Add complete frozen benchmark metrics and bounded multilingual metadata

.gitattributes

@@ -1,2 +1,4 @@
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-tokenizer.json filter=lfs diff=lfs merge=lfs -text
+models/semeval-propaganda/seeds/seed42/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+models/semeval-propaganda/seeds/seed43/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+models/semeval-propaganda/tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -4,10 +4,12 @@ base_model: Qwen/Qwen3.5-0.8B
 pipeline_tag: feature-extraction
 library_name: transformers
 language:
-  - en
+  - multilingual
 tags:
+  - multilingual
   - qwen3.5
   - classification-backbone
+  - text-classification
   - knowledge-distillation
   - model-compression
   - edge-ai
@@ -15,18 +17,34 @@ tags:
 
 # Qwen3.5 Standalone 4L Classification Base
 
-This repository publishes a **headless, classification-ready four-layer
-Qwen3.5 text backbone**. The root model has no task labels and no classification
-head. It is intended as a compact starting point for single-label or multilabel
-classifiers, not as a chat model or a ready-made universal classifier.
+This repository publishes a **headless, multilingual-input, classification-ready
+four-layer Qwen3.5 text backbone**. The root model has no task labels and no
+classification head. It is a compact starting point for user-trained
+single-label or multilabel classifiers, not a chat model or a ready-made
+universal classifier.
 
 The repository follows one integrated 1-1-1 layout:
 
-- **Root model:** task-agnostic four-layer base trained on unlabeled general
-  text.
-- **`models/semeval-propaganda/`:** the existing SemEval-derived propaganda
-  classifier as a separate task-specific lineage.
-- **`distillation/`:** a reusable 24L→8L→6L→4L platform for your own data.
+1. **Repository root:** task-agnostic four-layer base trained on unlabeled
+   general text.
+2. **`models/semeval-propaganda/`:** ready-to-use SemEval-derived propaganda
+   classifier preserved as a separate task-specific lineage.
+3. **`distillation/`:** reusable 24L→8L→6L→4L platform for user-owned data.
+
+## Model at a glance
+
+| Item | Root model |
+|---|---|
+| Upstream | [`Qwen/Qwen3.5-0.8B`](https://huggingface.co/Qwen/Qwen3.5-0.8B) |
+| Runtime class | `Qwen3_5TextModel` in Transformers 5.13.0 |
+| Text layers | 4: `linear_attention → full_attention → linear_attention → full_attention` |
+| Hidden / FFN | 1,024 / 3,584 |
+| Vocabulary | full upstream Qwen3.5 tokenizer, 248,320 entries |
+| Root parameters | 334,096,704 |
+| 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 |
 
 ## Load the headless base
 
@@ -38,10 +56,8 @@ tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=False)
 backbone = AutoModel.from_pretrained(model_id, trust_remote_code=False)
 ```
 
-The root uses `Qwen3_5TextModel` in Transformers 5.13.0. It has four text
-layers in the sequence `linear → full → linear → full`, hidden size 1,024, and
-the full upstream Qwen3.5 tokenizer. The published weight file contains no
-`score.weight`.
+The published root weight file contains no `score.weight`. A fresh task head and
+labeled fine-tuning are required before outputs have class meaning.
 
 ## Make your own classifier
 
@@ -58,72 +74,199 @@ qwen35-distill finetune \
   --output my-classifier-4l
 ```
 
-For a full task-specific compression run, first fine-tune a 24-layer teacher,
-then repeat `materialize-classifier` and `distill-classifier` for `24to8`,
-`8to6`, and `6to4`. Exact JSONL formats and commands are in
+For a full task-specific compression run, fine-tune a 24-layer teacher, then
+repeat `materialize-classifier` and `distill-classifier` for `24to8`, `8to6`,
+and `6to4`. Exact JSONL formats, split rules, manifests, and commands are in
 [`distillation/README.md`](distillation/README.md).
 
 ## How the root weights were made
 
-The text backbone and full tokenizer were extracted from
-[`Qwen/Qwen3.5-0.8B`](https://huggingface.co/Qwen/Qwen3.5-0.8B). Students were
-trained sequentially through 24→8→6→4 using hidden-boundary and final
-representation matching on unlabeled text.
+The text backbone and full tokenizer were extracted from the upstream
+Qwen3.5-0.8B model. Students were trained sequentially through 24→8→6→4 using
+hidden-boundary and final-representation matching on unlabeled text.
 
 The task-free training set was exactly 4,096 deterministically shuffled,
 non-empty rows from the first locally cached WikiText-103 raw training shard.
-No SemEval examples, labels, logits, thresholds, or evidence/span objectives
-were read while training these root weights.
+No SemEval examples, labels, logits, thresholds, classification heads, or
+evidence/span objectives were read while training the root weights.
 
-WikiText is not redistributed here. Its current Hugging Face page has a license
-wording discrepancy: metadata lists CC BY-SA 3.0 and GFDL, while the prose says
-CC BY-SA 4.0. Users should review the
-[`Salesforce/wikitext`](https://huggingface.co/datasets/Salesforce/wikitext)
-terms directly.
+| KD stage | Selected source layers | Mean total loss | Mean interface loss | Mean final loss |
+|---|---|---:|---:|---:|
+| 24→8 | `0,4,6,11,13,16,20,23` | 0.176227 | 0.215628 | 0.166264 |
+| 8→6 | `0,1,3,4,6,7` | 0.033821 | 0.044861 | 0.031147 |
+| 6→4 | `0,2,3,5` | 0.047020 | 0.070484 | 0.043120 |
 
-## One-task transfer check
+These are training-objective values, not downstream quality scores. Exact
+parameter counts, BF16 file sizes, hashes, and stage values are in
+[`benchmark/compression_ladder.csv`](benchmark/compression_ladder.csv) and
+[`provenance/task_agnostic_stages/`](provenance/task_agnostic_stages/).
 
-We attached a fresh 14-label head and used the frozen SemEval-derived
-article-level protocol only as a transfer probe.
+## Downstream benchmark contract
 
-| Initialization | Opened-test Macro-F1, mean ± sample SD |
-|---|---:|
-| Task-agnostic 4L base | **0.58748 ± 0.01579** |
-| Same 4L structural copy without task-free KD | 0.51809 ± 0.01631 |
-| Existing separate specialized 4L lineage | 0.58757 ± 0.00717 |
-
-Results use seeds 41/42/43 and a previously opened 55-article public test split.
-They are exploratory, not untouched confirmatory evidence. The derived task
-measures article-level presence of 14 techniques and is not identical to the
-official SemEval span- or fragment-level task. Three seed values are not a
+Because the root is headless, it cannot be scored as a classifier without
+adaptation. Its task-appropriate benchmark is therefore a **controlled transfer
+probe**:
+
+- attach a fresh randomly initialized 14-label classification head;
+- train all parameters for five epochs on the same SemEval-derived train split;
+- select one global threshold on the epoch-5 calibration split;
+- evaluate once per seed on the already-opened 55-article test split;
+- compare against the same four-layer structural copy without task-free KD.
+
+This derived task measures article-level presence of 14 propaganda techniques.
+It is not the official SemEval span- or fragment-level task. Results are
+exploratory because the public test split had already been opened.
+
+### Frozen protocol
+
+| Setting | Value |
+|---|---|
+| Seeds | 41, 42, 43 |
+| Train / calibration / test articles | 260 / 56 / 55 |
+| Test windows | 434 |
+| Epochs | 5, no early stopping |
+| Input windows | max length 256, overlap stride 128 |
+| Article aggregation | label-wise maximum probability |
+| Batch | micro 1, gradient accumulation 32, effective 32 |
+| Optimizer | AdamW, LR 2e-5, weight decay 0.01, gradient clip 1.0 |
+| Precision / device | BF16 / NVIDIA GeForce RTX 5070 Ti |
+| Threshold selection | epoch-5 calibration Macro-F1, then Micro-F1 |
+
+### Overall results by seed
+
+| Initialization | Seed | Threshold | Macro-F1 | Micro-F1 | Exact match | Predicted positive rate |
+|---|---:|---:|---:|---:|---:|---:|
+| Task-agnostic KD base | 41 | 0.15 | 0.58831 | 0.64681 | 0.05455 | 0.57143 |
+| Task-agnostic KD base | 42 | 0.30 | 0.60284 | 0.66552 | 0.03636 | 0.40909 |
+| Task-agnostic KD base | 43 | 0.10 | 0.57128 | 0.64653 | 0.00000 | 0.51558 |
+| Structural copy, no task-free KD | 41 | 0.05 | 0.51694 | 0.57740 | 0.00000 | 0.71299 |
+| Structural copy, no task-free KD | 42 | 0.20 | 0.53494 | 0.60834 | 0.00000 | 0.62078 |
+| Structural copy, no task-free KD | 43 | 0.20 | 0.50238 | 0.58564 | 0.00000 | 0.59610 |
+
+### Three-seed summary
+
+Values are mean ± sample SD across the three training seeds. The SD is not a
 confidence interval.
 
-This check supports transfer on one English classification task only. It does
-not establish universal classification quality, multilingual generalization,
-calibration, safety, or production suitability.
+| Initialization | Macro-F1 | Micro-F1 | Exact match | Predicted positive rate |
+|---|---:|---:|---:|---:|
+| **Task-agnostic KD base** | **0.58748 ± 0.01579** | **0.65295 ± 0.01088** | **0.03030 ± 0.02777** | 0.49870 ± 0.08248 |
+| Structural copy, no task-free KD | 0.51809 ± 0.01631 | 0.59046 ± 0.01603 | 0.00000 ± 0.00000 | 0.64329 ± 0.06161 |
+| Existing specialized 4L, separate lineage | 0.58757 ± 0.00717 | see nested reports | see nested reports | see nested reports |
+
+The task-agnostic KD initialization improved same-seed mean Macro-F1 by
+`+0.06939` over the structural-copy control. This is an observed comparison
+under this protocol, not a significance or universal-superiority claim.
+
+### Task-agnostic base: per-label test results
+
+Precision, recall, and F1 are three-seed means. Support is the fixed number of
+positive test articles, not a three-seed sum.
+
+| Label | Support | Precision | Recall | F1 |
+|---|---:|---:|---:|---:|
+| `Appeal_to_Authority` | 11 | 0.3959 | 0.5758 | 0.4657 |
+| `Appeal_to_fear-prejudice` | 24 | 0.5540 | 0.7361 | 0.6278 |
+| `Bandwagon,Reductio_ad_hitlerum` | 7 | 0.4852 | 0.8095 | 0.6007 |
+| `Black-and-White_Fallacy` | 12 | 0.3668 | 0.4722 | 0.4054 |
+| `Causal_Oversimplification` | 18 | 0.5429 | 0.7963 | 0.6447 |
+| `Doubt` | 22 | 0.5224 | 0.9848 | 0.6802 |
+| `Exaggeration,Minimisation` | 22 | 0.4621 | 0.7727 | 0.5727 |
+| `Flag-Waving` | 17 | 0.5643 | 0.8235 | 0.6465 |
+| `Loaded_Language` | 45 | 0.8649 | 0.9852 | 0.9209 |
+| `Name_Calling,Labeling` | 33 | 0.7381 | 0.9091 | 0.8125 |
+| `Repetition` | 23 | 0.5197 | 0.7536 | 0.6148 |
+| `Slogans` | 14 | 0.5161 | 0.7381 | 0.5835 |
+| `Thought-terminating_Cliches` | 7 | 0.3492 | 0.3333 | 0.3190 |
+| `Whataboutism,Straw_Men,Red_Herring` | 10 | 0.2762 | 0.5000 | 0.3304 |
+
+Per-seed values and sample SDs for all labels and all arms are available in
+[`benchmark/per_label_metrics.csv`](benchmark/per_label_metrics.csv) and
+[`benchmark/per_label_summary.csv`](benchmark/per_label_summary.csv).
+
+### Efficiency and resource measurements
+
+These numbers belong to the fresh-head transfer checkpoints, not a standalone
+headless-root latency benchmark.
+
+| Measure | Task-agnostic transfer model |
+|---|---:|
+| Parameters with 14-label head | 334,111,040 |
+| Test inference | 0.08131 ± 0.00243 s/article |
+| Test batch | 55 articles / 434 windows |
+| Peak CUDA allocated | 3,380,975,616 bytes (3.15 GiB) |
+| Total train-and-evaluate wall time | 1,591.69 s mean (26.53 min) |
+
+The timing is specific to the recorded RTX 5070 Ti, BF16, batch-1,
+sliding-window protocol and includes article aggregation. It is not a claim
+about CPU, mobile, UNO Q, or other deployment performance. Full seed-level
+resource fields are in
+[`benchmark/resource_metrics.csv`](benchmark/resource_metrics.csv).
+
+### Machine-readable benchmark bundle
+
+- [`benchmark/BENCHMARK_CARD.md`](benchmark/BENCHMARK_CARD.md): generated
+  benchmark report and interpretation
+- [`benchmark/summary.json`](benchmark/summary.json): complete aggregate
+- [`benchmark/seed_metrics.csv`](benchmark/seed_metrics.csv): overall metrics
+  per arm and seed
+- [`benchmark/arm_summary.csv`](benchmark/arm_summary.csv): three-seed means
+  and sample SDs
+- [`benchmark/per_label_metrics.csv`](benchmark/per_label_metrics.csv):
+  per-label values per arm and seed
+- [`benchmark/per_label_summary.csv`](benchmark/per_label_summary.csv):
+  per-label means and sample SDs
+- [`benchmark/resource_metrics.csv`](benchmark/resource_metrics.csv):
+  parameters, CUDA memory, wall time, and test timing
+- [`benchmark/reports/`](benchmark/reports/): normalized full frozen reports,
+  including article-level probability vectors
+
+## Multilingual scope
+
+The root retains the upstream Qwen3.5 tokenizer and multilingual architecture.
+The [official Qwen3.5-0.8B model card](https://huggingface.co/Qwen/Qwen3.5-0.8B)
+states expanded support for **201 languages and dialects** and reports
+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.
 
 ## Separate SemEval model
 
-The ready-to-use propaganda model is under
+The ready-to-use propaganda classifier is under
 [`models/semeval-propaganda/`](models/semeval-propaganda/). It preserves the
-currently published seed-41/42/43 checkpoints and links to the original
-repository at immutable revision
+published seed-41/42/43 checkpoints and links to the original repository at
+immutable revision
 [`e7e0ff1`](https://huggingface.co/mp-juuuns/qwen35-standalone4l-propaganda-classifier/tree/e7e0ff16828052687d2e8dd7849e7a521629cf38).
 
-That specialized model is **not** claimed to descend from this new task-free
-base. It has its own SemEval task-specific shrink/distillation and fine-tuning
-history and a reduced 128k vocabulary, while the root base retains the full
-upstream tokenizer.
+That specialized model is **not** claimed to descend from this task-free base.
+It has its own SemEval task-specific shrink/distillation and fine-tuning history
+and a reduced 128k vocabulary, while the root retains the full upstream
+tokenizer. Its complete reports remain under
+[`models/semeval-propaganda/benchmark/`](models/semeval-propaganda/benchmark/).
 
 ## Limitations
 
 - A classification head and labeled training are required before root-model
   predictions have task meaning.
-- The task-free corpus is small: 4,096 rows from one WikiText shard.
-- Only one downstream transfer task was measured.
+- Task-free KD used only 4,096 rows from one English WikiText shard.
+- Only one downstream task and one hardware/software setting were measured.
+- 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.
+- 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.
-- Long-context, multilingual, calibration, robustness, fairness, and device
-  performance of the new root base have not been established.
+- Long-context, calibration, robustness, fairness, and production safety have
+  not been established.
 - Do not use either model as a fact checker, safety oracle, or autonomous
   decision maker.
 
@@ -134,7 +277,7 @@ 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.
 
-Machine-readable training manifests, transfer summaries, hashes, and the
-release inventory are included under `provenance/`, `benchmark/`, and
-`release_manifest.json`.
-
+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)
+directly.

SHA256SUMS

@@ -1,13 +1,24 @@
 ee05350c7fda0af4da116020dd7b976d731de3beed4f5e717d5bcc95ead24550  .gitattributes
 77fd4710def9ec3c0f6225800e0235f15a425abd4a8b03559127fcd782612049  LICENSE
-4f5af67b76f19fca88e242298fa501aa6aa811b4f1f51b481d17abb46e0d39cf  README.md
-816c3249d3ec281ef73d34dc3ab7576a48415d250c32977eb94ca37f73b0b810  benchmark/reports/structural_control-seed41.json
-0b19eabe1e31d0012c754db3b31173f8a6fcdee30973f1a1bebd3a370d11064c  benchmark/reports/structural_control-seed42.json
-6dde34173462eac50141ff082c18194f3233f8c06853532d9b12756562e69e2e  benchmark/reports/structural_control-seed43.json
-6b3a2e0372a83e1e71c2eca9596efc8cd1a3e39c28e25e3373f195431b5858a4  benchmark/reports/task_agnostic-seed41.json
-42dc788983716be0e9847aad87ff79c7e7297884f2635a8d098639be77326639  benchmark/reports/task_agnostic-seed42.json
-7ee52dc94a152a2f767dcff171f3c9cc1072fabd3ebc7b14db53617b7b1d33d5  benchmark/reports/task_agnostic-seed43.json
-18c8d840dd2290ca25c517d52cbcf1ad6c6ae8d9c32c545236e96bd5ada7dab2  benchmark/semeval-transfer-summary.json
+5fcf007de6b6e828711c874310de1e042f1dada6989c70930ac0881285a9dc61  README.md
+bf4073de38c3905effd2dcce6f38894f325c8a0a79bedfb2779f0545c86b9c9f  benchmark/BENCHMARK_CARD.md
+e01bd7ac59a3212e44b77e5fcfbc7c41ec91d74c998306d2438209d6ef1a3aa7  benchmark/arm_summary.csv
+2c544c3907c4f592d724f658850b8c6b9634845cb5b50a66ebd2ea6c5d48ce0b  benchmark/compression_ladder.csv
+ffc21be32b2106c7bf1df3854097a975ac370da01ac1b732eb5c2ba9075ddeda  benchmark/per_label_metrics.csv
+c8e3876abf6cfe960c9332d3bc9d1fbde2f3213f3111351acf83b21bbb184049  benchmark/per_label_summary.csv
+c1bd379680abcf50b47b7d23e4aafbc8b5e7f40a0b0e9700bde21a8ce49af803  benchmark/reports/existing_specialized_separate_lineage-seed41.json
+a30e38433d008db4aa7e5fbfd477ab80a0daf50e970f8e7b04397efc1a47843a  benchmark/reports/existing_specialized_separate_lineage-seed42.json
+96466290d09d27b8bc044a9f9a95765ce12b4d7b02dfb0123e436e7535623353  benchmark/reports/existing_specialized_separate_lineage-seed43.json
+816c3249d3ec281ef73d34dc3ab7576a48415d250c32977eb94ca37f73b0b810  benchmark/reports/structural_copy_control-seed41.json
+0b19eabe1e31d0012c754db3b31173f8a6fcdee30973f1a1bebd3a370d11064c  benchmark/reports/structural_copy_control-seed42.json
+6dde34173462eac50141ff082c18194f3233f8c06853532d9b12756562e69e2e  benchmark/reports/structural_copy_control-seed43.json
+6b3a2e0372a83e1e71c2eca9596efc8cd1a3e39c28e25e3373f195431b5858a4  benchmark/reports/task_agnostic_base-seed41.json
+42dc788983716be0e9847aad87ff79c7e7297884f2635a8d098639be77326639  benchmark/reports/task_agnostic_base-seed42.json
+7ee52dc94a152a2f767dcff171f3c9cc1072fabd3ebc7b14db53617b7b1d33d5  benchmark/reports/task_agnostic_base-seed43.json
+bf5cefdec3b2f3646114bd18bbf8e5e39721bde768ba968d73f22e96fc9ca8cf  benchmark/resource_metrics.csv
+e0b08a958d25243158a03dff76ca153d2ded76cc3f1771e4046a0068688ad6a8  benchmark/seed_metrics.csv
+3d2c9fcf9c6cc7da1a885968b70a42eaa38659d317c9d3290a2b3066f3f9ee6a  benchmark/semeval-transfer-summary.json
+3d2c9fcf9c6cc7da1a885968b70a42eaa38659d317c9d3290a2b3066f3f9ee6a  benchmark/summary.json
 04b007131663760bf3e581e5a953be77044014e87efe1d2a6ca4b72ec0eac978  chat_template.jinja
 bb33f3a52d787e15db7d516c6de36a7ddd2c675e7964d5b9309131c4e5d365f0  config.json
 f5b7248f7781412e8b5d8f2fb3b0c309568e4e75baf55029b26a6fa985d9d608  distillation/README.md
@@ -25,11 +36,11 @@ a9c49af0ee14b307a08c096bcd1ae55fa9bc67bdb807fa9383fc313246e7edd7  distillation/q
 d9b135cf5b0adf4bbfc321200395fbd6112e49fd1a027b30262bcff1c4c0b654  distillation/qwen35_distill/losses.py
 00731ae7838317105f9fff960e3da210b2a55954feb37b3ac793bcfc81fdab2d  distillation/qwen35_distill/schema.py
 4dc432acb70e03848d2e5964e8550065490fda3857aac32fe3b79895ef72026f  distillation/qwen35_distill/training.py
-5a1d9e939cd0f350770bcc2a116bc427bb9971bc4fbbbda17ce568e9127185ef  docs/RELEASE_CONTRACT.md
+1b6e475b9d7d9af52bb8845de020303b1c6e89646d8e4f9be17dbe4ab2910417  docs/RELEASE_CONTRACT.md
 a9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d  merges.txt
 2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3  model.safetensors
 77fd4710def9ec3c0f6225800e0235f15a425abd4a8b03559127fcd782612049  models/semeval-propaganda/LICENSE
-8dbc2e641280c471b78064808ced7fffcb81c639784a5936b57175975a718f6f  models/semeval-propaganda/README.md
+bf4ba7cc2514502a28f424507c41c1cfce5089c051ba7500cf46c1d13860722f  models/semeval-propaganda/README.md
 0bae347bb54f8089a5fa5450f8366b5c56eb3e598beaf9a87f4385363a402bcb  models/semeval-propaganda/benchmark/final4l_aggregate.json
 c1bd379680abcf50b47b7d23e4aafbc8b5e7f40a0b0e9700bde21a8ce49af803  models/semeval-propaganda/benchmark/seed41.json
 a30e38433d008db4aa7e5fbfd477ab80a0daf50e970f8e7b04397efc1a47843a  models/semeval-propaganda/benchmark/seed42.json
@@ -59,7 +70,7 @@ ec734855870758cbfef99031dac2ded85b54c496532e5c514a9a4f63287c3493  provenance/bas
 b9ce99b993c4f4563bc2521fc731bfd82f5e8149c141133599fbd8f982ba43d9  provenance/task_agnostic_stages/24to8.json
 ec734855870758cbfef99031dac2ded85b54c496532e5c514a9a4f63287c3493  provenance/task_agnostic_stages/6to4.json
 65e5383afbbcfe0b7dd78a5411c319ed6bf2586659fb1892aca9ecbd5dfd4993  provenance/task_agnostic_stages/8to6.json
-143945a19235351e00048a99c74799e362eefec3b5e3f3a2ad4c8b64611034a4  release_manifest.json
+207e1df5765c838b7d355a0c34cf5ae2c55f90b419cff0f4bbf80528a48fd939  release_manifest.json
 7b48e7811e56e552fc4aef44189e1d7c9b1352c85d2b3f10ff0bdd1ea5cdd824  requirements.txt
 06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523  tokenizer.json
 5ab9bed0a4d27949672f65ba1141d6dd5b0514fb9091d3426506ecebb5d5e294  tokenizer_config.json

benchmark/BENCHMARK_CARD.md

@@ -0,0 +1,68 @@
+# Complete frozen transfer benchmark
+
+This report is generated from the nine frozen seed reports distributed in
+`benchmark/reports/`. It evaluates a headless base through fresh-head
+fine-tuning; it is not a zero-shot root-model score.
+
+## Scope
+
+- Derived task: article-level presence of 14 propaganda techniques.
+- Test: 55 previously opened English articles / 434 windows.
+- Seeds: 41, 42, 43.
+- Error bars: three-seed sample SD, not confidence intervals.
+- No multilingual downstream evaluation was performed.
+
+## Overall metrics
+
+| Arm | Macro-F1 | Micro-F1 | Exact match | Positive rate |
+|---|---:|---:|---:|---:|
+| Task-agnostic KD base | 0.58748 ± 0.01579 | 0.65295 ± 0.01088 | 0.03030 ± 0.02777 | 0.49870 ± 0.08248 |
+| Structural copy, no task-free KD | 0.51809 ± 0.01631 | 0.59046 ± 0.01603 | 0.00000 ± 0.00000 | 0.64329 ± 0.06161 |
+| Existing specialized 4L, separate lineage | 0.58757 ± 0.00717 | 0.64798 ± 0.00187 | 0.00606 ± 0.01050 | 0.58052 ± 0.02848 |
+
+The observed mean same-seed Macro-F1 delta between the task-agnostic
+base and structural control is `+0.06939`.
+The specialized arm is a separate historical lineage and is reference-only.
+
+## Task-agnostic base per-label metrics
+
+| Label | Support | Precision | Recall | F1 |
+|---|---:|---:|---:|---:|
+| `Appeal_to_Authority` | 11 | 0.3959 | 0.5758 | 0.4657 |
+| `Appeal_to_fear-prejudice` | 24 | 0.5540 | 0.7361 | 0.6278 |
+| `Bandwagon,Reductio_ad_hitlerum` | 7 | 0.4852 | 0.8095 | 0.6007 |
+| `Black-and-White_Fallacy` | 12 | 0.3668 | 0.4722 | 0.4054 |
+| `Causal_Oversimplification` | 18 | 0.5429 | 0.7963 | 0.6447 |
+| `Doubt` | 22 | 0.5224 | 0.9848 | 0.6802 |
+| `Exaggeration,Minimisation` | 22 | 0.4621 | 0.7727 | 0.5727 |
+| `Flag-Waving` | 17 | 0.5643 | 0.8235 | 0.6465 |
+| `Loaded_Language` | 45 | 0.8649 | 0.9852 | 0.9209 |
+| `Name_Calling,Labeling` | 33 | 0.7381 | 0.9091 | 0.8125 |
+| `Repetition` | 23 | 0.5197 | 0.7536 | 0.6148 |
+| `Slogans` | 14 | 0.5161 | 0.7381 | 0.5835 |
+| `Thought-terminating_Cliches` | 7 | 0.3492 | 0.3333 | 0.3190 |
+| `Whataboutism,Straw_Men,Red_Herring` | 10 | 0.2762 | 0.5000 | 0.3304 |
+
+## Resource scope
+
+Resource rows describe the fresh-head transfer checkpoints on the recorded
+RTX 5070 Ti BF16 sliding-window protocol. They are not a headless-root CPU
+or edge-device benchmark. See `resource_metrics.csv` for every seed.
+
+## Files
+
+- `summary.json`: aggregate and interpretation contract.
+- `seed_metrics.csv`: overall metrics and thresholds by arm/seed.
+- `arm_summary.csv`: means and sample SDs.
+- `per_label_metrics.csv`: precision/recall/F1/support by arm/seed/label.
+- `per_label_summary.csv`: per-label means and sample SDs.
+- `resource_metrics.csv`: parameters, timing, and CUDA memory.
+- `compression_ladder.csv`: 24L→8L→6L→4L size and KD evidence.
+- `reports/`: normalized full frozen reports, including probabilities.
+
+## Interpretation boundary
+
+The opened test and prior model-selection history prevent confirmatory
+inference. Three seeds do not represent 165 independent observations.
+Upstream support for 201 languages/dialects motivates multilingual metadata,
+but this four-layer root has no direct multilingual downstream score.

benchmark/arm_summary.csv

@@ -0,0 +1,13 @@
+arm,metric,values,mean,sample_sd,n
+task_agnostic_base,macro_f1,"[0.5883093293517536, 0.6028379036218096, 0.5712816197954101]",0.5874762842563245,0.01579462678670608,3
+task_agnostic_base,micro_f1,"[0.6468085106382978, 0.6655172413793103, 0.6465256797583082]",0.6529504772586388,0.010884055708931178,3
+task_agnostic_base,exact_match,"[0.05454545454545454, 0.03636363636363636, 0.0]",0.0303030303030303,0.027773186030035393,3
+task_agnostic_base,predicted_positive_rate,"[0.5714285714285714, 0.4090909090909091, 0.5155844155844156]",0.4987012987012987,0.08247520160679181,3
+structural_copy_control,macro_f1,"[0.5169424579262757, 0.5349405779540175, 0.5023778364496152]",0.5180869574433028,0.016311512588628545,3
+structural_copy_control,micro_f1,"[0.5773955773955775, 0.6083445491251682, 0.585635359116022]",0.5904584952122559,0.01602830914762526,3
+structural_copy_control,exact_match,"[0.0, 0.0, 0.0]",0.0,0.0,3
+structural_copy_control,predicted_positive_rate,"[0.712987012987013, 0.6207792207792208, 0.5961038961038961]",0.6432900432900432,0.06160737456395461,3
+existing_specialized_separate_lineage,macro_f1,"[0.5832848452288824, 0.5958405181160762, 0.5835785045218137]",0.5875679559555907,0.007165753447433348,3
+existing_specialized_separate_lineage,micro_f1,"[0.6494252873563218, 0.6458616010854816, 0.6486486486486486]",0.647978512363484,0.0018739735696546695,3
+existing_specialized_separate_lineage,exact_match,"[0.0, 0.0, 0.01818181818181818]",0.006060606060606061,0.01049727762162956,3
+existing_specialized_separate_lineage,predicted_positive_rate,"[0.5597402597402598, 0.612987012987013, 0.5688311688311688]",0.5805194805194805,0.02848274311618351,3

benchmark/compression_ladder.csv

@@ -0,0 +1,5 @@
+name,layers,parameters,weight_bytes,weight_mib,weight_sha256,kd_stage,mean_total_loss,mean_interface_loss,mean_final_loss
+24L upstream text teacher,24,752393024,1504825632,1435.1135559082031,30783da4651259c26b8a204821ea426b9526431e0c14da0d1c6fa284e7414e5e,,,,
+8L task-free KD,8,420318144,840647880,801.7042922973633,086fbaf9a4838ecde56cd3541c9c3212b22d0c5ee745cf2c1c5e0bb31e0ec22c,24to8,0.17622653172702485,0.21562811763578793,0.16626427527853593
+6L task-free KD,6,377207424,754423424,719.4742431640625,492749c20f77a42f6bc993f0927a82217b94f872b7eacf25612f7520c9bc28ce,8to6,0.03382104352249371,0.04486069045515251,0.031147060785087888
+4L task-free KD root,4,334096704,668198976,637.2442016601562,2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3,6to4,0.047020394468859195,0.07048421185527332,0.043119939065377366

benchmark/per_label_metrics.csv

@@ -0,0 +1,127 @@
+arm,seed,label,support,precision,recall,f1
+task_agnostic_base,41,Appeal_to_Authority,11,0.3333333333333333,0.5454545454545454,0.41379310344827586
+task_agnostic_base,41,Appeal_to_fear-prejudice,24,0.5757575757575758,0.7916666666666666,0.6666666666666667
+task_agnostic_base,41,"Bandwagon,Reductio_ad_hitlerum",7,0.5,0.8571428571428571,0.631578947368421
+task_agnostic_base,41,Black-and-White_Fallacy,12,0.3333333333333333,0.5833333333333334,0.4242424242424242
+task_agnostic_base,41,Causal_Oversimplification,18,0.5,0.6666666666666666,0.5714285714285715
+task_agnostic_base,41,Doubt,22,0.4782608695652174,1.0,0.6470588235294118
+task_agnostic_base,41,"Exaggeration,Minimisation",22,0.4523809523809524,0.8636363636363636,0.59375
+task_agnostic_base,41,Flag-Waving,17,0.4594594594594595,1.0,0.6296296296296297
+task_agnostic_base,41,Loaded_Language,45,0.9183673469387755,1.0,0.9574468085106383
+task_agnostic_base,41,"Name_Calling,Labeling",33,0.7333333333333333,1.0,0.846153846153846
+task_agnostic_base,41,Repetition,23,0.5428571428571428,0.8260869565217391,0.6551724137931034
+task_agnostic_base,41,Slogans,14,0.41379310344827586,0.8571428571428571,0.5581395348837208
+task_agnostic_base,41,Thought-terminating_Cliches,7,0.21428571428571427,0.42857142857142855,0.2857142857142857
+task_agnostic_base,41,"Whataboutism,Straw_Men,Red_Herring",10,0.22857142857142856,0.8,0.3555555555555555
+task_agnostic_base,42,Appeal_to_Authority,11,0.5384615384615384,0.6363636363636364,0.5833333333333334
+task_agnostic_base,42,Appeal_to_fear-prejudice,24,0.56,0.5833333333333334,0.5714285714285714
+task_agnostic_base,42,"Bandwagon,Reductio_ad_hitlerum",7,0.5555555555555556,0.7142857142857143,0.6250000000000001
+task_agnostic_base,42,Black-and-White_Fallacy,12,0.45454545454545453,0.4166666666666667,0.43478260869565216
+task_agnostic_base,42,Causal_Oversimplification,18,0.5769230769230769,0.8333333333333334,0.6818181818181818
+task_agnostic_base,42,Doubt,22,0.6,0.9545454545454546,0.7368421052631579
+task_agnostic_base,42,"Exaggeration,Minimisation",22,0.48148148148148145,0.5909090909090909,0.5306122448979591
+task_agnostic_base,42,Flag-Waving,17,0.7333333333333333,0.6470588235294118,0.6875
+task_agnostic_base,42,Loaded_Language,45,0.8461538461538461,0.9777777777777777,0.9072164948453608
+task_agnostic_base,42,"Name_Calling,Labeling",33,0.7428571428571429,0.7878787878787878,0.7647058823529412
+task_agnostic_base,42,Repetition,23,0.5,0.7391304347826086,0.5964912280701754
+task_agnostic_base,42,Slogans,14,0.75,0.6428571428571429,0.6923076923076924
+task_agnostic_base,42,Thought-terminating_Cliches,7,0.3333333333333333,0.2857142857142857,0.30769230769230765
+task_agnostic_base,42,"Whataboutism,Straw_Men,Red_Herring",10,0.26666666666666666,0.4,0.32
+task_agnostic_base,43,Appeal_to_Authority,11,0.3157894736842105,0.5454545454545454,0.39999999999999997
+task_agnostic_base,43,Appeal_to_fear-prejudice,24,0.5263157894736842,0.8333333333333334,0.6451612903225806
+task_agnostic_base,43,"Bandwagon,Reductio_ad_hitlerum",7,0.4,0.8571428571428571,0.5454545454545455
+task_agnostic_base,43,Black-and-White_Fallacy,12,0.3125,0.4166666666666667,0.35714285714285715
+task_agnostic_base,43,Causal_Oversimplification,18,0.5517241379310345,0.8888888888888888,0.6808510638297872
+task_agnostic_base,43,Doubt,22,0.4888888888888889,1.0,0.6567164179104478
+task_agnostic_base,43,"Exaggeration,Minimisation",22,0.4523809523809524,0.8636363636363636,0.59375
+task_agnostic_base,43,Flag-Waving,17,0.5,0.8235294117647058,0.6222222222222222
+task_agnostic_base,43,Loaded_Language,45,0.8301886792452831,0.9777777777777777,0.8979591836734695
+task_agnostic_base,43,"Name_Calling,Labeling",33,0.7380952380952381,0.9393939393939394,0.8266666666666667
+task_agnostic_base,43,Repetition,23,0.5161290322580645,0.6956521739130435,0.5925925925925926
+task_agnostic_base,43,Slogans,14,0.38461538461538464,0.7142857142857143,0.5
+task_agnostic_base,43,Thought-terminating_Cliches,7,0.5,0.2857142857142857,0.36363636363636365
+task_agnostic_base,43,"Whataboutism,Straw_Men,Red_Herring",10,0.3333333333333333,0.3,0.3157894736842105
+structural_copy_control,41,Appeal_to_Authority,11,0.2962962962962963,0.7272727272727273,0.42105263157894735
+structural_copy_control,41,Appeal_to_fear-prejudice,24,0.4666666666666667,0.875,0.608695652173913
+structural_copy_control,41,"Bandwagon,Reductio_ad_hitlerum",7,0.2857142857142857,0.8571428571428571,0.42857142857142855
+structural_copy_control,41,Black-and-White_Fallacy,12,0.2222222222222222,0.3333333333333333,0.26666666666666666
+structural_copy_control,41,Causal_Oversimplification,18,0.43243243243243246,0.8888888888888888,0.5818181818181818
+structural_copy_control,41,Doubt,22,0.4,1.0,0.5714285714285715
+structural_copy_control,41,"Exaggeration,Minimisation",22,0.425531914893617,0.9090909090909091,0.5797101449275361
+structural_copy_control,41,Flag-Waving,17,0.37777777777777777,1.0,0.5483870967741935
+structural_copy_control,41,Loaded_Language,45,0.8333333333333334,1.0,0.9090909090909091
+structural_copy_control,41,"Name_Calling,Labeling",33,0.6,1.0,0.7499999999999999
+structural_copy_control,41,Repetition,23,0.46,1.0,0.6301369863013699
+structural_copy_control,41,Slogans,14,0.2857142857142857,0.5714285714285714,0.38095238095238093
+structural_copy_control,41,Thought-terminating_Cliches,7,0.125,0.5714285714285714,0.20512820512820512
+structural_copy_control,41,"Whataboutism,Straw_Men,Red_Herring",10,0.22857142857142856,0.8,0.3555555555555555
+structural_copy_control,42,Appeal_to_Authority,11,0.1875,0.2727272727272727,0.2222222222222222
+structural_copy_control,42,Appeal_to_fear-prejudice,24,0.5384615384615384,0.875,0.6666666666666667
+structural_copy_control,42,"Bandwagon,Reductio_ad_hitlerum",7,0.35,1.0,0.5185185185185185
+structural_copy_control,42,Black-and-White_Fallacy,12,0.38461538461538464,0.8333333333333334,0.5263157894736842
+structural_copy_control,42,Causal_Oversimplification,18,0.4473684210526316,0.9444444444444444,0.6071428571428572
+structural_copy_control,42,Doubt,22,0.4583333333333333,1.0,0.6285714285714286
+structural_copy_control,42,"Exaggeration,Minimisation",22,0.46511627906976744,0.9090909090909091,0.6153846153846153
+structural_copy_control,42,Flag-Waving,17,0.43333333333333335,0.7647058823529411,0.5531914893617021
+structural_copy_control,42,Loaded_Language,45,0.8490566037735849,1.0,0.9183673469387755
+structural_copy_control,42,"Name_Calling,Labeling",33,0.6666666666666666,0.9696969696969697,0.7901234567901234
+structural_copy_control,42,Repetition,23,0.43137254901960786,0.9565217391304348,0.5945945945945946
+structural_copy_control,42,Slogans,14,0.3888888888888889,0.5,0.43750000000000006
+structural_copy_control,42,Thought-terminating_Cliches,7,0.11764705882352941,0.2857142857142857,0.16666666666666666
+structural_copy_control,42,"Whataboutism,Straw_Men,Red_Herring",10,0.16129032258064516,0.5,0.24390243902439024
+structural_copy_control,43,Appeal_to_Authority,11,0.2962962962962963,0.7272727272727273,0.42105263157894735
+structural_copy_control,43,Appeal_to_fear-prejudice,24,0.5263157894736842,0.8333333333333334,0.6451612903225806
+structural_copy_control,43,"Bandwagon,Reductio_ad_hitlerum",7,0.22727272727272727,0.7142857142857143,0.3448275862068965
+structural_copy_control,43,Black-and-White_Fallacy,12,0.3333333333333333,0.4166666666666667,0.3703703703703704
+structural_copy_control,43,Causal_Oversimplification,18,0.4594594594594595,0.9444444444444444,0.6181818181818182
+structural_copy_control,43,Doubt,22,0.4782608695652174,1.0,0.6470588235294118
+structural_copy_control,43,"Exaggeration,Minimisation",22,0.4186046511627907,0.8181818181818182,0.5538461538461539
+structural_copy_control,43,Flag-Waving,17,0.3333333333333333,0.6470588235294118,0.44
+structural_copy_control,43,Loaded_Language,45,0.8653846153846154,1.0,0.9278350515463918
+structural_copy_control,43,"Name_Calling,Labeling",33,0.6458333333333334,0.9393939393939394,0.7654320987654321
+structural_copy_control,43,Repetition,23,0.45454545454545453,0.8695652173913043,0.5970149253731344
+structural_copy_control,43,Slogans,14,0.3076923076923077,0.2857142857142857,0.29629629629629634
+structural_copy_control,43,Thought-terminating_Cliches,7,0.1,0.2857142857142857,0.14814814814814817
+structural_copy_control,43,"Whataboutism,Straw_Men,Red_Herring",10,0.19047619047619047,0.4,0.25806451612903225
+existing_specialized_separate_lineage,41,Appeal_to_Authority,11,0.5,0.6363636363636364,0.56
+existing_specialized_separate_lineage,41,Appeal_to_fear-prejudice,24,0.5,0.9166666666666666,0.6470588235294118
+existing_specialized_separate_lineage,41,"Bandwagon,Reductio_ad_hitlerum",7,0.4444444444444444,0.5714285714285714,0.5
+existing_specialized_separate_lineage,41,Black-and-White_Fallacy,12,0.4666666666666667,0.5833333333333334,0.5185185185185186
+existing_specialized_separate_lineage,41,Causal_Oversimplification,18,0.46153846153846156,0.6666666666666666,0.5454545454545455
+existing_specialized_separate_lineage,41,Doubt,22,0.5238095238095238,1.0,0.6875000000000001
+existing_specialized_separate_lineage,41,"Exaggeration,Minimisation",22,0.4418604651162791,0.8636363636363636,0.5846153846153846
+existing_specialized_separate_lineage,41,Flag-Waving,17,0.53125,1.0,0.6938775510204082
+existing_specialized_separate_lineage,41,Loaded_Language,45,0.8490566037735849,1.0,0.9183673469387755
+existing_specialized_separate_lineage,41,"Name_Calling,Labeling",33,0.717391304347826,1.0,0.8354430379746834
+existing_specialized_separate_lineage,41,Repetition,23,0.46938775510204084,1.0,0.6388888888888888
+existing_specialized_separate_lineage,41,Slogans,14,0.5,0.5,0.5
+existing_specialized_separate_lineage,41,Thought-terminating_Cliches,7,0.14285714285714285,0.5714285714285714,0.2285714285714286
+existing_specialized_separate_lineage,41,"Whataboutism,Straw_Men,Red_Herring",10,0.25,0.4,0.3076923076923077
+existing_specialized_separate_lineage,42,Appeal_to_Authority,11,0.3225806451612903,0.9090909090909091,0.4761904761904761
+existing_specialized_separate_lineage,42,Appeal_to_fear-prejudice,24,0.5405405405405406,0.8333333333333334,0.6557377049180328
+existing_specialized_separate_lineage,42,"Bandwagon,Reductio_ad_hitlerum",7,0.42857142857142855,0.8571428571428571,0.5714285714285714
+existing_specialized_separate_lineage,42,Black-and-White_Fallacy,12,0.47058823529411764,0.6666666666666666,0.5517241379310345
+existing_specialized_separate_lineage,42,Causal_Oversimplification,18,0.4444444444444444,0.8888888888888888,0.5925925925925926
+existing_specialized_separate_lineage,42,Doubt,22,0.5116279069767442,1.0,0.676923076923077
+existing_specialized_separate_lineage,42,"Exaggeration,Minimisation",22,0.4375,0.9545454545454546,0.6
+existing_specialized_separate_lineage,42,Flag-Waving,17,0.6,0.8823529411764706,0.7142857142857143
+existing_specialized_separate_lineage,42,Loaded_Language,45,0.8181818181818182,1.0,0.9
+existing_specialized_separate_lineage,42,"Name_Calling,Labeling",33,0.6808510638297872,0.9696969696969697,0.7999999999999999
+existing_specialized_separate_lineage,42,Repetition,23,0.4888888888888889,0.9565217391304348,0.6470588235294117
+existing_specialized_separate_lineage,42,Slogans,14,0.39285714285714285,0.7857142857142857,0.5238095238095237
+existing_specialized_separate_lineage,42,Thought-terminating_Cliches,7,0.2,0.8571428571428571,0.32432432432432434
+existing_specialized_separate_lineage,42,"Whataboutism,Straw_Men,Red_Herring",10,0.25,0.4,0.3076923076923077
+existing_specialized_separate_lineage,43,Appeal_to_Authority,11,0.3125,0.45454545454545453,0.3703703703703703
+existing_specialized_separate_lineage,43,Appeal_to_fear-prejudice,24,0.5365853658536586,0.9166666666666666,0.676923076923077
+existing_specialized_separate_lineage,43,"Bandwagon,Reductio_ad_hitlerum",7,0.46153846153846156,0.8571428571428571,0.6
+existing_specialized_separate_lineage,43,Black-and-White_Fallacy,12,0.46153846153846156,0.5,0.48000000000000004
+existing_specialized_separate_lineage,43,Causal_Oversimplification,18,0.4358974358974359,0.9444444444444444,0.5964912280701755
+existing_specialized_separate_lineage,43,Doubt,22,0.4782608695652174,1.0,0.6470588235294118
+existing_specialized_separate_lineage,43,"Exaggeration,Minimisation",22,0.4583333333333333,1.0,0.6285714285714286
+existing_specialized_separate_lineage,43,Flag-Waving,17,0.5416666666666666,0.7647058823529411,0.6341463414634146
+existing_specialized_separate_lineage,43,Loaded_Language,45,0.8490566037735849,1.0,0.9183673469387755
+existing_specialized_separate_lineage,43,"Name_Calling,Labeling",33,0.6888888888888889,0.9393939393939394,0.7948717948717948
+existing_specialized_separate_lineage,43,Repetition,23,0.47619047619047616,0.8695652173913043,0.6153846153846153
+existing_specialized_separate_lineage,43,Slogans,14,0.4074074074074074,0.7857142857142857,0.5365853658536585
+existing_specialized_separate_lineage,43,Thought-terminating_Cliches,7,0.21052631578947367,0.5714285714285714,0.3076923076923077
+existing_specialized_separate_lineage,43,"Whataboutism,Straw_Men,Red_Herring",10,0.3333333333333333,0.4,0.3636363636363636

benchmark/per_label_summary.csv

@@ -0,0 +1,43 @@
+arm,label,support,precision_mean,precision_sample_sd,recall_mean,recall_sample_sd,f1_mean,f1_sample_sd
+task_agnostic_base,Appeal_to_Authority,11,0.3958614484930274,0.1238064457010119,0.5757575757575757,0.05248638810814781,0.4657088122605364,0.1020990126979069
+task_agnostic_base,Appeal_to_fear-prejudice,24,0.5540244550770866,0.02525673990402998,0.7361111111111112,0.13393959390267993,0.6277521761392729,0.04994879016183189
+task_agnostic_base,"Bandwagon,Reductio_ad_hitlerum",7,0.4851851851851852,0.07882887649552875,0.8095238095238095,0.08247860988423221,0.6006778309409889,0.04793776258918466
+task_agnostic_base,Black-and-White_Fallacy,12,0.3667929292929293,0.07670649392711647,0.47222222222222227,0.09622504486493764,0.4053892966936445,0.04211369218048746
+task_agnostic_base,Causal_Oversimplification,18,0.5428824049513704,0.039216350192122156,0.7962962962962963,0.11564811108145183,0.6446992723588468,0.0634561308339175
+task_agnostic_base,Doubt,22,0.5223832528180354,0.06742780050556177,0.9848484848484849,0.026243194054073875,0.6802057822343391,0.04928561754686331
+task_agnostic_base,"Exaggeration,Minimisation",22,0.4620811287477954,0.016801198309751,0.7727272727272727,0.15745916432444337,0.5727040816326531,0.03645259990419198
+task_agnostic_base,Flag-Waving,17,0.5642642642642642,0.1478145677656334,0.8235294117647058,0.1764705882352941,0.6464506172839506,0.035742219952758225
+task_agnostic_base,Loaded_Language,45,0.8649032907793016,0.046984312428427724,0.9851851851851852,0.012830005981991702,0.9208741623431562,0.03200926878737141
+task_agnostic_base,"Name_Calling,Labeling",33,0.7380952380952381,0.0047619047619048005,0.9090909090909091,0.10925912955951485,0.8125087983911513,0.04252971464057227
+task_agnostic_base,Repetition,23,0.5196620583717357,0.021645909050194662,0.7536231884057971,0.06641414050660639,0.6147520781519571,0.03505927109416221
+task_agnostic_base,Slogans,14,0.5161361626878869,0.20305677930281646,0.7380952380952381,0.10910894511799614,0.5834824090638044,0.09862686289220693
+task_agnostic_base,Thought-terminating_Cliches,7,0.3492063492063492,0.143516994603374,0.3333333333333333,0.08247860988423225,0.319014319014319,0.04017590711117064
+task_agnostic_base,"Whataboutism,Straw_Men,Red_Herring",10,0.2761904761904762,0.05302632726504782,0.5,0.2645751311064591,0.330448343079922,0.021845164720973014
+structural_copy_control,Appeal_to_Authority,11,0.26003086419753085,0.06281357095350094,0.5757575757575758,0.262431940540739,0.35477582846003897,0.11479479036518875
+structural_copy_control,Appeal_to_fear-prejudice,24,0.5104813315339631,0.038427508996734555,0.8611111111111112,0.024056261216234387,0.6401745363877202,0.029305467347150463
+structural_copy_control,"Bandwagon,Reductio_ad_hitlerum",7,0.2876623376623376,0.061386823077618426,0.8571428571428571,0.14285714285714285,0.4306391777656145,0.08686392624022733
+structural_copy_control,Black-and-White_Fallacy,12,0.31339031339031337,0.08301311842925053,0.5277777777777778,0.2678791878053599,0.38778427550357375,0.13069755090515112
+structural_copy_control,Causal_Oversimplification,18,0.4464201043148412,0.013538446216471536,0.9259259259259259,0.03207501495497923,0.6023809523809524,0.018643640071557464
+structural_copy_control,Doubt,22,0.4455314009661836,0.04067072694857813,1.0,0.0,0.615686274509804,0.039427200694482675
+structural_copy_control,"Exaggeration,Minimisation",22,0.4364176150420584,0.025093958106097894,0.8787878787878788,0.05248638810814775,0.5829803047194351,0.030899288356345502
+structural_copy_control,Flag-Waving,17,0.3814814814814815,0.05010277503136551,0.803921568627451,0.17970885078258195,0.5138595287119652,0.06400931989326974
+structural_copy_control,Loaded_Language,45,0.8492581841638446,0.016026591848331686,1.0,0.0,0.9184311025253588,0.009372233868132662
+structural_copy_control,"Name_Calling,Labeling",33,0.6375,0.03410563654946854,0.9696969696969697,0.030303030303030276,0.7685185185185185,0.02023900779105559
+structural_copy_control,Repetition,23,0.44863933452168747,0.01520014564765896,0.9420289855072463,0.06641414050660639,0.607248835423033,0.019858627551810427
+structural_copy_control,Slogans,14,0.3274318274318274,0.05434598572773684,0.4523809523809524,0.1486904285332952,0.3715828924162258,0.07106660236343577
+structural_copy_control,Thought-terminating_Cliches,7,0.1142156862745098,0.012848374923608973,0.38095238095238093,0.1649572197684645,0.17331433998100665,0.029065880233243847
+structural_copy_control,"Whataboutism,Straw_Men,Red_Herring",10,0.19344598054275472,0.03373872471827226,0.5666666666666667,0.2081665999466133,0.28584083690299267,0.06078854830133676
+existing_specialized_separate_lineage,Appeal_to_Authority,11,0.37836021505376344,0.10546365636105155,0.6666666666666666,0.2287828616748712,0.46885361552028215,0.09502747629890008
+existing_specialized_separate_lineage,Appeal_to_fear-prejudice,24,0.525708635464733,0.02235198659615819,0.8888888888888888,0.048112522432468774,0.6599065351235072,0.015362381542320084
+existing_specialized_separate_lineage,"Bandwagon,Reductio_ad_hitlerum",7,0.44485144485144484,0.016487284575115283,0.7619047619047619,0.1649572197684645,0.5571428571428572,0.051507875363771265
+existing_specialized_separate_lineage,Black-and-White_Fallacy,12,0.4662644544997486,0.004538274146363688,0.5833333333333334,0.08333333333333331,0.516747552149851,0.03589484965697345
+existing_specialized_separate_lineage,Causal_Oversimplification,18,0.4472934472934473,0.013055771210700413,0.8333333333333333,0.1469861839480328,0.5781794553724379,0.028407563015813122
+existing_specialized_separate_lineage,Doubt,22,0.5045661001171619,0.023581174969843577,1.0,0.0,0.6704939668174963,0.020973132363659372
+existing_specialized_separate_lineage,"Exaggeration,Minimisation",22,0.44589793281653745,0.010987847993006481,0.9393939393939394,0.06943296507508849,0.6043956043956044,0.022305256187015846
+existing_specialized_separate_lineage,Flag-Waving,17,0.5576388888888889,0.03705367103320583,0.8823529411764706,0.11764705882352944,0.680769868923179,0.04164658769406446
+existing_specialized_separate_lineage,Loaded_Language,45,0.8387650085763294,0.01782556577257849,1.0,0.0,0.9122448979591837,0.01060439269940129
+existing_specialized_separate_lineage,"Name_Calling,Labeling",33,0.6957104190221673,0.019201490809269257,0.9696969696969697,0.030303030303030276,0.8101049442821594,0.022092733328802537
+existing_specialized_separate_lineage,Repetition,23,0.4781557067271353,0.009897987322673213,0.9420289855072463,0.06641414050660639,0.6337774426009719,0.016444118333485254
+existing_specialized_separate_lineage,Slogans,14,0.4334215167548501,0.05811581861282867,0.6904761904761905,0.1649572197684645,0.5201316298877274,0.018567913895656114
+existing_specialized_separate_lineage,Thought-terminating_Cliches,7,0.18446115288220552,0.03641251252864963,0.6666666666666666,0.1649572197684645,0.2868626868626869,0.0511620864429468
+existing_specialized_separate_lineage,"Whataboutism,Straw_Men,Red_Herring",10,0.2777777777777778,0.0481125224324688,0.4,0.0,0.32634032634032634,0.03229931575886014

benchmark/resource_metrics.csv

@@ -0,0 +1,10 @@
+arm,seed,parameters,trainable_parameters,peak_cuda_allocated_bytes,peak_cuda_reserved_bytes,total_wall_seconds,test_wall_seconds,test_seconds_per_article,device,dtype
+task_agnostic_base,41,334111040,334111040,3380975616,3519021056,1583.0297977919981,4.490653285000008,0.08164824154545468,NVIDIA GeForce RTX 5070 Ti,bfloat16
+task_agnostic_base,42,334111040,334111040,3380975616,3519021056,1588.617994428001,4.329529417009326,0.07871871667289683,NVIDIA GeForce RTX 5070 Ti,bfloat16
+task_agnostic_base,43,334111040,334111040,3380975616,3510632448,1603.4310500120046,4.595363147003809,0.08355205721825107,NVIDIA GeForce RTX 5070 Ti,bfloat16
+structural_copy_control,41,334111040,334111040,3380975616,3519021056,1575.5905925639963,4.332684432010865,0.07877608058201573,NVIDIA GeForce RTX 5070 Ti,bfloat16
+structural_copy_control,42,334111040,334111040,3380975616,3519021056,1623.5639530190092,4.569082793022972,0.08307423260041767,NVIDIA GeForce RTX 5070 Ti,bfloat16
+structural_copy_control,43,334111040,334111040,3380975616,3510632448,1629.195114912989,4.4899811029899865,0.08163602005436339,NVIDIA GeForce RTX 5070 Ti,bfloat16
+existing_specialized_separate_lineage,41,210903360,210903360,2143655936,2281701376,1634.488557530014,4.6637764379847795,0.0847959352360869,NVIDIA GeForce RTX 5070 Ti,bfloat16
+existing_specialized_separate_lineage,42,210903360,210903360,2143655936,2281701376,1652.9397718419787,4.728947311989032,0.08598086021798239,NVIDIA GeForce RTX 5070 Ti,bfloat16
+existing_specialized_separate_lineage,43,210903360,210903360,2143655936,2273312768,1643.488202479988,4.704143202980049,0.08552987641781908,NVIDIA GeForce RTX 5070 Ti,bfloat16

benchmark/seed_metrics.csv

@@ -0,0 +1,10 @@
+arm,seed,threshold,macro_f1,micro_f1,exact_match,predicted_positive_rate,tp,fp,fn,test_articles,test_windows
+task_agnostic_base,41,0.15,0.5883093293517536,0.6468085106382978,0.05454545454545454,0.5714285714285714,228,212,37,55,434
+task_agnostic_base,42,0.3,0.6028379036218096,0.6655172413793103,0.03636363636363636,0.4090909090909091,193,122,72,55,434
+task_agnostic_base,43,0.1,0.5712816197954101,0.6465256797583082,0.0,0.5155844155844156,214,183,51,55,434
+structural_copy_control,41,0.05,0.5169424579262757,0.5773955773955775,0.0,0.712987012987013,235,314,30,55,434
+structural_copy_control,42,0.2,0.5349405779540175,0.6083445491251682,0.0,0.6207792207792208,226,252,39,55,434
+structural_copy_control,43,0.2,0.5023778364496152,0.585635359116022,0.0,0.5961038961038961,212,247,53,55,434
+existing_specialized_separate_lineage,41,0.1,0.5832848452288824,0.6494252873563218,0.0,0.5597402597402598,226,205,39,55,434
+existing_specialized_separate_lineage,42,0.05,0.5958405181160762,0.6458616010854816,0.0,0.612987012987013,238,234,27,55,434
+existing_specialized_separate_lineage,43,0.05,0.5835785045218137,0.6486486486486486,0.01818181818181818,0.5688311688311688,228,210,37,55,434

benchmark/semeval-transfer-summary.json

@@ -1,47 +1,891 @@
 {
-  "schema": "standalone4l-transfer-summary-v1",
-  "claim_scope": "Exploratory one-task transfer evidence on a previously opened 55-article SemEval-derived test split; not confirmatory.",
-  "metric": "article-level multilabel Macro-F1",
+  "schema": "standalone4l-complete-transfer-benchmark-v2",
+  "status": "complete_from_frozen_reports",
+  "claim_scope": "Exploratory fresh-head transfer evidence on one English task and a previously opened 55-article SemEval-derived test split.",
+  "headless_root_requires_adaptation": true,
+  "test_previously_opened": true,
+  "test_articles": 55,
+  "test_windows": 434,
   "seeds": [
     41,
     42,
     43
   ],
-  "task_agnostic_base": {
-    "values": [
-      0.5883093293517536,
-      0.6028379036218096,
-      0.5712816197954101
-    ],
-    "mean": 0.5874762842563245,
-    "sample_sd": 0.01579462678670608
+  "three_seed_sd_is_not_confidence_interval": true,
+  "multilingual_downstream_evaluation_performed": false,
+  "upstream_multilingual_metadata": {
+    "source": "Qwen/Qwen3.5-0.8B official model card",
+    "languages_and_dialects_claimed_upstream": 201,
+    "scope": "inherited tokenizer and architecture coverage only"
   },
-  "structural_copy_control": {
-    "values": [
-      0.5169424579262757,
-      0.5349405779540175,
-      0.5023778364496152
+  "protocol": {
+    "epochs": 5,
+    "primary_epoch": 5,
+    "early_stopping": false,
+    "max_length": 256,
+    "stride": 128,
+    "window_target": "span-aligned technique labels",
+    "article_aggregation": "labelwise maximum probability",
+    "micro_batch_size": 1,
+    "gradient_accumulation": 32,
+    "effective_batch_size": 32,
+    "learning_rate": 2e-05,
+    "weight_decay": 0.01,
+    "gradient_clip": 1.0,
+    "optimizer": "AdamW",
+    "dtype": "bfloat16",
+    "thresholds": [
+      0.05,
+      0.1,
+      0.15,
+      0.2,
+      0.25,
+      0.3,
+      0.35,
+      0.4,
+      0.45,
+      0.5,
+      0.55,
+      0.6,
+      0.65,
+      0.7,
+      0.75,
+      0.8,
+      0.85,
+      0.9
     ],
-    "mean": 0.5180869574433028,
-    "sample_sd": 0.016311512588628545
+    "threshold_selection": "epoch-5 calibration macro-F1, then micro-F1"
   },
-  "same_seed_deltas": [
+  "arms": {
+    "task_agnostic_base": {
+      "display": "Task-agnostic KD base",
+      "lineage": "root base plus a fresh 14-label transfer head",
+      "metrics": {
+        "macro_f1": {
+          "values": [
+            0.5883093293517536,
+            0.6028379036218096,
+            0.5712816197954101
+          ],
+          "mean": 0.5874762842563245,
+          "sample_sd": 0.01579462678670608,
+          "n": 3
+        },
+        "micro_f1": {
+          "values": [
+            0.6468085106382978,
+            0.6655172413793103,
+            0.6465256797583082
+          ],
+          "mean": 0.6529504772586388,
+          "sample_sd": 0.010884055708931178,
+          "n": 3
+        },
+        "exact_match": {
+          "values": [
+            0.05454545454545454,
+            0.03636363636363636,
+            0.0
+          ],
+          "mean": 0.0303030303030303,
+          "sample_sd": 0.027773186030035393,
+          "n": 3
+        },
+        "predicted_positive_rate": {
+          "values": [
+            0.5714285714285714,
+            0.4090909090909091,
+            0.5155844155844156
+          ],
+          "mean": 0.4987012987012987,
+          "sample_sd": 0.08247520160679181,
+          "n": 3
+        }
+      },
+      "resources": {
+        "parameters": {
+          "values": [
+            334111040.0,
+            334111040.0,
+            334111040.0
+          ],
+          "mean": 334111040.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_allocated_bytes": {
+          "values": [
+            3380975616.0,
+            3380975616.0,
+            3380975616.0
+          ],
+          "mean": 3380975616.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_reserved_bytes": {
+          "values": [
+            3519021056.0,
+            3519021056.0,
+            3510632448.0
+          ],
+          "mean": 3516224853.3333335,
+          "sample_sd": 4843165.086926248,
+          "n": 3
+        },
+        "total_wall_seconds": {
+          "values": [
+            1583.0297977919981,
+            1588.617994428001,
+            1603.4310500120046
+          ],
+          "mean": 1591.692947410668,
+          "sample_sd": 10.54249851412311,
+          "n": 3
+        },
+        "test_seconds_per_article": {
+          "values": [
+            0.08164824154545468,
+            0.07871871667289683,
+            0.08355205721825107
+          ],
+          "mean": 0.08130633847886752,
+          "sample_sd": 0.0024347419754560395,
+          "n": 3
+        }
+      },
+      "thresholds": [
+        0.15,
+        0.3,
+        0.1
+      ]
+    },
+    "structural_copy_control": {
+      "display": "Structural copy, no task-free KD",
+      "lineage": "same four-layer structure without task-free KD",
+      "metrics": {
+        "macro_f1": {
+          "values": [
+            0.5169424579262757,
+            0.5349405779540175,
+            0.5023778364496152
+          ],
+          "mean": 0.5180869574433028,
+          "sample_sd": 0.016311512588628545,
+          "n": 3
+        },
+        "micro_f1": {
+          "values": [
+            0.5773955773955775,
+            0.6083445491251682,
+            0.585635359116022
+          ],
+          "mean": 0.5904584952122559,
+          "sample_sd": 0.01602830914762526,
+          "n": 3
+        },
+        "exact_match": {
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ],
+          "mean": 0.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "predicted_positive_rate": {
+          "values": [
+            0.712987012987013,
+            0.6207792207792208,
+            0.5961038961038961
+          ],
+          "mean": 0.6432900432900432,
+          "sample_sd": 0.06160737456395461,
+          "n": 3
+        }
+      },
+      "resources": {
+        "parameters": {
+          "values": [
+            334111040.0,
+            334111040.0,
+            334111040.0
+          ],
+          "mean": 334111040.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_allocated_bytes": {
+          "values": [
+            3380975616.0,
+            3380975616.0,
+            3380975616.0
+          ],
+          "mean": 3380975616.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_reserved_bytes": {
+          "values": [
+            3519021056.0,
+            3519021056.0,
+            3510632448.0
+          ],
+          "mean": 3516224853.3333335,
+          "sample_sd": 4843165.086926248,
+          "n": 3
+        },
+        "total_wall_seconds": {
+          "values": [
+            1575.5905925639963,
+            1623.5639530190092,
+            1629.195114912989
+          ],
+          "mean": 1609.4498868319981,
+          "sample_sd": 29.457874198835384,
+          "n": 3
+        },
+        "test_seconds_per_article": {
+          "values": [
+            0.07877608058201573,
+            0.08307423260041767,
+            0.08163602005436339
+          ],
+          "mean": 0.08116211107893226,
+          "sample_sd": 0.002187914527825257,
+          "n": 3
+        }
+      },
+      "thresholds": [
+        0.05,
+        0.2,
+        0.2
+      ]
+    },
+    "existing_specialized_separate_lineage": {
+      "display": "Existing specialized 4L, separate lineage",
+      "lineage": "historical task-specific shrink/distillation lineage",
+      "metrics": {
+        "macro_f1": {
+          "values": [
+            0.5832848452288824,
+            0.5958405181160762,
+            0.5835785045218137
+          ],
+          "mean": 0.5875679559555907,
+          "sample_sd": 0.007165753447433348,
+          "n": 3
+        },
+        "micro_f1": {
+          "values": [
+            0.6494252873563218,
+            0.6458616010854816,
+            0.6486486486486486
+          ],
+          "mean": 0.647978512363484,
+          "sample_sd": 0.0018739735696546695,
+          "n": 3
+        },
+        "exact_match": {
+          "values": [
+            0.0,
+            0.0,
+            0.01818181818181818
+          ],
+          "mean": 0.006060606060606061,
+          "sample_sd": 0.01049727762162956,
+          "n": 3
+        },
+        "predicted_positive_rate": {
+          "values": [
+            0.5597402597402598,
+            0.612987012987013,
+            0.5688311688311688
+          ],
+          "mean": 0.5805194805194805,
+          "sample_sd": 0.02848274311618351,
+          "n": 3
+        }
+      },
+      "resources": {
+        "parameters": {
+          "values": [
+            210903360.0,
+            210903360.0,
+            210903360.0
+          ],
+          "mean": 210903360.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_allocated_bytes": {
+          "values": [
+            2143655936.0,
+            2143655936.0,
+            2143655936.0
+          ],
+          "mean": 2143655936.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_reserved_bytes": {
+          "values": [
+            2281701376.0,
+            2281701376.0,
+            2273312768.0
+          ],
+          "mean": 2278905173.3333335,
+          "sample_sd": 4843165.086926248,
+          "n": 3
+        },
+        "total_wall_seconds": {
+          "values": [
+            1634.488557530014,
+            1652.9397718419787,
+            1643.488202479988
+          ],
+          "mean": 1643.6388439506602,
+          "sample_sd": 9.226529522850198,
+          "n": 3
+        },
+        "test_seconds_per_article": {
+          "values": [
+            0.0847959352360869,
+            0.08598086021798239,
+            0.08552987641781908
+          ],
+          "mean": 0.08543555729062946,
+          "sample_sd": 0.0005980667826383295,
+          "n": 3
+        }
+      },
+      "thresholds": [
+        0.1,
+        0.05,
+        0.05
+      ]
+    }
+  },
+  "same_seed_macro_f1_deltas_base_minus_control": [
     0.07136687142547793,
     0.06789732566779205,
     0.06890378334579494
   ],
-  "mean_delta": 0.06938932681302164,
+  "mean_macro_f1_delta_base_minus_control": 0.06938932681302164,
   "publication_gate_passed": true,
-  "existing_specialized_separate_lineage": {
-    "values": [
-      0.5832848452288824,
-      0.5958405181160762,
-      0.5835785045218137
-    ],
-    "mean": 0.5875679559555907,
-    "sample_sd": 0.007165593191426643
-  },
-  "test_articles": 55,
-  "test_previously_opened": true,
-  "three_seed_sd_is_not_confidence_interval": true
+  "per_label_summary": [
+    {
+      "arm": "task_agnostic_base",
+      "label": "Appeal_to_Authority",
+      "support": 11,
+      "precision_mean": 0.3958614484930274,
+      "precision_sample_sd": 0.1238064457010119,
+      "recall_mean": 0.5757575757575757,
+      "recall_sample_sd": 0.05248638810814781,
+      "f1_mean": 0.4657088122605364,
+      "f1_sample_sd": 0.1020990126979069
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Appeal_to_fear-prejudice",
+      "support": 24,
+      "precision_mean": 0.5540244550770866,
+      "precision_sample_sd": 0.02525673990402998,
+      "recall_mean": 0.7361111111111112,
+      "recall_sample_sd": 0.13393959390267993,
+      "f1_mean": 0.6277521761392729,
+      "f1_sample_sd": 0.04994879016183189
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Bandwagon,Reductio_ad_hitlerum",
+      "support": 7,
+      "precision_mean": 0.4851851851851852,
+      "precision_sample_sd": 0.07882887649552875,
+      "recall_mean": 0.8095238095238095,
+      "recall_sample_sd": 0.08247860988423221,
+      "f1_mean": 0.6006778309409889,
+      "f1_sample_sd": 0.04793776258918466
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Black-and-White_Fallacy",
+      "support": 12,
+      "precision_mean": 0.3667929292929293,
+      "precision_sample_sd": 0.07670649392711647,
+      "recall_mean": 0.47222222222222227,
+      "recall_sample_sd": 0.09622504486493764,
+      "f1_mean": 0.4053892966936445,
+      "f1_sample_sd": 0.04211369218048746
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Causal_Oversimplification",
+      "support": 18,
+      "precision_mean": 0.5428824049513704,
+      "precision_sample_sd": 0.039216350192122156,
+      "recall_mean": 0.7962962962962963,
+      "recall_sample_sd": 0.11564811108145183,
+      "f1_mean": 0.6446992723588468,
+      "f1_sample_sd": 0.0634561308339175
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Doubt",
+      "support": 22,
+      "precision_mean": 0.5223832528180354,
+      "precision_sample_sd": 0.06742780050556177,
+      "recall_mean": 0.9848484848484849,
+      "recall_sample_sd": 0.026243194054073875,
+      "f1_mean": 0.6802057822343391,
+      "f1_sample_sd": 0.04928561754686331
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Exaggeration,Minimisation",
+      "support": 22,
+      "precision_mean": 0.4620811287477954,
+      "precision_sample_sd": 0.016801198309751,
+      "recall_mean": 0.7727272727272727,
+      "recall_sample_sd": 0.15745916432444337,
+      "f1_mean": 0.5727040816326531,
+      "f1_sample_sd": 0.03645259990419198
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Flag-Waving",
+      "support": 17,
+      "precision_mean": 0.5642642642642642,
+      "precision_sample_sd": 0.1478145677656334,
+      "recall_mean": 0.8235294117647058,
+      "recall_sample_sd": 0.1764705882352941,
+      "f1_mean": 0.6464506172839506,
+      "f1_sample_sd": 0.035742219952758225
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Loaded_Language",
+      "support": 45,
+      "precision_mean": 0.8649032907793016,
+      "precision_sample_sd": 0.046984312428427724,
+      "recall_mean": 0.9851851851851852,
+      "recall_sample_sd": 0.012830005981991702,
+      "f1_mean": 0.9208741623431562,
+      "f1_sample_sd": 0.03200926878737141
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Name_Calling,Labeling",
+      "support": 33,
+      "precision_mean": 0.7380952380952381,
+      "precision_sample_sd": 0.0047619047619048005,
+      "recall_mean": 0.9090909090909091,
+      "recall_sample_sd": 0.10925912955951485,
+      "f1_mean": 0.8125087983911513,
+      "f1_sample_sd": 0.04252971464057227
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Repetition",
+      "support": 23,
+      "precision_mean": 0.5196620583717357,
+      "precision_sample_sd": 0.021645909050194662,
+      "recall_mean": 0.7536231884057971,
+      "recall_sample_sd": 0.06641414050660639,
+      "f1_mean": 0.6147520781519571,
+      "f1_sample_sd": 0.03505927109416221
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Slogans",
+      "support": 14,
+      "precision_mean": 0.5161361626878869,
+      "precision_sample_sd": 0.20305677930281646,
+      "recall_mean": 0.7380952380952381,
+      "recall_sample_sd": 0.10910894511799614,
+      "f1_mean": 0.5834824090638044,
+      "f1_sample_sd": 0.09862686289220693
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Thought-terminating_Cliches",
+      "support": 7,
+      "precision_mean": 0.3492063492063492,
+      "precision_sample_sd": 0.143516994603374,
+      "recall_mean": 0.3333333333333333,
+      "recall_sample_sd": 0.08247860988423225,
+      "f1_mean": 0.319014319014319,
+      "f1_sample_sd": 0.04017590711117064
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Whataboutism,Straw_Men,Red_Herring",
+      "support": 10,
+      "precision_mean": 0.2761904761904762,
+      "precision_sample_sd": 0.05302632726504782,
+      "recall_mean": 0.5,
+      "recall_sample_sd": 0.2645751311064591,
+      "f1_mean": 0.330448343079922,
+      "f1_sample_sd": 0.021845164720973014
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Appeal_to_Authority",
+      "support": 11,
+      "precision_mean": 0.26003086419753085,
+      "precision_sample_sd": 0.06281357095350094,
+      "recall_mean": 0.5757575757575758,
+      "recall_sample_sd": 0.262431940540739,
+      "f1_mean": 0.35477582846003897,
+      "f1_sample_sd": 0.11479479036518875
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Appeal_to_fear-prejudice",
+      "support": 24,
+      "precision_mean": 0.5104813315339631,
+      "precision_sample_sd": 0.038427508996734555,
+      "recall_mean": 0.8611111111111112,
+      "recall_sample_sd": 0.024056261216234387,
+      "f1_mean": 0.6401745363877202,
+      "f1_sample_sd": 0.029305467347150463
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Bandwagon,Reductio_ad_hitlerum",
+      "support": 7,
+      "precision_mean": 0.2876623376623376,
+      "precision_sample_sd": 0.061386823077618426,
+      "recall_mean": 0.8571428571428571,
+      "recall_sample_sd": 0.14285714285714285,
+      "f1_mean": 0.4306391777656145,
+      "f1_sample_sd": 0.08686392624022733
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Black-and-White_Fallacy",
+      "support": 12,
+      "precision_mean": 0.31339031339031337,
+      "precision_sample_sd": 0.08301311842925053,
+      "recall_mean": 0.5277777777777778,
+      "recall_sample_sd": 0.2678791878053599,
+      "f1_mean": 0.38778427550357375,
+      "f1_sample_sd": 0.13069755090515112
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Causal_Oversimplification",
+      "support": 18,
+      "precision_mean": 0.4464201043148412,
+      "precision_sample_sd": 0.013538446216471536,
+      "recall_mean": 0.9259259259259259,
+      "recall_sample_sd": 0.03207501495497923,
+      "f1_mean": 0.6023809523809524,
+      "f1_sample_sd": 0.018643640071557464
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Doubt",
+      "support": 22,
+      "precision_mean": 0.4455314009661836,
+      "precision_sample_sd": 0.04067072694857813,
+      "recall_mean": 1.0,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.615686274509804,
+      "f1_sample_sd": 0.039427200694482675
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Exaggeration,Minimisation",
+      "support": 22,
+      "precision_mean": 0.4364176150420584,
+      "precision_sample_sd": 0.025093958106097894,
+      "recall_mean": 0.8787878787878788,
+      "recall_sample_sd": 0.05248638810814775,
+      "f1_mean": 0.5829803047194351,
+      "f1_sample_sd": 0.030899288356345502
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Flag-Waving",
+      "support": 17,
+      "precision_mean": 0.3814814814814815,
+      "precision_sample_sd": 0.05010277503136551,
+      "recall_mean": 0.803921568627451,
+      "recall_sample_sd": 0.17970885078258195,
+      "f1_mean": 0.5138595287119652,
+      "f1_sample_sd": 0.06400931989326974
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Loaded_Language",
+      "support": 45,
+      "precision_mean": 0.8492581841638446,
+      "precision_sample_sd": 0.016026591848331686,
+      "recall_mean": 1.0,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.9184311025253588,
+      "f1_sample_sd": 0.009372233868132662
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Name_Calling,Labeling",
+      "support": 33,
+      "precision_mean": 0.6375,
+      "precision_sample_sd": 0.03410563654946854,
+      "recall_mean": 0.9696969696969697,
+      "recall_sample_sd": 0.030303030303030276,
+      "f1_mean": 0.7685185185185185,
+      "f1_sample_sd": 0.02023900779105559
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Repetition",
+      "support": 23,
+      "precision_mean": 0.44863933452168747,
+      "precision_sample_sd": 0.01520014564765896,
+      "recall_mean": 0.9420289855072463,
+      "recall_sample_sd": 0.06641414050660639,
+      "f1_mean": 0.607248835423033,
+      "f1_sample_sd": 0.019858627551810427
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Slogans",
+      "support": 14,
+      "precision_mean": 0.3274318274318274,
+      "precision_sample_sd": 0.05434598572773684,
+      "recall_mean": 0.4523809523809524,
+      "recall_sample_sd": 0.1486904285332952,
+      "f1_mean": 0.3715828924162258,
+      "f1_sample_sd": 0.07106660236343577
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Thought-terminating_Cliches",
+      "support": 7,
+      "precision_mean": 0.1142156862745098,
+      "precision_sample_sd": 0.012848374923608973,
+      "recall_mean": 0.38095238095238093,
+      "recall_sample_sd": 0.1649572197684645,
+      "f1_mean": 0.17331433998100665,
+      "f1_sample_sd": 0.029065880233243847
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Whataboutism,Straw_Men,Red_Herring",
+      "support": 10,
+      "precision_mean": 0.19344598054275472,
+      "precision_sample_sd": 0.03373872471827226,
+      "recall_mean": 0.5666666666666667,
+      "recall_sample_sd": 0.2081665999466133,
+      "f1_mean": 0.28584083690299267,
+      "f1_sample_sd": 0.06078854830133676
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Appeal_to_Authority",
+      "support": 11,
+      "precision_mean": 0.37836021505376344,
+      "precision_sample_sd": 0.10546365636105155,
+      "recall_mean": 0.6666666666666666,
+      "recall_sample_sd": 0.2287828616748712,
+      "f1_mean": 0.46885361552028215,
+      "f1_sample_sd": 0.09502747629890008
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Appeal_to_fear-prejudice",
+      "support": 24,
+      "precision_mean": 0.525708635464733,
+      "precision_sample_sd": 0.02235198659615819,
+      "recall_mean": 0.8888888888888888,
+      "recall_sample_sd": 0.048112522432468774,
+      "f1_mean": 0.6599065351235072,
+      "f1_sample_sd": 0.015362381542320084
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Bandwagon,Reductio_ad_hitlerum",
+      "support": 7,
+      "precision_mean": 0.44485144485144484,
+      "precision_sample_sd": 0.016487284575115283,
+      "recall_mean": 0.7619047619047619,
+      "recall_sample_sd": 0.1649572197684645,
+      "f1_mean": 0.5571428571428572,
+      "f1_sample_sd": 0.051507875363771265
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Black-and-White_Fallacy",
+      "support": 12,
+      "precision_mean": 0.4662644544997486,
+      "precision_sample_sd": 0.004538274146363688,
+      "recall_mean": 0.5833333333333334,
+      "recall_sample_sd": 0.08333333333333331,
+      "f1_mean": 0.516747552149851,
+      "f1_sample_sd": 0.03589484965697345
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Causal_Oversimplification",
+      "support": 18,
+      "precision_mean": 0.4472934472934473,
+      "precision_sample_sd": 0.013055771210700413,
+      "recall_mean": 0.8333333333333333,
+      "recall_sample_sd": 0.1469861839480328,
+      "f1_mean": 0.5781794553724379,
+      "f1_sample_sd": 0.028407563015813122
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Doubt",
+      "support": 22,
+      "precision_mean": 0.5045661001171619,
+      "precision_sample_sd": 0.023581174969843577,
+      "recall_mean": 1.0,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.6704939668174963,
+      "f1_sample_sd": 0.020973132363659372
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Exaggeration,Minimisation",
+      "support": 22,
+      "precision_mean": 0.44589793281653745,
+      "precision_sample_sd": 0.010987847993006481,
+      "recall_mean": 0.9393939393939394,
+      "recall_sample_sd": 0.06943296507508849,
+      "f1_mean": 0.6043956043956044,
+      "f1_sample_sd": 0.022305256187015846
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Flag-Waving",
+      "support": 17,
+      "precision_mean": 0.5576388888888889,
+      "precision_sample_sd": 0.03705367103320583,
+      "recall_mean": 0.8823529411764706,
+      "recall_sample_sd": 0.11764705882352944,
+      "f1_mean": 0.680769868923179,
+      "f1_sample_sd": 0.04164658769406446
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Loaded_Language",
+      "support": 45,
+      "precision_mean": 0.8387650085763294,
+      "precision_sample_sd": 0.01782556577257849,
+      "recall_mean": 1.0,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.9122448979591837,
+      "f1_sample_sd": 0.01060439269940129
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Name_Calling,Labeling",
+      "support": 33,
+      "precision_mean": 0.6957104190221673,
+      "precision_sample_sd": 0.019201490809269257,
+      "recall_mean": 0.9696969696969697,
+      "recall_sample_sd": 0.030303030303030276,
+      "f1_mean": 0.8101049442821594,
+      "f1_sample_sd": 0.022092733328802537
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Repetition",
+      "support": 23,
+      "precision_mean": 0.4781557067271353,
+      "precision_sample_sd": 0.009897987322673213,
+      "recall_mean": 0.9420289855072463,
+      "recall_sample_sd": 0.06641414050660639,
+      "f1_mean": 0.6337774426009719,
+      "f1_sample_sd": 0.016444118333485254
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Slogans",
+      "support": 14,
+      "precision_mean": 0.4334215167548501,
+      "precision_sample_sd": 0.05811581861282867,
+      "recall_mean": 0.6904761904761905,
+      "recall_sample_sd": 0.1649572197684645,
+      "f1_mean": 0.5201316298877274,
+      "f1_sample_sd": 0.018567913895656114
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Thought-terminating_Cliches",
+      "support": 7,
+      "precision_mean": 0.18446115288220552,
+      "precision_sample_sd": 0.03641251252864963,
+      "recall_mean": 0.6666666666666666,
+      "recall_sample_sd": 0.1649572197684645,
+      "f1_mean": 0.2868626868626869,
+      "f1_sample_sd": 0.0511620864429468
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Whataboutism,Straw_Men,Red_Herring",
+      "support": 10,
+      "precision_mean": 0.2777777777777778,
+      "precision_sample_sd": 0.0481125224324688,
+      "recall_mean": 0.4,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.32634032634032634,
+      "f1_sample_sd": 0.03229931575886014
+    }
+  ],
+  "compression_ladder": [
+    {
+      "name": "24L upstream text teacher",
+      "layers": 24,
+      "parameters": 752393024,
+      "weight_bytes": 1504825632,
+      "weight_mib": 1435.1135559082031,
+      "weight_sha256": "30783da4651259c26b8a204821ea426b9526431e0c14da0d1c6fa284e7414e5e",
+      "kd_stage": "",
+      "mean_total_loss": "",
+      "mean_interface_loss": "",
+      "mean_final_loss": ""
+    },
+    {
+      "name": "8L task-free KD",
+      "layers": 8,
+      "parameters": 420318144,
+      "weight_bytes": 840647880,
+      "weight_mib": 801.7042922973633,
+      "weight_sha256": "086fbaf9a4838ecde56cd3541c9c3212b22d0c5ee745cf2c1c5e0bb31e0ec22c",
+      "kd_stage": "24to8",
+      "mean_total_loss": 0.17622653172702485,
+      "mean_interface_loss": 0.21562811763578793,
+      "mean_final_loss": 0.16626427527853593
+    },
+    {
+      "name": "6L task-free KD",
+      "layers": 6,
+      "parameters": 377207424,
+      "weight_bytes": 754423424,
+      "weight_mib": 719.4742431640625,
+      "weight_sha256": "492749c20f77a42f6bc993f0927a82217b94f872b7eacf25612f7520c9bc28ce",
+      "kd_stage": "8to6",
+      "mean_total_loss": 0.03382104352249371,
+      "mean_interface_loss": 0.04486069045515251,
+      "mean_final_loss": 0.031147060785087888
+    },
+    {
+      "name": "4L task-free KD root",
+      "layers": 4,
+      "parameters": 334096704,
+      "weight_bytes": 668198976,
+      "weight_mib": 637.2442016601562,
+      "weight_sha256": "2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3",
+      "kd_stage": "6to4",
+      "mean_total_loss": 0.047020394468859195,
+      "mean_interface_loss": 0.07048421185527332,
+      "mean_final_loss": 0.043119939065377366
+    }
+  ]
 }

benchmark/summary.json

@@ -0,0 +1,891 @@
+{
+  "schema": "standalone4l-complete-transfer-benchmark-v2",
+  "status": "complete_from_frozen_reports",
+  "claim_scope": "Exploratory fresh-head transfer evidence on one English task and a previously opened 55-article SemEval-derived test split.",
+  "headless_root_requires_adaptation": true,
+  "test_previously_opened": true,
+  "test_articles": 55,
+  "test_windows": 434,
+  "seeds": [
+    41,
+    42,
+    43
+  ],
+  "three_seed_sd_is_not_confidence_interval": true,
+  "multilingual_downstream_evaluation_performed": false,
+  "upstream_multilingual_metadata": {
+    "source": "Qwen/Qwen3.5-0.8B official model card",
+    "languages_and_dialects_claimed_upstream": 201,
+    "scope": "inherited tokenizer and architecture coverage only"
+  },
+  "protocol": {
+    "epochs": 5,
+    "primary_epoch": 5,
+    "early_stopping": false,
+    "max_length": 256,
+    "stride": 128,
+    "window_target": "span-aligned technique labels",
+    "article_aggregation": "labelwise maximum probability",
+    "micro_batch_size": 1,
+    "gradient_accumulation": 32,
+    "effective_batch_size": 32,
+    "learning_rate": 2e-05,
+    "weight_decay": 0.01,
+    "gradient_clip": 1.0,
+    "optimizer": "AdamW",
+    "dtype": "bfloat16",
+    "thresholds": [
+      0.05,
+      0.1,
+      0.15,
+      0.2,
+      0.25,
+      0.3,
+      0.35,
+      0.4,
+      0.45,
+      0.5,
+      0.55,
+      0.6,
+      0.65,
+      0.7,
+      0.75,
+      0.8,
+      0.85,
+      0.9
+    ],
+    "threshold_selection": "epoch-5 calibration macro-F1, then micro-F1"
+  },
+  "arms": {
+    "task_agnostic_base": {
+      "display": "Task-agnostic KD base",
+      "lineage": "root base plus a fresh 14-label transfer head",
+      "metrics": {
+        "macro_f1": {
+          "values": [
+            0.5883093293517536,
+            0.6028379036218096,
+            0.5712816197954101
+          ],
+          "mean": 0.5874762842563245,
+          "sample_sd": 0.01579462678670608,
+          "n": 3
+        },
+        "micro_f1": {
+          "values": [
+            0.6468085106382978,
+            0.6655172413793103,
+            0.6465256797583082
+          ],
+          "mean": 0.6529504772586388,
+          "sample_sd": 0.010884055708931178,
+          "n": 3
+        },
+        "exact_match": {
+          "values": [
+            0.05454545454545454,
+            0.03636363636363636,
+            0.0
+          ],
+          "mean": 0.0303030303030303,
+          "sample_sd": 0.027773186030035393,
+          "n": 3
+        },
+        "predicted_positive_rate": {
+          "values": [
+            0.5714285714285714,
+            0.4090909090909091,
+            0.5155844155844156
+          ],
+          "mean": 0.4987012987012987,
+          "sample_sd": 0.08247520160679181,
+          "n": 3
+        }
+      },
+      "resources": {
+        "parameters": {
+          "values": [
+            334111040.0,
+            334111040.0,
+            334111040.0
+          ],
+          "mean": 334111040.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_allocated_bytes": {
+          "values": [
+            3380975616.0,
+            3380975616.0,
+            3380975616.0
+          ],
+          "mean": 3380975616.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_reserved_bytes": {
+          "values": [
+            3519021056.0,
+            3519021056.0,
+            3510632448.0
+          ],
+          "mean": 3516224853.3333335,
+          "sample_sd": 4843165.086926248,
+          "n": 3
+        },
+        "total_wall_seconds": {
+          "values": [
+            1583.0297977919981,
+            1588.617994428001,
+            1603.4310500120046
+          ],
+          "mean": 1591.692947410668,
+          "sample_sd": 10.54249851412311,
+          "n": 3
+        },
+        "test_seconds_per_article": {
+          "values": [
+            0.08164824154545468,
+            0.07871871667289683,
+            0.08355205721825107
+          ],
+          "mean": 0.08130633847886752,
+          "sample_sd": 0.0024347419754560395,
+          "n": 3
+        }
+      },
+      "thresholds": [
+        0.15,
+        0.3,
+        0.1
+      ]
+    },
+    "structural_copy_control": {
+      "display": "Structural copy, no task-free KD",
+      "lineage": "same four-layer structure without task-free KD",
+      "metrics": {
+        "macro_f1": {
+          "values": [
+            0.5169424579262757,
+            0.5349405779540175,
+            0.5023778364496152
+          ],
+          "mean": 0.5180869574433028,
+          "sample_sd": 0.016311512588628545,
+          "n": 3
+        },
+        "micro_f1": {
+          "values": [
+            0.5773955773955775,
+            0.6083445491251682,
+            0.585635359116022
+          ],
+          "mean": 0.5904584952122559,
+          "sample_sd": 0.01602830914762526,
+          "n": 3
+        },
+        "exact_match": {
+          "values": [
+            0.0,
+            0.0,
+            0.0
+          ],
+          "mean": 0.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "predicted_positive_rate": {
+          "values": [
+            0.712987012987013,
+            0.6207792207792208,
+            0.5961038961038961
+          ],
+          "mean": 0.6432900432900432,
+          "sample_sd": 0.06160737456395461,
+          "n": 3
+        }
+      },
+      "resources": {
+        "parameters": {
+          "values": [
+            334111040.0,
+            334111040.0,
+            334111040.0
+          ],
+          "mean": 334111040.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_allocated_bytes": {
+          "values": [
+            3380975616.0,
+            3380975616.0,
+            3380975616.0
+          ],
+          "mean": 3380975616.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_reserved_bytes": {
+          "values": [
+            3519021056.0,
+            3519021056.0,
+            3510632448.0
+          ],
+          "mean": 3516224853.3333335,
+          "sample_sd": 4843165.086926248,
+          "n": 3
+        },
+        "total_wall_seconds": {
+          "values": [
+            1575.5905925639963,
+            1623.5639530190092,
+            1629.195114912989
+          ],
+          "mean": 1609.4498868319981,
+          "sample_sd": 29.457874198835384,
+          "n": 3
+        },
+        "test_seconds_per_article": {
+          "values": [
+            0.07877608058201573,
+            0.08307423260041767,
+            0.08163602005436339
+          ],
+          "mean": 0.08116211107893226,
+          "sample_sd": 0.002187914527825257,
+          "n": 3
+        }
+      },
+      "thresholds": [
+        0.05,
+        0.2,
+        0.2
+      ]
+    },
+    "existing_specialized_separate_lineage": {
+      "display": "Existing specialized 4L, separate lineage",
+      "lineage": "historical task-specific shrink/distillation lineage",
+      "metrics": {
+        "macro_f1": {
+          "values": [
+            0.5832848452288824,
+            0.5958405181160762,
+            0.5835785045218137
+          ],
+          "mean": 0.5875679559555907,
+          "sample_sd": 0.007165753447433348,
+          "n": 3
+        },
+        "micro_f1": {
+          "values": [
+            0.6494252873563218,
+            0.6458616010854816,
+            0.6486486486486486
+          ],
+          "mean": 0.647978512363484,
+          "sample_sd": 0.0018739735696546695,
+          "n": 3
+        },
+        "exact_match": {
+          "values": [
+            0.0,
+            0.0,
+            0.01818181818181818
+          ],
+          "mean": 0.006060606060606061,
+          "sample_sd": 0.01049727762162956,
+          "n": 3
+        },
+        "predicted_positive_rate": {
+          "values": [
+            0.5597402597402598,
+            0.612987012987013,
+            0.5688311688311688
+          ],
+          "mean": 0.5805194805194805,
+          "sample_sd": 0.02848274311618351,
+          "n": 3
+        }
+      },
+      "resources": {
+        "parameters": {
+          "values": [
+            210903360.0,
+            210903360.0,
+            210903360.0
+          ],
+          "mean": 210903360.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_allocated_bytes": {
+          "values": [
+            2143655936.0,
+            2143655936.0,
+            2143655936.0
+          ],
+          "mean": 2143655936.0,
+          "sample_sd": 0.0,
+          "n": 3
+        },
+        "peak_cuda_reserved_bytes": {
+          "values": [
+            2281701376.0,
+            2281701376.0,
+            2273312768.0
+          ],
+          "mean": 2278905173.3333335,
+          "sample_sd": 4843165.086926248,
+          "n": 3
+        },
+        "total_wall_seconds": {
+          "values": [
+            1634.488557530014,
+            1652.9397718419787,
+            1643.488202479988
+          ],
+          "mean": 1643.6388439506602,
+          "sample_sd": 9.226529522850198,
+          "n": 3
+        },
+        "test_seconds_per_article": {
+          "values": [
+            0.0847959352360869,
+            0.08598086021798239,
+            0.08552987641781908
+          ],
+          "mean": 0.08543555729062946,
+          "sample_sd": 0.0005980667826383295,
+          "n": 3
+        }
+      },
+      "thresholds": [
+        0.1,
+        0.05,
+        0.05
+      ]
+    }
+  },
+  "same_seed_macro_f1_deltas_base_minus_control": [
+    0.07136687142547793,
+    0.06789732566779205,
+    0.06890378334579494
+  ],
+  "mean_macro_f1_delta_base_minus_control": 0.06938932681302164,
+  "publication_gate_passed": true,
+  "per_label_summary": [
+    {
+      "arm": "task_agnostic_base",
+      "label": "Appeal_to_Authority",
+      "support": 11,
+      "precision_mean": 0.3958614484930274,
+      "precision_sample_sd": 0.1238064457010119,
+      "recall_mean": 0.5757575757575757,
+      "recall_sample_sd": 0.05248638810814781,
+      "f1_mean": 0.4657088122605364,
+      "f1_sample_sd": 0.1020990126979069
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Appeal_to_fear-prejudice",
+      "support": 24,
+      "precision_mean": 0.5540244550770866,
+      "precision_sample_sd": 0.02525673990402998,
+      "recall_mean": 0.7361111111111112,
+      "recall_sample_sd": 0.13393959390267993,
+      "f1_mean": 0.6277521761392729,
+      "f1_sample_sd": 0.04994879016183189
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Bandwagon,Reductio_ad_hitlerum",
+      "support": 7,
+      "precision_mean": 0.4851851851851852,
+      "precision_sample_sd": 0.07882887649552875,
+      "recall_mean": 0.8095238095238095,
+      "recall_sample_sd": 0.08247860988423221,
+      "f1_mean": 0.6006778309409889,
+      "f1_sample_sd": 0.04793776258918466
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Black-and-White_Fallacy",
+      "support": 12,
+      "precision_mean": 0.3667929292929293,
+      "precision_sample_sd": 0.07670649392711647,
+      "recall_mean": 0.47222222222222227,
+      "recall_sample_sd": 0.09622504486493764,
+      "f1_mean": 0.4053892966936445,
+      "f1_sample_sd": 0.04211369218048746
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Causal_Oversimplification",
+      "support": 18,
+      "precision_mean": 0.5428824049513704,
+      "precision_sample_sd": 0.039216350192122156,
+      "recall_mean": 0.7962962962962963,
+      "recall_sample_sd": 0.11564811108145183,
+      "f1_mean": 0.6446992723588468,
+      "f1_sample_sd": 0.0634561308339175
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Doubt",
+      "support": 22,
+      "precision_mean": 0.5223832528180354,
+      "precision_sample_sd": 0.06742780050556177,
+      "recall_mean": 0.9848484848484849,
+      "recall_sample_sd": 0.026243194054073875,
+      "f1_mean": 0.6802057822343391,
+      "f1_sample_sd": 0.04928561754686331
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Exaggeration,Minimisation",
+      "support": 22,
+      "precision_mean": 0.4620811287477954,
+      "precision_sample_sd": 0.016801198309751,
+      "recall_mean": 0.7727272727272727,
+      "recall_sample_sd": 0.15745916432444337,
+      "f1_mean": 0.5727040816326531,
+      "f1_sample_sd": 0.03645259990419198
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Flag-Waving",
+      "support": 17,
+      "precision_mean": 0.5642642642642642,
+      "precision_sample_sd": 0.1478145677656334,
+      "recall_mean": 0.8235294117647058,
+      "recall_sample_sd": 0.1764705882352941,
+      "f1_mean": 0.6464506172839506,
+      "f1_sample_sd": 0.035742219952758225
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Loaded_Language",
+      "support": 45,
+      "precision_mean": 0.8649032907793016,
+      "precision_sample_sd": 0.046984312428427724,
+      "recall_mean": 0.9851851851851852,
+      "recall_sample_sd": 0.012830005981991702,
+      "f1_mean": 0.9208741623431562,
+      "f1_sample_sd": 0.03200926878737141
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Name_Calling,Labeling",
+      "support": 33,
+      "precision_mean": 0.7380952380952381,
+      "precision_sample_sd": 0.0047619047619048005,
+      "recall_mean": 0.9090909090909091,
+      "recall_sample_sd": 0.10925912955951485,
+      "f1_mean": 0.8125087983911513,
+      "f1_sample_sd": 0.04252971464057227
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Repetition",
+      "support": 23,
+      "precision_mean": 0.5196620583717357,
+      "precision_sample_sd": 0.021645909050194662,
+      "recall_mean": 0.7536231884057971,
+      "recall_sample_sd": 0.06641414050660639,
+      "f1_mean": 0.6147520781519571,
+      "f1_sample_sd": 0.03505927109416221
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Slogans",
+      "support": 14,
+      "precision_mean": 0.5161361626878869,
+      "precision_sample_sd": 0.20305677930281646,
+      "recall_mean": 0.7380952380952381,
+      "recall_sample_sd": 0.10910894511799614,
+      "f1_mean": 0.5834824090638044,
+      "f1_sample_sd": 0.09862686289220693
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Thought-terminating_Cliches",
+      "support": 7,
+      "precision_mean": 0.3492063492063492,
+      "precision_sample_sd": 0.143516994603374,
+      "recall_mean": 0.3333333333333333,
+      "recall_sample_sd": 0.08247860988423225,
+      "f1_mean": 0.319014319014319,
+      "f1_sample_sd": 0.04017590711117064
+    },
+    {
+      "arm": "task_agnostic_base",
+      "label": "Whataboutism,Straw_Men,Red_Herring",
+      "support": 10,
+      "precision_mean": 0.2761904761904762,
+      "precision_sample_sd": 0.05302632726504782,
+      "recall_mean": 0.5,
+      "recall_sample_sd": 0.2645751311064591,
+      "f1_mean": 0.330448343079922,
+      "f1_sample_sd": 0.021845164720973014
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Appeal_to_Authority",
+      "support": 11,
+      "precision_mean": 0.26003086419753085,
+      "precision_sample_sd": 0.06281357095350094,
+      "recall_mean": 0.5757575757575758,
+      "recall_sample_sd": 0.262431940540739,
+      "f1_mean": 0.35477582846003897,
+      "f1_sample_sd": 0.11479479036518875
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Appeal_to_fear-prejudice",
+      "support": 24,
+      "precision_mean": 0.5104813315339631,
+      "precision_sample_sd": 0.038427508996734555,
+      "recall_mean": 0.8611111111111112,
+      "recall_sample_sd": 0.024056261216234387,
+      "f1_mean": 0.6401745363877202,
+      "f1_sample_sd": 0.029305467347150463
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Bandwagon,Reductio_ad_hitlerum",
+      "support": 7,
+      "precision_mean": 0.2876623376623376,
+      "precision_sample_sd": 0.061386823077618426,
+      "recall_mean": 0.8571428571428571,
+      "recall_sample_sd": 0.14285714285714285,
+      "f1_mean": 0.4306391777656145,
+      "f1_sample_sd": 0.08686392624022733
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Black-and-White_Fallacy",
+      "support": 12,
+      "precision_mean": 0.31339031339031337,
+      "precision_sample_sd": 0.08301311842925053,
+      "recall_mean": 0.5277777777777778,
+      "recall_sample_sd": 0.2678791878053599,
+      "f1_mean": 0.38778427550357375,
+      "f1_sample_sd": 0.13069755090515112
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Causal_Oversimplification",
+      "support": 18,
+      "precision_mean": 0.4464201043148412,
+      "precision_sample_sd": 0.013538446216471536,
+      "recall_mean": 0.9259259259259259,
+      "recall_sample_sd": 0.03207501495497923,
+      "f1_mean": 0.6023809523809524,
+      "f1_sample_sd": 0.018643640071557464
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Doubt",
+      "support": 22,
+      "precision_mean": 0.4455314009661836,
+      "precision_sample_sd": 0.04067072694857813,
+      "recall_mean": 1.0,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.615686274509804,
+      "f1_sample_sd": 0.039427200694482675
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Exaggeration,Minimisation",
+      "support": 22,
+      "precision_mean": 0.4364176150420584,
+      "precision_sample_sd": 0.025093958106097894,
+      "recall_mean": 0.8787878787878788,
+      "recall_sample_sd": 0.05248638810814775,
+      "f1_mean": 0.5829803047194351,
+      "f1_sample_sd": 0.030899288356345502
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Flag-Waving",
+      "support": 17,
+      "precision_mean": 0.3814814814814815,
+      "precision_sample_sd": 0.05010277503136551,
+      "recall_mean": 0.803921568627451,
+      "recall_sample_sd": 0.17970885078258195,
+      "f1_mean": 0.5138595287119652,
+      "f1_sample_sd": 0.06400931989326974
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Loaded_Language",
+      "support": 45,
+      "precision_mean": 0.8492581841638446,
+      "precision_sample_sd": 0.016026591848331686,
+      "recall_mean": 1.0,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.9184311025253588,
+      "f1_sample_sd": 0.009372233868132662
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Name_Calling,Labeling",
+      "support": 33,
+      "precision_mean": 0.6375,
+      "precision_sample_sd": 0.03410563654946854,
+      "recall_mean": 0.9696969696969697,
+      "recall_sample_sd": 0.030303030303030276,
+      "f1_mean": 0.7685185185185185,
+      "f1_sample_sd": 0.02023900779105559
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Repetition",
+      "support": 23,
+      "precision_mean": 0.44863933452168747,
+      "precision_sample_sd": 0.01520014564765896,
+      "recall_mean": 0.9420289855072463,
+      "recall_sample_sd": 0.06641414050660639,
+      "f1_mean": 0.607248835423033,
+      "f1_sample_sd": 0.019858627551810427
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Slogans",
+      "support": 14,
+      "precision_mean": 0.3274318274318274,
+      "precision_sample_sd": 0.05434598572773684,
+      "recall_mean": 0.4523809523809524,
+      "recall_sample_sd": 0.1486904285332952,
+      "f1_mean": 0.3715828924162258,
+      "f1_sample_sd": 0.07106660236343577
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Thought-terminating_Cliches",
+      "support": 7,
+      "precision_mean": 0.1142156862745098,
+      "precision_sample_sd": 0.012848374923608973,
+      "recall_mean": 0.38095238095238093,
+      "recall_sample_sd": 0.1649572197684645,
+      "f1_mean": 0.17331433998100665,
+      "f1_sample_sd": 0.029065880233243847
+    },
+    {
+      "arm": "structural_copy_control",
+      "label": "Whataboutism,Straw_Men,Red_Herring",
+      "support": 10,
+      "precision_mean": 0.19344598054275472,
+      "precision_sample_sd": 0.03373872471827226,
+      "recall_mean": 0.5666666666666667,
+      "recall_sample_sd": 0.2081665999466133,
+      "f1_mean": 0.28584083690299267,
+      "f1_sample_sd": 0.06078854830133676
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Appeal_to_Authority",
+      "support": 11,
+      "precision_mean": 0.37836021505376344,
+      "precision_sample_sd": 0.10546365636105155,
+      "recall_mean": 0.6666666666666666,
+      "recall_sample_sd": 0.2287828616748712,
+      "f1_mean": 0.46885361552028215,
+      "f1_sample_sd": 0.09502747629890008
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Appeal_to_fear-prejudice",
+      "support": 24,
+      "precision_mean": 0.525708635464733,
+      "precision_sample_sd": 0.02235198659615819,
+      "recall_mean": 0.8888888888888888,
+      "recall_sample_sd": 0.048112522432468774,
+      "f1_mean": 0.6599065351235072,
+      "f1_sample_sd": 0.015362381542320084
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Bandwagon,Reductio_ad_hitlerum",
+      "support": 7,
+      "precision_mean": 0.44485144485144484,
+      "precision_sample_sd": 0.016487284575115283,
+      "recall_mean": 0.7619047619047619,
+      "recall_sample_sd": 0.1649572197684645,
+      "f1_mean": 0.5571428571428572,
+      "f1_sample_sd": 0.051507875363771265
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Black-and-White_Fallacy",
+      "support": 12,
+      "precision_mean": 0.4662644544997486,
+      "precision_sample_sd": 0.004538274146363688,
+      "recall_mean": 0.5833333333333334,
+      "recall_sample_sd": 0.08333333333333331,
+      "f1_mean": 0.516747552149851,
+      "f1_sample_sd": 0.03589484965697345
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Causal_Oversimplification",
+      "support": 18,
+      "precision_mean": 0.4472934472934473,
+      "precision_sample_sd": 0.013055771210700413,
+      "recall_mean": 0.8333333333333333,
+      "recall_sample_sd": 0.1469861839480328,
+      "f1_mean": 0.5781794553724379,
+      "f1_sample_sd": 0.028407563015813122
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Doubt",
+      "support": 22,
+      "precision_mean": 0.5045661001171619,
+      "precision_sample_sd": 0.023581174969843577,
+      "recall_mean": 1.0,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.6704939668174963,
+      "f1_sample_sd": 0.020973132363659372
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Exaggeration,Minimisation",
+      "support": 22,
+      "precision_mean": 0.44589793281653745,
+      "precision_sample_sd": 0.010987847993006481,
+      "recall_mean": 0.9393939393939394,
+      "recall_sample_sd": 0.06943296507508849,
+      "f1_mean": 0.6043956043956044,
+      "f1_sample_sd": 0.022305256187015846
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Flag-Waving",
+      "support": 17,
+      "precision_mean": 0.5576388888888889,
+      "precision_sample_sd": 0.03705367103320583,
+      "recall_mean": 0.8823529411764706,
+      "recall_sample_sd": 0.11764705882352944,
+      "f1_mean": 0.680769868923179,
+      "f1_sample_sd": 0.04164658769406446
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Loaded_Language",
+      "support": 45,
+      "precision_mean": 0.8387650085763294,
+      "precision_sample_sd": 0.01782556577257849,
+      "recall_mean": 1.0,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.9122448979591837,
+      "f1_sample_sd": 0.01060439269940129
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Name_Calling,Labeling",
+      "support": 33,
+      "precision_mean": 0.6957104190221673,
+      "precision_sample_sd": 0.019201490809269257,
+      "recall_mean": 0.9696969696969697,
+      "recall_sample_sd": 0.030303030303030276,
+      "f1_mean": 0.8101049442821594,
+      "f1_sample_sd": 0.022092733328802537
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Repetition",
+      "support": 23,
+      "precision_mean": 0.4781557067271353,
+      "precision_sample_sd": 0.009897987322673213,
+      "recall_mean": 0.9420289855072463,
+      "recall_sample_sd": 0.06641414050660639,
+      "f1_mean": 0.6337774426009719,
+      "f1_sample_sd": 0.016444118333485254
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Slogans",
+      "support": 14,
+      "precision_mean": 0.4334215167548501,
+      "precision_sample_sd": 0.05811581861282867,
+      "recall_mean": 0.6904761904761905,
+      "recall_sample_sd": 0.1649572197684645,
+      "f1_mean": 0.5201316298877274,
+      "f1_sample_sd": 0.018567913895656114
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Thought-terminating_Cliches",
+      "support": 7,
+      "precision_mean": 0.18446115288220552,
+      "precision_sample_sd": 0.03641251252864963,
+      "recall_mean": 0.6666666666666666,
+      "recall_sample_sd": 0.1649572197684645,
+      "f1_mean": 0.2868626868626869,
+      "f1_sample_sd": 0.0511620864429468
+    },
+    {
+      "arm": "existing_specialized_separate_lineage",
+      "label": "Whataboutism,Straw_Men,Red_Herring",
+      "support": 10,
+      "precision_mean": 0.2777777777777778,
+      "precision_sample_sd": 0.0481125224324688,
+      "recall_mean": 0.4,
+      "recall_sample_sd": 0.0,
+      "f1_mean": 0.32634032634032634,
+      "f1_sample_sd": 0.03229931575886014
+    }
+  ],
+  "compression_ladder": [
+    {
+      "name": "24L upstream text teacher",
+      "layers": 24,
+      "parameters": 752393024,
+      "weight_bytes": 1504825632,
+      "weight_mib": 1435.1135559082031,
+      "weight_sha256": "30783da4651259c26b8a204821ea426b9526431e0c14da0d1c6fa284e7414e5e",
+      "kd_stage": "",
+      "mean_total_loss": "",
+      "mean_interface_loss": "",
+      "mean_final_loss": ""
+    },
+    {
+      "name": "8L task-free KD",
+      "layers": 8,
+      "parameters": 420318144,
+      "weight_bytes": 840647880,
+      "weight_mib": 801.7042922973633,
+      "weight_sha256": "086fbaf9a4838ecde56cd3541c9c3212b22d0c5ee745cf2c1c5e0bb31e0ec22c",
+      "kd_stage": "24to8",
+      "mean_total_loss": 0.17622653172702485,
+      "mean_interface_loss": 0.21562811763578793,
+      "mean_final_loss": 0.16626427527853593
+    },
+    {
+      "name": "6L task-free KD",
+      "layers": 6,
+      "parameters": 377207424,
+      "weight_bytes": 754423424,
+      "weight_mib": 719.4742431640625,
+      "weight_sha256": "492749c20f77a42f6bc993f0927a82217b94f872b7eacf25612f7520c9bc28ce",
+      "kd_stage": "8to6",
+      "mean_total_loss": 0.03382104352249371,
+      "mean_interface_loss": 0.04486069045515251,
+      "mean_final_loss": 0.031147060785087888
+    },
+    {
+      "name": "4L task-free KD root",
+      "layers": 4,
+      "parameters": 334096704,
+      "weight_bytes": 668198976,
+      "weight_mib": 637.2442016601562,
+      "weight_sha256": "2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3",
+      "kd_stage": "6to4",
+      "mean_total_loss": 0.047020394468859195,
+      "mean_interface_loss": 0.07048421185527332,
+      "mean_final_loss": 0.043119939065377366
+    }
+  ]
+}

docs/RELEASE_CONTRACT.md

@@ -14,14 +14,13 @@
    SemEval examples, labels, logits, thresholds, or evidence objectives.
 2. `models/semeval-propaganda/`: the existing three-seed task-specific
    classifier copied byte-for-byte as a separate historical lineage. It is not
-   described as a fine-tuned child of the new root base.
+   described as a fine-tuned child of the root base.
 3. `distillation/`: the reusable 24L→8L→6L→4L platform for unlabeled
-   representation distillation and user-owned single-label or multilabel
-   classification data.
+   representation distillation and user-owned single-label or multilabel data.
 
 ## Base training contract
 
-- Source: the text backbone and full tokenizer of `Qwen/Qwen3.5-0.8B`.
+- Source: text backbone and full tokenizer of `Qwen/Qwen3.5-0.8B`.
 - Structural defaults:
   - 24→8: `0,4,6,11,13,16,20,23`
   - 8→6: `0,1,3,4,6,7`
@@ -31,43 +30,81 @@
 - Objective: aligned hidden-boundary and final-representation distillation.
 - One epoch per stage, seed 41, maximum length 128, batch size 1, gradient
   accumulation 8, learning rate 2e-5, BF16 CUDA.
+- Root artifact: 334,096,704 parameters and 668,198,976 BF16 weight bytes.
 - No classification head is published at the root.
 
-The WikiText dataset page currently has inconsistent license wording: metadata
-lists CC BY-SA 3.0 and GFDL, while the prose says CC BY-SA 4.0. The release
-redistributes no WikiText text and discloses both statements rather than
-silently selecting one.
+WikiText source text is not redistributed. Its dataset page's metadata and
+prose currently disagree on the license version, so users must review the
+upstream terms directly.
 
-## Transfer gate
+## Benchmark evidence contract
 
-The only downstream validation task is an article-level multilabel task derived
-from SemEval-2020 Task 11. It is not the official span- or fragment-level task.
-The public test split contains 55 previously opened articles, so results are
-exploratory rather than confirmatory.
+The headless root is evaluated through a fresh-head transfer probe, not by
+pretending it is already a classifier. Both the task-agnostic base and the
+structural-copy control use the same frozen protocol:
 
-| Initialization | Seed 41 | Seed 42 | Seed 43 | Mean ± sample SD |
-|---|---:|---:|---:|---:|
-| Task-agnostic KD base | 0.58831 | 0.60284 | 0.57128 | 0.58748 ± 0.01579 |
-| Structural 4L copy, no task-free KD | 0.51694 | 0.53494 | 0.50238 | 0.51809 ± 0.01631 |
+- SemEval-2020 Task 11 annotations converted to article-level 14-label targets;
+- train/calibration/test article counts 260/56/55;
+- seeds 41/42/43; five epochs; no early stopping;
+- 256-token windows, stride 128, label-wise maximum article aggregation;
+- micro batch 1, accumulation 32, AdamW, LR 2e-5, BF16;
+- one global threshold selected on epoch-5 calibration Macro-F1 then Micro-F1;
+- 55 previously opened test articles and 434 test windows.
 
-The preregistered publication gate was that the task-agnostic base must not be
-worse than the structural-copy control by more than 0.01 mean Macro-F1. The
-observed mean difference was +0.06939, so the gate passed. Three training seeds
-are not a confidence interval and no significance claim is made.
+| Initialization | Macro-F1 mean ± SD | Micro-F1 mean ± SD | Exact match mean ± SD |
+|---|---:|---:|---:|
+| Task-agnostic KD base | 0.58748 ± 0.01579 | 0.65295 ± 0.01088 | 0.03030 ± 0.02777 |
+| Structural copy, no task-free KD | 0.51809 ± 0.01631 | 0.59046 ± 0.01603 | 0.00000 ± 0.00000 |
+
+The observed same-seed mean Macro-F1 difference is +0.06939. All values and
+sample SDs must be derived from the six frozen full reports at build time; the
+card, JSON, and CSV tables must numerically close against those reports.
+
+The existing specialized 4L lineage may appear only as a clearly labeled
+reference because it uses a separate task-specific shrink/distillation history
+and reduced vocabulary. It is not evidence that the task-free base produced
+that checkpoint.
+
+The report bundle must include:
+
+- per-arm, per-seed Macro-F1, Micro-F1, exact match, positive rate, threshold;
+- three-seed means and sample SDs;
+- per-label precision, recall, F1, and fixed test support;
+- parameters, CUDA allocation/reservation, total wall time, and test timing;
+- compression-ladder depth, parameter count, BF16 size, hashes, and KD losses;
+- normalized full frozen reports with article-level probability vectors.
+
+The test is already opened, per-label supports are small, and three seeds are
+not a confidence interval. No confirmatory, significance, universal
+classification, or population-generalization claim is permitted.
+
+## Multilingual metadata contract
+
+The root keeps the full upstream tokenizer. The official Qwen3.5 card states
+support for 201 languages and dialects and reports upstream multilingual
+benchmarks. Therefore `language: multilingual` and a `multilingual` tag are
+permitted as inherited input/architecture metadata.
+
+This release's task-free KD corpus and only downstream probe are English.
+Accordingly, the card must state prominently that the four-layer root has no
+direct multilingual downstream evaluation and that tokenizer coverage is not a
+multilingual classification-quality result. It may not claim verified quality
+across 201 languages.
 
 ## Release gates
 
 - Root `model.safetensors` contains no `score.weight`.
-- Root loads with `AutoModel`, and a temporary classification head can be
-  attached with `AutoModelForSequenceClassification`.
+- Root loads with `AutoModel`; a temporary head can be attached with
+  `AutoModelForSequenceClassification`.
 - Root weight SHA-256 is
   `2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3`.
 - Base manifests say `labels_read=false` and `semeval_used=false`.
-- Nested specialized checkpoint hashes equal the already published release.
+- Root metadata contains `multilingual` and contains no propaganda task tag.
+- Benchmark CSV/JSON/card values close against all frozen reports.
+- Nested specialized checkpoint hashes equal the existing release.
 - The platform validates label order and ID/group split leakage, never
   overwrites output directories, hashes inputs/outputs, verifies immutable
   teachers, and reloads fresh checkpoints.
 - The release contains no SemEval or WikiText source data.
 - The old specialized repository revision and inventory are checked before and
-  after creation of the new repository.
-
+  after updating only the new repository.

models/semeval-propaganda/README.md

@@ -23,5 +23,7 @@ model = AutoModelForSequenceClassification.from_pretrained(
 ```
 
 The files below are byte-identical copies of the current release bundle. The
-original release manifest and locked thresholds are included for provenance.
-No SemEval source article or annotation file is redistributed.
+original release manifest, locked thresholds, three full reports, and aggregate
+are included for provenance. The repository-level `benchmark/` directory also
+places this lineage beside the task-agnostic transfer probe as a clearly marked
+reference. No SemEval source article or annotation file is redistributed.

release_manifest.json

@@ -1,5 +1,5 @@
 {
-  "schema": "qwen35-unified-classification-base-release-v1",
+  "schema": "qwen35-unified-classification-base-release-v2",
   "status": "complete",
   "repo_id": "mp-juuuns/qwen35-standalone4l-classification-base",
   "old_repo_preserved": {
@@ -8,6 +8,8 @@
   },
   "root_role": "task_agnostic_headless_classification_backbone",
   "root_weight_sha256": "2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3",
+  "root_multilingual_metadata_scope": "inherited tokenizer/architecture coverage; no direct multilingual downstream evaluation",
+  "benchmark_schema": "standalone4l-complete-transfer-benchmark-v2",
   "nested_specialized_is_separate_lineage": true,
   "source_data_redistributed": false,
   "files_before_manifest": [
@@ -28,57 +30,134 @@
     {
       "path": "README.md",
       "role": "root model card",
-      "bytes": 5628,
-      "sha256": "4f5af67b76f19fca88e242298fa501aa6aa811b4f1f51b481d17abb46e0d39cf",
+      "bytes": 12909,
+      "sha256": "5fcf007de6b6e828711c874310de1e042f1dada6989c70930ac0881285a9dc61",
       "source": "docs/huggingface/UNIFIED_BASE_RELEASE_MODEL_CARD.md"
     },
     {
-      "path": "benchmark/reports/structural_control-seed41.json",
-      "role": "normalized transfer report",
+      "path": "benchmark/BENCHMARK_CARD.md",
+      "role": "generated benchmark report",
+      "bytes": 3173,
+      "sha256": "bf4073de38c3905effd2dcce6f38894f325c8a0a79bedfb2779f0545c86b9c9f",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/arm_summary.csv",
+      "role": "three-seed metric summary",
+      "bytes": 1655,
+      "sha256": "e01bd7ac59a3212e44b77e5fcfbc7c41ec91d74c998306d2438209d6ef1a3aa7",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/compression_ladder.csv",
+      "role": "24L to 4L compression evidence",
+      "bytes": 827,
+      "sha256": "2c544c3907c4f592d724f658850b8c6b9634845cb5b50a66ebd2ea6c5d48ce0b",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/per_label_metrics.csv",
+      "role": "per-label metrics by arm and seed",
+      "bytes": 12963,
+      "sha256": "ffc21be32b2106c7bf1df3854097a975ac370da01ac1b732eb5c2ba9075ddeda",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/per_label_summary.csv",
+      "role": "per-label three-seed summary",
+      "bytes": 7009,
+      "sha256": "c8e3876abf6cfe960c9332d3bc9d1fbde2f3213f3111351acf83b21bbb184049",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/reports/existing_specialized_separate_lineage-seed41.json",
+      "role": "normalized frozen benchmark report",
+      "bytes": 115793,
+      "sha256": "c1bd379680abcf50b47b7d23e4aafbc8b5e7f40a0b0e9700bde21a8ce49af803",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/reports/existing_specialized_separate_lineage-seed42.json",
+      "role": "normalized frozen benchmark report",
+      "bytes": 115780,
+      "sha256": "a30e38433d008db4aa7e5fbfd477ab80a0daf50e970f8e7b04397efc1a47843a",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/reports/existing_specialized_separate_lineage-seed43.json",
+      "role": "normalized frozen benchmark report",
+      "bytes": 115985,
+      "sha256": "96466290d09d27b8bc044a9f9a95765ce12b4d7b02dfb0123e436e7535623353",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/reports/structural_copy_control-seed41.json",
+      "role": "normalized frozen benchmark report",
       "bytes": 115612,
       "sha256": "816c3249d3ec281ef73d34dc3ab7576a48415d250c32977eb94ca37f73b0b810",
       "source": "generated"
     },
     {
-      "path": "benchmark/reports/structural_control-seed42.json",
-      "role": "normalized transfer report",
+      "path": "benchmark/reports/structural_copy_control-seed42.json",
+      "role": "normalized frozen benchmark report",
       "bytes": 115225,
       "sha256": "0b19eabe1e31d0012c754db3b31173f8a6fcdee30973f1a1bebd3a370d11064c",
       "source": "generated"
     },
     {
-      "path": "benchmark/reports/structural_control-seed43.json",
-      "role": "normalized transfer report",
+      "path": "benchmark/reports/structural_copy_control-seed43.json",
+      "role": "normalized frozen benchmark report",
       "bytes": 115744,
       "sha256": "6dde34173462eac50141ff082c18194f3233f8c06853532d9b12756562e69e2e",
       "source": "generated"
     },
     {
-      "path": "benchmark/reports/task_agnostic-seed41.json",
-      "role": "normalized transfer report",
+      "path": "benchmark/reports/task_agnostic_base-seed41.json",
+      "role": "normalized frozen benchmark report",
       "bytes": 115885,
       "sha256": "6b3a2e0372a83e1e71c2eca9596efc8cd1a3e39c28e25e3373f195431b5858a4",
       "source": "generated"
     },
     {
-      "path": "benchmark/reports/task_agnostic-seed42.json",
-      "role": "normalized transfer report",
+      "path": "benchmark/reports/task_agnostic_base-seed42.json",
+      "role": "normalized frozen benchmark report",
       "bytes": 115674,
       "sha256": "42dc788983716be0e9847aad87ff79c7e7297884f2635a8d098639be77326639",
       "source": "generated"
     },
     {
-      "path": "benchmark/reports/task_agnostic-seed43.json",
-      "role": "normalized transfer report",
+      "path": "benchmark/reports/task_agnostic_base-seed43.json",
+      "role": "normalized frozen benchmark report",
       "bytes": 115062,
       "sha256": "7ee52dc94a152a2f767dcff171f3c9cc1072fabd3ebc7b14db53617b7b1d33d5",
       "source": "generated"
     },
+    {
+      "path": "benchmark/resource_metrics.csv",
+      "role": "resource measurements by arm and seed",
+      "bytes": 1649,
+      "sha256": "bf5cefdec3b2f3646114bd18bbf8e5e39721bde768ba968d73f22e96fc9ca8cf",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/seed_metrics.csv",
+      "role": "overall metrics by arm and seed",
+      "bytes": 1175,
+      "sha256": "e0b08a958d25243158a03dff76ca153d2ded76cc3f1771e4046a0068688ad6a8",
+      "source": "generated"
+    },
     {
       "path": "benchmark/semeval-transfer-summary.json",
-      "role": "machine-readable one-task transfer summary",
-      "bytes": 1190,
-      "sha256": "18c8d840dd2290ca25c517d52cbcf1ad6c6ae8d9c32c545236e96bd5ada7dab2",
+      "role": "compatibility alias for complete transfer summary",
+      "bytes": 26913,
+      "sha256": "3d2c9fcf9c6cc7da1a885968b70a42eaa38659d317c9d3290a2b3066f3f9ee6a",
+      "source": "generated"
+    },
+    {
+      "path": "benchmark/summary.json",
+      "role": "complete machine-readable transfer benchmark",
+      "bytes": 26913,
+      "sha256": "3d2c9fcf9c6cc7da1a885968b70a42eaa38659d317c9d3290a2b3066f3f9ee6a",
       "source": "generated"
     },
     {
@@ -203,8 +282,8 @@
     {
       "path": "docs/RELEASE_CONTRACT.md",
       "role": "release contract",
-      "bytes": 3399,
-      "sha256": "5a1d9e939cd0f350770bcc2a116bc427bb9971bc4fbbbda17ce568e9127185ef",
+      "bytes": 5381,
+      "sha256": "1b6e475b9d7d9af52bb8845de020303b1c6e89646d8e4f9be17dbe4ab2910417",
       "source": "docs/huggingface/UNIFIED_BASE_RELEASE_CONTRACT.md"
     },
     {
@@ -231,8 +310,8 @@
     {
       "path": "models/semeval-propaganda/README.md",
       "role": "specialized lineage card",
-      "bytes": 1231,
-      "sha256": "8dbc2e641280c471b78064808ced7fffcb81c639784a5936b57175975a718f6f",
+      "bytes": 1403,
+      "sha256": "bf4ba7cc2514502a28f424507c41c1cfce5089c051ba7500cf46c1d13860722f",
       "source": "generated"
     },
     {