mean
+pooling alone improved it by `+0.08962` in `9/9` folds. The incremental KD
+contrast was `+0.02285` in `8/9` folds, but its interval
+`[-0.00155, +0.05059]` crosses zero. Therefore the strongest isolated evidence
+is for masked mean pooling; independent KD and teacher superiority are not
+established.
+
+![MALINT 4L repeated quality metrics](benchmark/figures/cross_task/06_malint_4l_quality.png)
+
+All three Qwen student arms above are four-layer models. This is evidence that
+a 4L backbone can support a materially better document interface under this
+internal protocol. It is **not** a MALINT 4L-vs-8L/24L depth ablation. It is
+also post-selection development evidence, not untouched confirmation or SOTA.
+The older validation-selected rightmost 4L has an official fixed-threshold test
+weighted F1 of `0.53749` with row-bootstrap interval `[0.49102, 0.58403]`; the
+changed Mean+KD artifact did not reopen that test, so the two protocols must not
+be compared as if they were one held-out leaderboard.
+
+### Quality–resource trade-off / 품질–자원 절충
+
+A separate label-free same-input resource run used an RTX 5070 Ti, BF16,
+sequence length 256, fresh processes, three counterbalanced rounds per batch,
+and CUDA-resident token IDs/masks. Timing is model forward only.
+
+| Batch | Rightmost ms/example | Mean+KD ms/example | Latency change | Rightmost peak allocated | Mean+KD peak allocated | Peak change |
+|---:|---:|---:|---:|---:|---:|---:|
+| 1 | 18.47463 | 19.31878 | **+4.569%** | 1,373,150,208 B | 1,374,199,296 B | **+0.076%** |
+| 4 | 4.58975 | 4.67101 | **+1.770%** | 1,454,388,736 B | 1,455,437,824 B | **+0.072%** |
+
+Mean+KD adds exactly five parameters. It is **not faster**; the supported claim
+is higher internal F1 at a small measured forward-latency and peak-allocation
+cost. Ambient `nvidia-smi` utilization is not reported as model occupancy.
+
+![MALINT 4L quality and resource trade-off](benchmark/figures/cross_task/07_malint_quality_resource.png)
+
+한국어 해석: MALINT에서 확인된 핵심은 “4L보다 더 깊은 모델을 이겼다”가 아니라,
+**깊이를 4L로 고정한 상태에서 문서 전체 mean pooling interface가 rightmost-token
+interface보다 반복 내부평가 F1을 높였다**는 것이다. Mean+KD는 batch 1/4에서 각각
+`4.569%`/`1.770%` 느렸고 peak allocation도 소폭 증가했다. 따라서 빠른 모델이
+아니라, 거의 같은 크기·memory에서 제한된 latency 비용으로 내부 품질을 높인 적응
+결과로 설명해야 한다.
+
+### Cross-task boundary / task 간 비교 경계
+
+Propaganda and MALINT have different labels, split histories, thresholds,
+metrics, and evaluation protocols. Their F1 values are not pooled or ranked
+against each other. Together they support only a bounded statement: this 4L
+platform has been adapted successfully to two English multilabel settings,
+while generalization, multilingual quality, pure depth superiority, and SOTA
+remain unproven.
+
 ## Downstream benchmark contract
 
 Because the root is headless, it cannot be scored as a classifier without
@@ -245,6 +387,10 @@ hardware.
   including article-level probability vectors
 - [`benchmark/figures/`](benchmark/figures/): Matplotlib PNG/PDF figures,
   per-figure source CSVs, and a hash manifest
+- [`benchmark/figures/cross_task/manifest.json`](benchmark/figures/cross_task/manifest.json):
+  immutable input/output hashes for figures 05–07
+- [`benchmark/figures/cross_task/source_summary.json`](benchmark/figures/cross_task/source_summary.json):
+  aggregate-only source values and claim boundaries for figures 05–07
 
 ## Multilingual scope
 
@@ -255,12 +401,13 @@ multilingual upstream benchmarks such as MMMLU, MMLU-ProX, NOVA-63, INCLUDE,
 Global PIQA, PolyMATH, WMT24++, and MAXIFE. This is why this repository carries
 the Hugging Face `multilingual` language and model tag.
 
-However, this four-layer root was distilled on English WikiText and its only
-downstream transfer probe is English SemEval-derived data. **Tokenizer coverage
-and inherited architecture support do not establish retained classification
-quality in all 201 languages.** No multilingual downstream score is reported
-for this root release. Users should fine-tune and evaluate on each intended
-language and domain before making performance claims.
+However, this four-layer root was distilled on English WikiText and its current
+downstream evidence is limited to two English multilabel settings:
+SemEval-derived propaganda and MALINT. **Tokenizer coverage and inherited
+architecture support do not establish retained classification quality in all
+201 languages.** No multilingual downstream score is reported for this root
+release. Users should fine-tune and evaluate on each intended language and
+domain before making performance claims.
 
 ## Separate SemEval model
 
@@ -281,12 +428,19 @@ tokenizer. Its complete reports remain under
 - A classification head and labeled training are required before root-model
   predictions have task meaning.
 - Task-free KD used only 4,096 rows from one English WikiText shard.
-- Only one downstream task and one hardware/software setting were measured.
+- Only two English multilabel downstream settings and a limited set of
+  hardware/software conditions were measured.
+- Cross-task evidence now covers two English multilabel settings, but they use
+  different protocols and cannot be pooled into one score.
 - The 55-article test split was previously opened; results are exploratory.
 - Per-label support ranges from 7 to 45 articles and rare-label estimates are
   unstable.
 - Three seeds describe run-to-run variation; they are not a confidence
   interval.
+- The changed MALINT model has post-selection repeated-development evidence but
+  no newly opened official held-out score; MALINT is not a depth ablation.
+- MALINT forward latency excludes tokenization, loading, disk, host-to-device
+  transfer, CPU, Mac, Arduino, and end-to-end application costs.
 - Multilingual input support has not been validated as multilingual downstream
   classification quality for this four-layer root.
 - Default layer maps are documented structural choices, not universal optima.
@@ -302,6 +456,12 @@ and data terms. Qwen3.5 is attributed to Qwen. Transformers, PyTorch,
 Hugging Face Hub, WikiText, and SemEval remain the work of their respective
 authors. No WikiText or SemEval source records are redistributed.
 
+No MALINT row is redistributed. The MALINT task/data attribution follows
+Modzelewski et al., *MALicious INTent Dataset and Inoculating LLMs for Enhanced
+Disinformation Detection*, EACL 2026
+([ACL Anthology](https://aclanthology.org/2026.eacl-long.144/)); the official
+dataset repository states CC BY 4.0 for the dataset.
+
 The WikiText page currently has a license wording discrepancy: metadata lists
 CC BY-SA 3.0 and GFDL, while prose says CC BY-SA 4.0. Users should inspect
 [`Salesforce/wikitext`](https://huggingface.co/datasets/Salesforce/wikitext)

benchmark/figures/README.md

@@ -1,6 +1,8 @@
 # Matplotlib benchmark figures
 
-All figures are deterministically rendered from the frozen v2 benchmark bundle.
+Figures 01–04 are deterministically rendered from the frozen v2 task-agnostic
+transfer bundle. Figures 05–07 are recomputed from frozen propaganda, UNO Q,
+and MALINT aggregate records.
 PNG files are intended for the Hugging Face model card; PDF files are vector
 exports for reports. Each plot has a compact source CSV beside it.
 
@@ -8,6 +10,25 @@ exports for reports. Each plot has a compact source CSV beside it.
 2. Per-label precision/recall/F1 for the task-agnostic base.
 3. Macro-F1 against article inference time and peak CUDA allocation.
 4. Parameters and BF16 size across the 24L→8L→6L→4L task-free ladder.
+5. Controlled UNO Q depth resources plus descriptive propaganda 24L/8L/4L
+   quality and resources.
+6. MALINT repeated-OOF weighted/macro/micro F1, subset exact match, and
+   label-wise accuracy.
+7. MALINT internal weighted-F1 gain beside measured CUDA-forward latency and
+   peak-allocation changes.
 
 The 55-article test split was previously opened. Error bars are not confidence
 intervals. The specialized checkpoint is a separate reference lineage.
+
+`manifest.json` and `source_summary.json` cover figures 01–04.
+`cross_task/manifest.json` and `cross_task/source_summary.json` cover figures
+05–07 and bind every frozen input by SHA-256. The latter bundle is aggregate
+only and contains no raw text, IDs, row labels, logits, probabilities, or
+predictions.
+
+## 한국어
+
+01–04는 기존 task-free base/propaganda transfer 결과이고, 05–07은 4L 선택 근거와
+MALINT 내부 반복평가를 추가한다. MALINT의 Qwen student는 모두 4L이므로 06은
+4L-vs-8L/24L depth 비교가 아니다. 07은 Mean+KD가 더 빠르지 않으며 제한된 latency와
+peak CUDA allocation 비용이 있음을 함께 표시한다.

benchmark/figures/cross_task/05_why_4l_edge_tradeoff.csv

@@ -0,0 +1,34 @@
+scope,model,layers,metric,value,sample_sd,unit,normalized_reference,normalized_value,n,claim_boundary
+uno_q_controlled_v128k_f16_depth_chain,8L,8,artifact_mib,100.0,,percent_of_8l,8L=100,100.0,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,6L,6,artifact_mib,85.59176355424236,,percent_of_8l,8L=100,85.59176355424236,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,4L proxy,4,artifact_mib,72.2607217304372,,percent_of_8l,8L=100,72.2607217304372,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,8L,8,latency_pp128_ms,100.0,,percent_of_8l,8L=100,100.0,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,6L,6,latency_pp128_ms,76.36565126547553,,percent_of_8l,8L=100,76.36565126547553,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,4L proxy,4,latency_pp128_ms,50.16216836129361,,percent_of_8l,8L=100,50.16216836129361,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,8L,8,peak_rss_mib,100.0,,percent_of_8l,8L=100,100.0,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,6L,6,peak_rss_mib,87.50983221571457,,percent_of_8l,8L=100,87.50983221571457,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,4L proxy,4,peak_rss_mib,78.19200407435702,,percent_of_8l,8L=100,78.19200407435702,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,8L,8,mean_core_seconds_per_round,100.0,,percent_of_8l,8L=100,100.0,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,6L,6,mean_core_seconds_per_round,76.16702179915283,,percent_of_8l,8L=100,76.16702179915283,3,4L is the historical early-path proxy
+uno_q_controlled_v128k_f16_depth_chain,4L proxy,4,mean_core_seconds_per_round,51.03561371028788,,percent_of_8l,8L=100,51.03561371028788,3,4L is the historical early-path proxy
+propaganda_operational_endpoint_quality,24L,24,macro_f1,0.5642638295960175,0.006258972649035587,score,,,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_quality,24L,24,micro_f1,0.6264844890231888,0.01398932248989461,score,,,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_quality,24L,24,subset_exact_match,0.0303030303030303,0.027773186030035393,score,,,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,24L,24,test_seconds_per_article,0.586362602903109,0.005979023788731576,seconds_per_article,24L=100,100.0,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,24L,24,peak_cuda_allocated_bytes,7635339264.0,0.0,bytes,24L=100,100.0,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,24L,24,parameters,752407360,,count,24L=100,100.0,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,24L,24,total_wall_seconds,12842.063612925334,94.80225434811064,seconds,24L=100,99.99999999999999,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_quality,8L,8,macro_f1,0.5366873350531403,0.017264532146217884,score,,,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_quality,8L,8,micro_f1,0.6238830315003392,0.013623163834245318,score,,,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_quality,8L,8,subset_exact_match,0.006060606060606061,0.01049727762162956,score,,,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,8L,8,test_seconds_per_article,0.206657676709195,0.002071265185964238,seconds_per_article,24L=100,35.24400698237286,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,8L,8,peak_cuda_allocated_bytes,4257894400.0,3632373.815194686,bytes,24L=100,55.765621575921635,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,8L,8,parameters,420332480,,count,24L=100,55.865014398583234,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,8L,8,total_wall_seconds,4291.582853937677,44.49609583978631,seconds,24L=100,33.418171590571056,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_quality,4L,4,macro_f1,0.5875679559555907,0.007165753447433348,score,,,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_quality,4L,4,micro_f1,0.647978512363484,0.0018739735696546695,score,,,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_quality,4L,4,subset_exact_match,0.006060606060606061,0.01049727762162956,score,,,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,4L,4,test_seconds_per_article,0.08543555729062946,0.0005980667826383295,seconds_per_article,24L=100,14.570430799582711,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,4L,4,peak_cuda_allocated_bytes,2143655936.0,0.0,bytes,24L=100,28.075451029493376,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,4L,4,parameters,210903360,,count,24L=100,28.030475406301182,3,same task protocol; lineage/vocabulary differ
+propaganda_operational_endpoint_resource,4L,4,total_wall_seconds,1643.6388439506602,9.226529522850198,seconds,24L=100,12.798868573555138,3,same task protocol; lineage/vocabulary differ

benchmark/figures/cross_task/06_malint_4l_quality.csv

@@ -0,0 +1,21 @@
+arm,arm_label,role,layers,metric,mean,sample_sd,n_repeats,threshold
+rightmost_hard,Rightmost + hard,qwen_4l_student,4,weighted_f1,0.522182806412052,0.014974460565445307,3,0.3
+rightmost_hard,Rightmost + hard,qwen_4l_student,4,macro_f1,0.5109624469249431,0.016275524962473233,3,0.3
+rightmost_hard,Rightmost + hard,qwen_4l_student,4,micro_f1,0.5320256865023298,0.010757300158510568,3,0.3
+rightmost_hard,Rightmost + hard,qwen_4l_student,4,subset_exact_match,0.6403030303030303,0.003674047167570365,3,0.3
+rightmost_hard,Rightmost + hard,qwen_4l_student,4,label_accuracy,0.8746666666666667,0.0012106051124507965,3,0.3
+mean_hard,Mean pool + hard,qwen_4l_student,4,weighted_f1,0.6118049035859879,0.008860176348616246,3,0.3
+mean_hard,Mean pool + hard,qwen_4l_student,4,macro_f1,0.6055064323204021,0.007533985719586246,3,0.3
+mean_hard,Mean pool + hard,qwen_4l_student,4,micro_f1,0.6176900341883843,0.007461802499370809,3,0.3
+mean_hard,Mean pool + hard,qwen_4l_student,4,subset_exact_match,0.6696969696969697,0.005169309730070291,3,0.3
+mean_hard,Mean pool + hard,qwen_4l_student,4,label_accuracy,0.8876363636363637,0.0017533910474532954,3,0.3
+mean_kd,Mean pool + KD,qwen_4l_student,4,weighted_f1,0.6346533875063399,0.011966996735426847,3,0.3
+mean_kd,Mean pool + KD,qwen_4l_student,4,macro_f1,0.6281768437686408,0.014225316120600004,3,0.3
+mean_kd,Mean pool + KD,qwen_4l_student,4,micro_f1,0.6389632888388834,0.009850482866928375,3,0.3
+mean_kd,Mean pool + KD,qwen_4l_student,4,subset_exact_match,0.6557575757575758,0.003192622349349309,3,0.3
+mean_kd,Mean pool + KD,qwen_4l_student,4,label_accuracy,0.8757575757575757,0.0020436018545560413,3,0.3
+teacher,RoBERTa teacher,external_teacher,,weighted_f1,0.6046800630077942,0.01579900095351724,3,0.3
+teacher,RoBERTa teacher,external_teacher,,macro_f1,0.5955721586042237,0.01637303032583919,3,0.3
+teacher,RoBERTa teacher,external_teacher,,micro_f1,0.6046188233688233,0.01776761734134435,3,0.3
+teacher,RoBERTa teacher,external_teacher,,subset_exact_match,0.62,0.005961307749365459,3,0.3
+teacher,RoBERTa teacher,external_teacher,,label_accuracy,0.8507878787878788,0.01087469324477906,3,0.3

benchmark/figures/cross_task/07_malint_quality_resource.csv

@@ -0,0 +1,15 @@
+panel,batch,arm,metric,value,sample_sd,unit,ratio_vs_rightmost,percent_change,scope
+quality,,rightmost_hard,weighted_f1,0.522182806412052,0.014974460565445307,score,1.0,0.0,post_selection_internal_repeated_cv
+quality,,mean_kd,weighted_f1,0.6346533875063399,0.011966996735426847,score,1.2153854545060183,21.53854545060183,post_selection_internal_repeated_cv
+resource,1,existing_rightmost,latency_ms_per_example,18.474629627574572,,ms_per_example,1.0,0.0,same_input_cuda_forward_only
+resource,1,existing_rightmost,peak_allocated_bytes,1373150208,,bytes,1.0,0.0,same_input_cuda_forward_only
+resource,1,existing_rightmost,parameter_count,334101824,,count,1.0,0.0,same_input_cuda_forward_only
+resource,1,mean_kd,latency_ms_per_example,19.318781786253957,,ms_per_example,1.0456925078172845,4.569250781728451,same_input_cuda_forward_only
+resource,1,mean_kd,peak_allocated_bytes,1374199296,,bytes,1.0007640009038254,0.07640009038254458,same_input_cuda_forward_only
+resource,1,mean_kd,parameter_count,334101829,,count,1.0000000149654975,1.496549750790166e-06,same_input_cuda_forward_only
+resource,4,existing_rightmost,latency_ms_per_example,4.589751928502863,,ms_per_example,1.0,0.0,same_input_cuda_forward_only
+resource,4,existing_rightmost,peak_allocated_bytes,1454388736,,bytes,1.0,0.0,same_input_cuda_forward_only
+resource,4,existing_rightmost,parameter_count,334101824,,count,1.0,0.0,same_input_cuda_forward_only
+resource,4,mean_kd,latency_ms_per_example,4.671012829289292,,ms_per_example,1.0177048568315403,1.7704856831540328,same_input_cuda_forward_only
+resource,4,mean_kd,peak_allocated_bytes,1455437824,,bytes,1.0007213257185181,0.07213257185181288,same_input_cuda_forward_only
+resource,4,mean_kd,parameter_count,334101829,,count,1.0000000149654975,1.496549750790166e-06,same_input_cuda_forward_only

benchmark/figures/cross_task/README.md

@@ -0,0 +1,16 @@
+# Cross-task Matplotlib figures
+
+This directory contains aggregate-only, deterministic plots for the existing
+Qwen3.5 standalone 4L classification-base model card. PNG files are for the
+Hugging Face README; PDF files are vector exports; each figure has a source CSV.
+
+1. `05_why_4l_edge_tradeoff`: controlled UNO Q depth-chain resources plus
+   descriptive propaganda endpoint quality/resources.
+2. `06_malint_4l_quality`: repeated-development F1 and accuracy-type metrics.
+3. `07_malint_quality_resource`: the primary internal quality contrast and its
+   measured CUDA-forward latency/peak-allocation cost.
+
+Scientific boundaries are printed inside the figures. In particular, MALINT
+does not compare 4L against 8L or 24L; all Qwen student arms keep four layers.
+No raw text, IDs, row labels, logits, probabilities, or predictions are copied
+into this public bundle.

benchmark/figures/cross_task/manifest.json

@@ -0,0 +1,158 @@
+{
+  "figures": [
+    "05_why_4l_edge_tradeoff",
+    "06_malint_4l_quality",
+    "07_malint_quality_resource"
+  ],
+  "files_before_manifest": [
+    {
+      "bytes": 5442,
+      "path": "05_why_4l_edge_tradeoff.csv",
+      "sha256": "afd203646513c49140a25ef2a562a064286a4fa5efb495e6a73f83f0106e7933"
+    },
+    {
+      "bytes": 27224,
+      "path": "05_why_4l_edge_tradeoff.pdf",
+      "sha256": "22f531ea586f9c79ac8f114345cec56e74ea5ed97aa3726f7310018666ae6312"
+    },
+    {
+      "bytes": 178965,
+      "path": "05_why_4l_edge_tradeoff.png",
+      "sha256": "cdc0a8a98cd13ad35a68b980d2e0750e2424b8932f7d27dc0ecd548fed823c74"
+    },
+    {
+      "bytes": 2118,
+      "path": "06_malint_4l_quality.csv",
+      "sha256": "25848237b9c77596a8c4b6224fa111f86c816e065804ccfdd6a4fdf6c63ef8d1"
+    },
+    {
+      "bytes": 26351,
+      "path": "06_malint_4l_quality.pdf",
+      "sha256": "9a1ae78b4e5871f5d5c5321db6c5d1fe0bf349a7869977b2649b44ccea633d04"
+    },
+    {
+      "bytes": 181392,
+      "path": "06_malint_4l_quality.png",
+      "sha256": "2c5b22415d311ca5fa2b198d55b77f3570b57595d01f0b5031abde59ffc3e9e6"
+    },
+    {
+      "bytes": 1800,
+      "path": "07_malint_quality_resource.csv",
+      "sha256": "aac7d3400cf8dcb3ac8b6fc317c27ded5343231c93476997011a3bb56b5e4d6c"
+    },
+    {
+      "bytes": 29309,
+      "path": "07_malint_quality_resource.pdf",
+      "sha256": "498338991976e64038096a12361f63830ced3b9c6eea3e1bdaed8d0e4efac606"
+    },
+    {
+      "bytes": 163562,
+      "path": "07_malint_quality_resource.png",
+      "sha256": "1cb13cc695f3d59ae9ce75599b86867da624d6e0aae0897e0db8686f6db984ba"
+    },
+    {
+      "bytes": 868,
+      "path": "README.md",
+      "sha256": "f3e79965f14ec0fd20d11c361d25c3a3abc4c9a20b5dbce9fac0660bb12789f0"
+    },
+    {
+      "bytes": 29752,
+      "path": "source_summary.json",
+      "sha256": "20ea292f85757440d8641199003c47674deb0c6aacea0ef8f9f5fcd2480bb5d0"
+    }
+  ],
+  "inputs": [
+    {
+      "bytes": 58857,
+      "path": "results/malint_qwen35_4l_repeated_cv/20260814-r3x3-terminal-v1/aggregate-v1/aggregate_result.json",
+      "role": "malint_repeated_aggregate",
+      "sha256": "e3d659c2f5aca645f040d03c57e4effa9eb2616f622aceb503c71236b6b01efa"
+    },
+    {
+      "bytes": 325527,
+      "path": "results/malint_qwen35_4l_repeated_cv/20260814-r3x3-terminal-v1/aggregate-v1/aggregate_oof_predictions.npz",
+      "role": "malint_repeated_oof",
+      "sha256": "a1d7827ebe4beecb49676ca42e0c986c5f4101c394d55c9965161481c347a3f1"
+    },
+    {
+      "bytes": 15407,
+      "path": "results/malint_qwen35_4l_resource_benchmark/20260814-matched-bf16-v2/resource_benchmark.json",
+      "role": "malint_resource_aggregate",
+      "sha256": "5337e44c907a05590c0ea3da6a0760baff22c3f930ed81c069c257060fd2e725"
+    },
+    {
+      "bytes": 3144,
+      "path": "results/final_broad_benchmark/20260729-v1/report/quality_matrix.csv",
+      "role": "propaganda_quality_matrix",
+      "sha256": "c6e316f7081e2a6dcda1ec4e95b57cdcc357ae49ed8b3f5a876c5b51e0df722f"
+    },
+    {
+      "bytes": 115661,
+      "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-full24l-commonhead/seed41/report.json",
+      "role": "propaganda_seed_report_01",
+      "sha256": "1c572fb728755dbde0460e6b8eb9a84d078b128f8a42a9148a17c40fc0b27d82"
+    },
+    {
+      "bytes": 116024,
+      "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-full24l-commonhead/seed42/report.json",
+      "role": "propaganda_seed_report_02",
+      "sha256": "055edfee92e39928aa9577f239c5b0389f3dda375a96fc69a2c000a0ed6edc36"
+    },
+    {
+      "bytes": 115639,
+      "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-full24l-commonhead/seed43-retry2/report.json",
+      "role": "propaganda_seed_report_03",
+      "sha256": "84b2c316556db4aa5fdf7c08433a4808a5fe9ef6f620da6472f4f2f73dee2e4b"
+    },
+    {
+      "bytes": 114924,
+      "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-c8q8l-commonhead/seed41/report.json",
+      "role": "propaganda_seed_report_04",
+      "sha256": "9234258589ffb24eaa92859639a66aeb3e55190f8aa82f798d88e791a282bf9c"
+    },
+    {
+      "bytes": 115800,
+      "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-c8q8l-commonhead/seed42/report.json",
+      "role": "propaganda_seed_report_05",
+      "sha256": "8161dd1378081de9960136fddab800994529aa3c754674ca41bb4aa71515b3c6"
+    },
+    {
+      "bytes": 116039,
+      "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-c8q8l-commonhead/seed43/report.json",
+      "role": "propaganda_seed_report_06",
+      "sha256": "ce0c76a2aa97a5b7c8cdf820af0bd4f8a26bf50d460612ca1db30bd497ca3d9d"
+    },
+    {
+      "bytes": 115985,
+      "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-standalone4l-commonhead/seed41/report.json",
+      "role": "propaganda_seed_report_07",
+      "sha256": "ec5314140fff0bdd46e87ee00ad67bee7e61d82c7cfa34add1510dd7d5bdb3c0"
+    },
+    {
+      "bytes": 115972,
+      "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-standalone4l-commonhead/seed42/report.json",
+      "role": "propaganda_seed_report_08",
+      "sha256": "e5c45a7dc41bebf4550da7c5763d0759c6d0922da9efda4048564c11c6d2f3af"
+    },
+    {
+      "bytes": 116177,
+      "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-standalone4l-commonhead/seed43/report.json",
+      "role": "propaganda_seed_report_09",
+      "sha256": "e05bc2146d720706780fc45a1fb05911d5c2926caa0bcb12ffead946028eef0d"
+    },
+    {
+      "bytes": 827,
+      "path": "results/standalone4l_base/20260731-v1/huggingface_release_v3/benchmark/compression_ladder.csv",
+      "role": "task_free_compression_ladder",
+      "sha256": "2c544c3907c4f592d724f658850b8c6b9634845cb5b50a66ebd2ea6c5d48ce0b"
+    },
+    {
+      "bytes": 13963,
+      "path": "results/unoq_lineage_matrix/20260731-v1/analysis/analysis.json",
+      "role": "uno_q_depth_analysis",
+      "sha256": "c899a19e4433e3a40547a11cfcfc3b0a7b6f00c436aa8c1d742a5dc0399bf649"
+    }
+  ],
+  "schema": "qwen35-4l-cross-task-base-card-figures-manifest-v1",
+  "status": "complete"
+}

benchmark/figures/cross_task/source_summary.json

@@ -0,0 +1,972 @@
+{
+  "claim_boundaries": [
+    "The controlled UNO Q 4L point is a historical early-path proxy, not the final common-head quality endpoint.",
+    "The propaganda 24L/8L/4L endpoints differ in lineage and vocabulary; their comparison is descriptive, not a pure depth ablation.",
+    "All Qwen MALINT student arms are 4L; MALINT supports a 4L interface result, not 4L-versus-8L/24L depth superiority.",
+    "The MALINT result is post-selection internal repeated-fold evidence, not untouched held-out confirmation or SOTA evidence.",
+    "Mean+KD incurred measured CUDA-forward latency and peak-allocation increases; it is not claimed to be faster.",
+    "No ambient GPU utilization value is used as model occupancy."
+  ],
+  "malint_quality": [
+    {
+      "arm": "rightmost_hard",
+      "arm_label": "Rightmost + hard",
+      "layers": 4,
+      "mean": 0.522182806412052,
+      "metric": "weighted_f1",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.014974460565445307,
+      "threshold": 0.3
+    },
+    {
+      "arm": "rightmost_hard",
+      "arm_label": "Rightmost + hard",
+      "layers": 4,
+      "mean": 0.5109624469249431,
+      "metric": "macro_f1",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.016275524962473233,
+      "threshold": 0.3
+    },
+    {
+      "arm": "rightmost_hard",
+      "arm_label": "Rightmost + hard",
+      "layers": 4,
+      "mean": 0.5320256865023298,
+      "metric": "micro_f1",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.010757300158510568,
+      "threshold": 0.3
+    },
+    {
+      "arm": "rightmost_hard",
+      "arm_label": "Rightmost + hard",
+      "layers": 4,
+      "mean": 0.6403030303030303,
+      "metric": "subset_exact_match",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.003674047167570365,
+      "threshold": 0.3
+    },
+    {
+      "arm": "rightmost_hard",
+      "arm_label": "Rightmost + hard",
+      "layers": 4,
+      "mean": 0.8746666666666667,
+      "metric": "label_accuracy",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.0012106051124507965,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_hard",
+      "arm_label": "Mean pool + hard",
+      "layers": 4,
+      "mean": 0.6118049035859879,
+      "metric": "weighted_f1",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.008860176348616246,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_hard",
+      "arm_label": "Mean pool + hard",
+      "layers": 4,
+      "mean": 0.6055064323204021,
+      "metric": "macro_f1",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.007533985719586246,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_hard",
+      "arm_label": "Mean pool + hard",
+      "layers": 4,
+      "mean": 0.6176900341883843,
+      "metric": "micro_f1",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.007461802499370809,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_hard",
+      "arm_label": "Mean pool + hard",
+      "layers": 4,
+      "mean": 0.6696969696969697,
+      "metric": "subset_exact_match",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.005169309730070291,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_hard",
+      "arm_label": "Mean pool + hard",
+      "layers": 4,
+      "mean": 0.8876363636363637,
+      "metric": "label_accuracy",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.0017533910474532954,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_kd",
+      "arm_label": "Mean pool + KD",
+      "layers": 4,
+      "mean": 0.6346533875063399,
+      "metric": "weighted_f1",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.011966996735426847,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_kd",
+      "arm_label": "Mean pool + KD",
+      "layers": 4,
+      "mean": 0.6281768437686408,
+      "metric": "macro_f1",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.014225316120600004,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_kd",
+      "arm_label": "Mean pool + KD",
+      "layers": 4,
+      "mean": 0.6389632888388834,
+      "metric": "micro_f1",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.009850482866928375,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_kd",
+      "arm_label": "Mean pool + KD",
+      "layers": 4,
+      "mean": 0.6557575757575758,
+      "metric": "subset_exact_match",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.003192622349349309,
+      "threshold": 0.3
+    },
+    {
+      "arm": "mean_kd",
+      "arm_label": "Mean pool + KD",
+      "layers": 4,
+      "mean": 0.8757575757575757,
+      "metric": "label_accuracy",
+      "n_repeats": 3,
+      "role": "qwen_4l_student",
+      "sample_sd": 0.0020436018545560413,
+      "threshold": 0.3
+    },
+    {
+      "arm": "teacher",
+      "arm_label": "RoBERTa teacher",
+      "layers": "",
+      "mean": 0.6046800630077942,
+      "metric": "weighted_f1",
+      "n_repeats": 3,
+      "role": "external_teacher",
+      "sample_sd": 0.01579900095351724,
+      "threshold": 0.3
+    },
+    {
+      "arm": "teacher",
+      "arm_label": "RoBERTa teacher",
+      "layers": "",
+      "mean": 0.5955721586042237,
+      "metric": "macro_f1",
+      "n_repeats": 3,
+      "role": "external_teacher",
+      "sample_sd": 0.01637303032583919,
+      "threshold": 0.3
+    },
+    {
+      "arm": "teacher",
+      "arm_label": "RoBERTa teacher",
+      "layers": "",
+      "mean": 0.6046188233688233,
+      "metric": "micro_f1",
+      "n_repeats": 3,
+      "role": "external_teacher",
+      "sample_sd": 0.01776761734134435,
+      "threshold": 0.3
+    },
+    {
+      "arm": "teacher",
+      "arm_label": "RoBERTa teacher",
+      "layers": "",
+      "mean": 0.62,
+      "metric": "subset_exact_match",
+      "n_repeats": 3,
+      "role": "external_teacher",
+      "sample_sd": 0.005961307749365459,
+      "threshold": 0.3
+    },
+    {
+      "arm": "teacher",
+      "arm_label": "RoBERTa teacher",
+      "layers": "",
+      "mean": 0.8507878787878788,
+      "metric": "label_accuracy",
+      "n_repeats": 3,
+      "role": "external_teacher",
+      "sample_sd": 0.01087469324477906,
+      "threshold": 0.3
+    }
+  ],
+  "malint_quality_contract": {
+    "claim_boundary": "Post-selection internal repeated-fold robustness and matched ablation only. This result is not untouched confirmation, official held-out evidence, all-label superiority, or SOTA evidence.",
+    "contrasts": {
+      "conditional_kd_mean_kd_minus_mean_hard": {
+        "corrected_resampled_t_interval": {
+          "confidence": 0.95,
+          "critical_t": 2.306004135204166,
+          "degrees_of_freedom": 8,
+          "formula": "sqrt((1/n + mean(n_eval/n_train)) * sample_variance)",
+          "interval": [
+            -0.0015478418507784523,
+            0.05059146195651548
+          ],
+          "mean": 0.024521810052868513,
+          "mean_eval_train_ratio": 0.5000006195533516,
+          "n": 9,
+          "sample_variance": 0.00020913642539219796,
+          "standard_error": 0.011305119321192738,
+          "variance_correction": 0.6111117306644627
+        },
+        "fold_deltas": [
+          0.04524702943141845,
+          0.017956307877355804,
+          -0.0053068243793217285,
+          0.015864642448092825,
+          0.022064206421293142,
+          0.03213041489613799,
+          0.02962871459506977,
+          0.026719118927120977,
+          0.03639268025864939
+        ],
+        "fold_mean": 0.024521810052868513,
+        "fold_sample_standard_deviation": 0.014461549895920491,
+        "name": "conditional_kd_mean_kd_minus_mean_hard",
+        "per_label_mean_delta": {
+          "CPV": 0.0006438268748058898,
+          "PASV": 0.027106289374744647,
+          "PSSA": 0.058741661936162225,
+          "UCPI": 0.0247139382339762,
+          "UIOA": 0.0021463408215049995
+        },
+        "per_label_repeat_deltas": [
+          [
+            -0.01842723004694835,
+            0.04865531704909942,
+            0.015803620583568212,
+            -0.013921424268790505,
+            0.03679497238444562
+          ],
+          [
+            0.025183733463988223,
+            0.070038654494613,
+            -0.0026296349009481412,
+            0.021790722761596548,
+            0.0004450534649754756
+          ],
+          [
+            -0.004825022792622202,
+            0.05753101426477425,
+            -0.006734963218105072,
+            0.0734495696314279,
+            0.0369017888525075
+          ]
+        ],
+        "positive_fold_tail": {
+          "null_probability": 0.5,
+          "one_sided_probability_at_least_observed": 0.01953125,
+          "positive": 8,
+          "total": 9
+        },
+        "repeat_oof_deltas": [
+          0.018545164631776245,
+          0.020840853227636957,
+          0.029159433901642884
+        ],
+        "repeat_oof_t_interval": {
+          "confidence": 0.95,
+          "critical_t": 4.302652729749462,
+          "degrees_of_freedom": 2,
+          "interval": [
+            0.00897537793561895,
+            0.03672158990508511
+          ],
+          "mean": 0.02284848392035203,
+          "n": 3,
+          "sample_standard_deviation": 0.005584676694496962,
+          "standard_error": 0.0032243145929048505
+        }
+      },
+      "mean_kd_minus_teacher": {
+        "corrected_resampled_t_interval": {
+          "confidence": 0.95,
+          "critical_t": 2.306004135204166,
+          "degrees_of_freedom": 8,
+          "formula": "sqrt((1/n + mean(n_eval/n_train)) * sample_variance)",
+          "interval": [
+            -0.03180974553628697,
+            0.08853586985785872
+          ],
+          "mean": 0.02836306216078588,
+          "mean_eval_train_ratio": 0.5000006195533516,
+          "n": 9,
+          "sample_variance": 0.0011141913151793519,
+          "standard_error": 0.026093972156622066,
+          "variance_correction": 0.6111117306644627
+        },
+        "fold_deltas": [
+          -0.03162698747672954,
+          0.06379890775996899,
+          0.056964918230746675,
+          0.04693837159160652,
+          0.001345287770746495,
+          0.014058020214096523,
+          0.00030085368163068615,
+          0.05808798200431842,
+          0.04540020567068814
+        ],
+        "fold_mean": 0.02836306216078588,
+        "fold_sample_standard_deviation": 0.0333795044178213,
+        "name": "mean_kd_minus_teacher",
+        "per_label_mean_delta": {
+          "CPV": 0.03245133981428311,
+          "PASV": 0.04215462872506418,
+          "PSSA": 0.030550793916265178,
+          "UCPI": 0.006157663285133859,
+          "UIOA": 0.05170900008133931
+        },
+        "per_label_repeat_deltas": [
+          [
+            0.010081541882876188,
+            0.09069144184384492,
+            0.058604490711147705,
+            -0.005178527872112815,
+            0.0012970008399624255
+          ],
+          [
+            0.05816392692322364,
+            -0.013100009776126686,
+            0.017979371387083054,
+            0.03578643578643581,
+            0.01576007056748019
+          ],
+          [
+            0.029108550636749497,
+            0.014060949681077295,
+            0.07854313814578717,
+            0.09585597826086956,
+            0.0014159184479589637
+          ]
+        ],
+        "positive_fold_tail": {
+          "null_probability": 0.5,
+          "one_sided_probability_at_least_observed": 0.01953125,
+          "positive": 8,
+          "total": 9
+        },
+        "repeat_oof_deltas": [
+          0.032162246558158025,
+          0.020416569493259984,
+          0.037341157444219175
+        ],
+        "repeat_oof_t_interval": {
+          "confidence": 0.95,
+          "critical_t": 4.302652729749462,
+          "degrees_of_freedom": 2,
+          "interval": [
+            0.00843082924454409,
+            0.05151581975254736
+          ],
+          "mean": 0.029973324498545728,
+          "n": 3,
+          "sample_standard_deviation": 0.008672021342497449,
+          "standard_error": 0.005006793856509082
+        }
+      },
+      "pooling_mean_hard_minus_rightmost_hard": {
+        "corrected_resampled_t_interval": {
+          "confidence": 0.95,
+          "critical_t": 2.306004135204166,
+          "degrees_of_freedom": 8,
+          "formula": "sqrt((1/n + mean(n_eval/n_train)) * sample_variance)",
+          "interval": [
+            0.056762831891190316,
+            0.12504455147590826
+          ],
+          "mean": 0.0909036916835493,
+          "mean_eval_train_ratio": 0.5000006195533516,
+          "n": 9,
+          "sample_variance": 0.00035868079546522353,
+          "standard_error": 0.014805203195932804,
+          "variance_correction": 0.6111117306644627
+        },
+        "fold_deltas": [
+          0.06678423059606675,
+          0.08178035353463753,
+          0.08672762777739151,
+          0.09344774812516743,
+          0.0878998694863643,
+          0.07415570161139795,
+          0.11094545718958193,
+          0.08708564234043492,
+          0.12930659449090132
+        ],
+        "fold_mean": 0.0909036916835493,
+        "fold_sample_standard_deviation": 0.01893886996273071,
+        "name": "pooling_mean_hard_minus_rightmost_hard",
+        "per_label_mean_delta": {
+          "CPV": 0.14860808426516173,
+          "PASV": 0.09778905085334237,
+          "PSSA": 0.07923875253833933,
+          "UCPI": 0.055020095369494,
+          "UIOA": 0.09206394395095778
+        },
+        "per_label_repeat_deltas": [
+          [
+            0.09436936936936935,
+            0.07977821265492502,
+            0.10417829690385194,
+            0.07872402692420255,
+            0.03863953001237319
+          ],
+          [
+            0.15668016194331985,
+            0.04167051934042232,
+            0.06934192200557099,
+            0.11942791523123497,
+            0.06908520415955133
+          ],
+          [
+            0.194774721482796,
+            0.11626752561967063,
+            0.1026716129434504,
+            0.09521521040458958,
+            0.05733555193655748
+          ]
+        ],
+        "positive_fold_tail": {
+          "null_probability": 0.5,
+          "one_sided_probability_at_least_observed": 0.001953125,
+          "positive": 9,
+          "total": 9
+        },
+        "repeat_oof_deltas": [
+          0.07574451062131737,
+          0.08540665760778032,
+          0.10771512329271016
+        ],
+        "repeat_oof_t_interval": {
+          "confidence": 0.95,
+          "critical_t": 4.302652729749462,
+          "degrees_of_freedom": 2,
+          "interval": [
+            0.048890004893968025,
+            0.13035418945390387
+          ],
+          "mean": 0.08962209717393595,
+          "n": 3,
+          "sample_standard_deviation": 0.01639687368671199,
+          "standard_error": 0.009466739436891459
+        }
+      },
+      "primary_mean_kd_minus_rightmost_hard": {
+        "corrected_resampled_t_interval": {
+          "confidence": 0.95,
+          "critical_t": 2.306004135204166,
+          "degrees_of_freedom": 8,
+          "formula": "sqrt((1/n + mean(n_eval/n_train)) * sample_variance)",
+          "interval": [
+            0.07160480852626178,
+            0.15924619494657383
+          ],
+          "mean": 0.1154255017364178,
+          "mean_eval_train_ratio": 0.5000006195533516,
+          "n": 9,
+          "sample_variance": 0.0005909050521203029,
+          "standard_error": 0.019002868442938105,
+          "variance_correction": 0.6111117306644627
+        },
+        "fold_deltas": [
+          0.1120312600274852,
+          0.09973666141199333,
+          0.08142080339806979,
+          0.10931239057326025,
+          0.10996407590765744,
+          0.10628611650753594,
+          0.1405741717846517,
+          0.1138047612675559,
+          0.1656992747495507
+        ],
+        "fold_mean": 0.1154255017364178,
+        "fold_sample_standard_deviation": 0.024308538666902683,
+        "name": "primary_mean_kd_minus_rightmost_hard",
+        "per_label_mean_delta": {
+          "CPV": 0.14925191113996764,
+          "PASV": 0.12489534022808702,
+          "PSSA": 0.13798041447450154,
+          "UCPI": 0.0797340336034702,
+          "UIOA": 0.09421028477246278
+        },
+        "per_label_repeat_deltas": [
+          [
+            0.075942139322421,
+            0.12843352970402444,
+            0.11998191748742015,
+            0.06480260265541204,
+            0.07543450239681881
+          ],
+          [
+            0.18186389540730807,
+            0.11170917383503531,
+            0.06671228710462285,
+            0.14121863799283152,
+            0.0695302576245268
+          ],
+          [
+            0.1899496986901738,
+            0.17379853988444488,
+            0.09593664972534532,
+            0.16866478003601748,
+            0.09423734078906498
+          ]
+        ],
+        "positive_fold_tail": {
+          "null_probability": 0.5,
+          "one_sided_probability_at_least_observed": 0.001953125,
+          "positive": 9,
+          "total": 9
+        },
+        "repeat_oof_deltas": [
+          0.09428967525309362,
+          0.10624751083541728,
+          0.13687455719435304
+        ],
+        "repeat_oof_t_interval": {
+          "confidence": 0.95,
+          "critical_t": 4.302652729749462,
+          "degrees_of_freedom": 2,
+          "interval": [
+            0.05790922361228608,
+            0.1670319385762899
+          ],
+          "mean": 0.11247058109428798,
+          "n": 3,
+          "sample_standard_deviation": 0.02196390209122417,
+          "standard_error": 0.01268086478482286
+        }
+      }
+    },
+    "promotion_gate": {
+      "fold_gate_passed": true,
+      "observed_positive_fold_deltas": 9,
+      "observed_positive_repeat_oof_deltas": 3,
+      "passed": true,
+      "repeat_gate_passed": true,
+      "required_positive_fold_deltas": 7,
+      "required_positive_repeat_oof_deltas": 3
+    },
+    "study": {
+      "fold_count": 9,
+      "fold_runner_protocol": "malint-qwen35-4l-repeated-matched-cv/1.0",
+      "fold_runner_sha256": "c71af81f1182ce860c39c107dd19eff02f1addd8dd646e3eebaadedc4503067c",
+      "heldout_split_accessed": false,
+      "post_selection_internal_evidence": true,
+      "raw_text_written": false,
+      "repeat_count": 3,
+      "rows_per_repeat": 1100,
+      "terminal_evaluation_only": true,
+      "threshold": 0.3
+    }
+  },
+  "malint_resources": {
+    "all_resource_parity_gates_passed": true,
+    "claim_boundary": {
+      "development_labels_read": false,
+      "heldout_access": false,
+      "quality_scored": false,
+      "resource_only": true,
+      "sota_claim_supported": false,
+      "timing_scope": "CUDA model forward with CUDA-resident token IDs and masks; tokenization, loading, disk, and host-to-device transfer excluded"
+    },
+    "comparisons": {
+      "1": {
+        "existing_over_mean_kd_throughput_ratio": 1.0456925078172845,
+        "mean_kd_minus_existing_parameters": 5,
+        "mean_kd_minus_existing_serialized_weight_bytes": -220,
+        "mean_kd_over_existing_latency_ratio": 1.0456925078172845,
+        "mean_kd_over_existing_peak_allocated_ratio": 1.0007640009038254,
+        "paired_round_latency_ratio_summary": {
+          "maximum": 1.0527066487704841,
+          "mean": 1.0488242542281458,
+          "median": 1.0511611345543765,
+          "minimum": 1.042604979359577,
+          "n": 3,
+          "p05": 1.043460594879057,
+          "p95": 1.0525520973488733,
+          "sample_standard_deviation": 0.005441202851039159
+        },
+        "paired_round_latency_ratios": [
+          1.0511611345543765,
+          1.042604979359577,
+          1.0527066487704841
+        ],
+        "resource_parity_gate": {
+          "four_layers_and_length_256_pass": true,
+          "latency_pass": true,
+          "latency_ratio": 1.0456925078172845,
+          "latency_ratio_limit": 1.05,
+          "parameter_pass": true,
+          "parameter_relative_difference": 1.4965497464629228e-08,
+          "parameter_relative_limit": 0.0001,
+          "passed": true,
+          "peak_allocated_ratio": 1.0007640009038254,
+          "peak_memory_pass": true,
+          "peak_memory_ratio_limit": 1.05
+        }
+      },
+      "4": {
+        "existing_over_mean_kd_throughput_ratio": 1.0177048568315403,
+        "mean_kd_minus_existing_parameters": 5,
+        "mean_kd_minus_existing_serialized_weight_bytes": -220,
+        "mean_kd_over_existing_latency_ratio": 1.0177048568315403,
+        "mean_kd_over_existing_peak_allocated_ratio": 1.0007213257185181,
+        "paired_round_latency_ratio_summary": {
+          "maximum": 1.0378811577131521,
+          "mean": 1.0299336811276214,
+          "median": 1.0272510844529572,
+          "minimum": 1.0246688012167542,
+          "n": 3,
+          "p05": 1.0249270295403745,
+          "p95": 1.0368181503871325,
+          "sample_standard_deviation": 0.007002773360331318
+        },
+        "paired_round_latency_ratios": [
+          1.0272510844529572,
+          1.0378811577131521,
+          1.0246688012167542
+        ],
+        "resource_parity_gate": {
+          "four_layers_and_length_256_pass": true,
+          "latency_pass": true,
+          "latency_ratio": 1.0177048568315403,
+          "latency_ratio_limit": 1.05,
+          "parameter_pass": true,
+          "parameter_relative_difference": 1.4965497464629228e-08,
+          "parameter_relative_limit": 0.0001,
+          "passed": true,
+          "peak_allocated_ratio": 1.0007213257185181,
+          "peak_memory_pass": true,
+          "peak_memory_ratio_limit": 1.05
+        }
+      }
+    },
+    "rows": [
+      {
+        "arm": "existing_rightmost",
+        "batch": 1,
+        "latency_ms_per_example": 18.474629627574572,
+        "parameter_count": 334101824,
+        "peak_allocated_bytes": 1373150208,
+        "worker_count": 3
+      },
+      {
+        "arm": "mean_kd",
+        "batch": 1,
+        "latency_ms_per_example": 19.318781786253957,
+        "parameter_count": 334101829,
+        "peak_allocated_bytes": 1374199296,
+        "worker_count": 3
+      },
+      {
+        "arm": "existing_rightmost",
+        "batch": 4,
+        "latency_ms_per_example": 4.589751928502863,
+        "parameter_count": 334101824,
+        "peak_allocated_bytes": 1454388736,
+        "worker_count": 3
+      },
+      {
+        "arm": "mean_kd",
+        "batch": 4,
+        "latency_ms_per_example": 4.671012829289292,
+        "parameter_count": 334101829,
+        "peak_allocated_bytes": 1455437824,
+        "worker_count": 3
+      }
+    ]
+  },
+  "privacy": {
+    "aggregate_only": true,
+    "contains_article_ids": false,
+    "contains_logits_probabilities_or_predictions": false,
+    "contains_raw_text": false,
+    "contains_row_labels": false
+  },
+  "propaganda_operational_endpoints": [
+    {
+      "arm": "qwen35-full24l-commonhead",
+      "label": "24L",
+      "layers": 24,
+      "lineage_note": "same five-epoch common-head protocol",
+      "macro_f1": {
+        "mean": 0.5642638295960175,
+        "n": 3,
+        "sample_sd": 0.006258972649035587
+      },
+      "micro_f1": {
+        "mean": 0.6264844890231888,
+        "n": 3,
+        "sample_sd": 0.01398932248989461
+      },
+      "parameters": 752407360,
+      "peak_cuda_allocated_bytes": {
+        "mean": 7635339264.0,
+        "n": 3,
+        "sample_sd": 0.0
+      },
+      "subset_exact_match": {
+        "mean": 0.0303030303030303,
+        "n": 3,
+        "sample_sd": 0.027773186030035393
+      },
+      "test_articles": 55,
+      "test_seconds_per_article": {
+        "mean": 0.586362602903109,
+        "n": 3,
+        "sample_sd": 0.005979023788731576
+      },
+      "total_wall_seconds": {
+        "mean": 12842.063612925334,
+        "n": 3,
+        "sample_sd": 94.80225434811064
+      }
+    },
+    {
+      "arm": "qwen35-c8q8l-commonhead",
+      "label": "8L",
+      "layers": 8,
+      "lineage_note": "same five-epoch common-head protocol",
+      "macro_f1": {
+        "mean": 0.5366873350531403,
+        "n": 3,
+        "sample_sd": 0.017264532146217884
+      },
+      "micro_f1": {
+        "mean": 0.6238830315003392,
+        "n": 3,
+        "sample_sd": 0.013623163834245318
+      },
+      "parameters": 420332480,
+      "peak_cuda_allocated_bytes": {
+        "mean": 4257894400.0,
+        "n": 3,
+        "sample_sd": 3632373.815194686
+      },
+      "subset_exact_match": {
+        "mean": 0.006060606060606061,
+        "n": 3,
+        "sample_sd": 0.01049727762162956
+      },
+      "test_articles": 55,
+      "test_seconds_per_article": {
+        "mean": 0.206657676709195,
+        "n": 3,
+        "sample_sd": 0.002071265185964238
+      },
+      "total_wall_seconds": {
+        "mean": 4291.582853937677,
+        "n": 3,
+        "sample_sd": 44.49609583978631
+      }
+    },
+    {
+      "arm": "qwen35-standalone4l-commonhead",
+      "label": "4L",
+      "layers": 4,
+      "lineage_note": "same five-epoch common-head protocol",
+      "macro_f1": {
+        "mean": 0.5875679559555907,
+        "n": 3,
+        "sample_sd": 0.007165753447433348
+      },
+      "micro_f1": {
+        "mean": 0.647978512363484,
+        "n": 3,
+        "sample_sd": 0.0018739735696546695
+      },
+      "parameters": 210903360,
+      "peak_cuda_allocated_bytes": {
+        "mean": 2143655936.0,
+        "n": 3,
+        "sample_sd": 0.0
+      },
+      "subset_exact_match": {
+        "mean": 0.006060606060606061,
+        "n": 3,
+        "sample_sd": 0.01049727762162956
+      },
+      "test_articles": 55,
+      "test_seconds_per_article": {
+        "mean": 0.08543555729062946,
+        "n": 3,
+        "sample_sd": 0.0005980667826383295
+      },
+      "total_wall_seconds": {
+        "mean": 1643.6388439506602,
+        "n": 3,
+        "sample_sd": 9.226529522850198
+      }
+    }
+  ],
+  "schema": "qwen35-4l-cross-task-base-card-figures-v1",
+  "status": "complete",
+  "task_free_compression_ladder": [
+    {
+      "layers": 24,
+      "name": "24L upstream text teacher",
+      "parameters": 752393024,
+      "weight_bytes": 1504825632,
+      "weight_mib": 1435.1135559082031
+    },
+    {
+      "layers": 8,
+      "name": "8L task-free KD",
+      "parameters": 420318144,
+      "weight_bytes": 840647880,
+      "weight_mib": 801.7042922973633
+    },
+    {
+      "layers": 6,
+      "name": "6L task-free KD",
+      "parameters": 377207424,
+      "weight_bytes": 754423424,
+      "weight_mib": 719.4742431640625
+    },
+    {
+      "layers": 4,
+      "name": "4L task-free KD root",
+      "parameters": 334096704,
+      "weight_bytes": 668198976,
+      "weight_mib": 637.2442016601562
+    }
+  ],
+  "uno_q_controlled_depth": {
+    "boundary": "Matched V128k/F16 runtime comparison. The 4L arm is the historical early-path proxy, not the final common-head quality model.",
+    "claim_boundaries": [
+      "Only layer_effect_primary is used for same-vocabulary F16 depth-effect estimates.",
+      "Only vocabulary_effect_primary is used for exact-remap vocabulary-effect estimates.",
+      "Operational and cross-generation groups are descriptive because precision, heads, layer maps, training lineages, or architectures differ.",
+      "llama-bench embedding-mode latency measures backbone/classifier compute but does not validate evidence-span quality.",
+      "CPU occupancy is instantaneous saturation; core-seconds is a compute-duration proxy, not measured energy.",
+      "No power meter was used, so energy consumption is not reported.",
+      "Temperature is device- and ambient-specific and is descriptive, not a universal hardware constant."
+    ],
+    "rows": [
+      {
+        "layers": 8,
+        "metric": "artifact_mib",
+        "metric_label": "Artifact size",
+        "model": "8L",
+        "normalized_8l_percent": 100.0
+      },
+      {
+        "layers": 6,
+        "metric": "artifact_mib",
+        "metric_label": "Artifact size",
+        "model": "6L",
+        "normalized_8l_percent": 85.59176355424236
+      },
+      {
+        "layers": 4,
+        "metric": "artifact_mib",
+        "metric_label": "Artifact size",
+        "model": "4L proxy",
+        "normalized_8l_percent": 72.2607217304372
+      },
+      {
+        "layers": 8,
+        "metric": "latency_pp128_ms",
+        "metric_label": "PP128 latency",
+        "model": "8L",
+        "normalized_8l_percent": 100.0
+      },
+      {
+        "layers": 6,
+        "metric": "latency_pp128_ms",
+        "metric_label": "PP128 latency",
+        "model": "6L",
+        "normalized_8l_percent": 76.36565126547553
+      },
+      {
+        "layers": 4,
+        "metric": "latency_pp128_ms",
+        "metric_label": "PP128 latency",
+        "model": "4L proxy",
+        "normalized_8l_percent": 50.16216836129361
+      },
+      {
+        "layers": 8,
+        "metric": "peak_rss_mib",
+        "metric_label": "Peak RSS",
+        "model": "8L",
+        "normalized_8l_percent": 100.0
+      },
+      {
+        "layers": 6,
+        "metric": "peak_rss_mib",
+        "metric_label": "Peak RSS",
+        "model": "6L",
+        "normalized_8l_percent": 87.50983221571457
+      },
+      {
+        "layers": 4,
+        "metric": "peak_rss_mib",
+        "metric_label": "Peak RSS",
+        "model": "4L proxy",
+        "normalized_8l_percent": 78.19200407435702
+      },
+      {
+        "layers": 8,
+        "metric": "mean_core_seconds_per_round",
+        "metric_label": "Core-seconds",
+        "model": "8L",
+        "normalized_8l_percent": 100.0
+      },
+      {
+        "layers": 6,
+        "metric": "mean_core_seconds_per_round",
+        "metric_label": "Core-seconds",
+        "model": "6L",
+        "normalized_8l_percent": 76.16702179915283
+      },
+      {
+        "layers": 4,
+        "metric": "mean_core_seconds_per_round",
+        "metric_label": "Core-seconds",
+        "model": "4L proxy",
+        "normalized_8l_percent": 51.03561371028788
+      }
+    ]
+  }
+}