# 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.