--- dataset_info: - config_name: shard_0 features: - name: original dtype: string - name: prompt struct: - name: chat_turns list: string - name: use_multiturn dtype: bool - name: examples list: 'null' - name: metadata struct: - name: PROMPT_TYPE dtype: string - name: response_0 dtype: string - name: response_1 dtype: string - name: final_response dtype: string - name: source_row dtype: int64 - name: encoder_index dtype: int64 - name: encoder_model dtype: string - name: encoder_generation_params dtype: string - name: generator_model dtype: string - name: generation_params dtype: string - name: decoder_index dtype: int64 - name: decoder_model dtype: string - name: source_shard dtype: int64 - name: jaccard_1 dtype: float64 - name: jaccard_2 dtype: float64 - name: levenshtein dtype: float64 - name: softngram dtype: float64 - name: cosdist dtype: float64 - name: bertscore_precision dtype: float64 - name: bertscore_recall dtype: float64 - name: bertscore dtype: float64 - name: moverscore dtype: float64 - name: reranker dtype: float64 - name: original_editlens_bucket_roberta_large dtype: int64 - name: original_editlens_score_roberta_large dtype: float64 - name: final_response_editlens_bucket_roberta_large dtype: int64 - name: final_response_editlens_score_roberta_large dtype: float64 splits: - name: train num_bytes: 235642579 num_examples: 17393 download_size: 139775674 dataset_size: 235642579 - config_name: shard_1 features: - name: original dtype: string - name: prompt struct: - name: chat_turns list: string - name: use_multiturn dtype: bool - name: examples list: 'null' - name: metadata struct: - name: PROMPT_TYPE dtype: string - name: response_0 dtype: string - name: response_1 dtype: string - name: final_response dtype: string - name: source_row dtype: int64 - name: encoder_index dtype: int64 - name: encoder_model dtype: string - name: encoder_generation_params dtype: string - name: generator_model dtype: string - name: generation_params dtype: string - name: decoder_index dtype: int64 - name: decoder_model dtype: string - name: source_shard dtype: int64 - name: jaccard_1 dtype: float64 - name: jaccard_2 dtype: float64 - name: levenshtein dtype: float64 - name: softngram dtype: float64 - name: cosdist dtype: float64 - name: bertscore_precision dtype: float64 - name: bertscore_recall dtype: float64 - name: bertscore dtype: float64 - name: moverscore dtype: float64 - name: reranker dtype: float64 - name: original_editlens_bucket_roberta_large dtype: int64 - name: original_editlens_score_roberta_large dtype: float64 - name: final_response_editlens_bucket_roberta_large dtype: int64 - name: final_response_editlens_score_roberta_large dtype: float64 splits: - name: train num_bytes: 239723758 num_examples: 17573 download_size: 140521510 dataset_size: 239723758 - config_name: shard_2 features: - name: original dtype: string - name: prompt struct: - name: chat_turns list: string - name: use_multiturn dtype: bool - name: examples list: 'null' - name: metadata struct: - name: PROMPT_TYPE dtype: string - name: response_0 dtype: string - name: response_1 dtype: string - name: final_response dtype: string - name: source_row dtype: int64 - name: encoder_index dtype: int64 - name: encoder_model dtype: string - name: encoder_generation_params dtype: string - name: generator_model dtype: string - name: generation_params dtype: string - name: decoder_index dtype: int64 - name: decoder_model dtype: string - name: source_shard dtype: int64 - name: jaccard_1 dtype: float64 - name: jaccard_2 dtype: float64 - name: levenshtein dtype: float64 - name: softngram dtype: float64 - name: cosdist dtype: float64 - name: bertscore_precision dtype: float64 - name: bertscore_recall dtype: float64 - name: bertscore dtype: float64 - name: moverscore dtype: float64 - name: reranker dtype: float64 - name: original_editlens_bucket_roberta_large dtype: int64 - name: original_editlens_score_roberta_large dtype: float64 - name: final_response_editlens_bucket_roberta_large dtype: int64 - name: final_response_editlens_score_roberta_large dtype: float64 splits: - name: train num_bytes: 234538054 num_examples: 17805 download_size: 135964661 dataset_size: 234538054 - config_name: shard_3 features: - name: original dtype: string - name: prompt struct: - name: chat_turns list: string - name: use_multiturn dtype: bool - name: examples list: 'null' - name: metadata struct: - name: PROMPT_TYPE dtype: string - name: response_0 dtype: string - name: response_1 dtype: string - name: final_response dtype: string - name: source_row dtype: int64 - name: encoder_index dtype: int64 - name: encoder_model dtype: string - name: encoder_generation_params dtype: string - name: generator_model dtype: string - name: generation_params dtype: string - name: decoder_index dtype: int64 - name: decoder_model dtype: string - name: source_shard dtype: int64 - name: jaccard_1 dtype: float64 - name: jaccard_2 dtype: float64 - name: levenshtein dtype: float64 - name: softngram dtype: float64 - name: cosdist dtype: float64 - name: bertscore_precision dtype: float64 - name: bertscore_recall dtype: float64 - name: bertscore dtype: float64 - name: moverscore dtype: float64 - name: reranker dtype: float64 - name: original_editlens_bucket_roberta_large dtype: int64 - name: original_editlens_score_roberta_large dtype: float64 - name: final_response_editlens_bucket_roberta_large dtype: int64 - name: final_response_editlens_score_roberta_large dtype: float64 splits: - name: train num_bytes: 226257989 num_examples: 17530 download_size: 135942633 dataset_size: 226257989 - config_name: shard_4 features: - name: original dtype: string - name: prompt struct: - name: chat_turns list: string - name: use_multiturn dtype: bool - name: examples list: 'null' - name: metadata struct: - name: PROMPT_TYPE dtype: string - name: response_0 dtype: string - name: response_1 dtype: string - name: final_response dtype: string - name: source_row dtype: int64 - name: encoder_index dtype: int64 - name: encoder_model dtype: string - name: encoder_generation_params dtype: string - name: generator_model dtype: string - name: generation_params dtype: string - name: decoder_index dtype: int64 - name: decoder_model dtype: string - name: source_shard dtype: int64 - name: jaccard_1 dtype: float64 - name: jaccard_2 dtype: float64 - name: levenshtein dtype: float64 - name: softngram dtype: float64 - name: cosdist dtype: float64 - name: bertscore_precision dtype: float64 - name: bertscore_recall dtype: float64 - name: bertscore dtype: float64 - name: moverscore dtype: float64 - name: reranker dtype: float64 - name: original_editlens_bucket_roberta_large dtype: int64 - name: original_editlens_score_roberta_large dtype: float64 - name: final_response_editlens_bucket_roberta_large dtype: int64 - name: final_response_editlens_score_roberta_large dtype: float64 splits: - name: train num_bytes: 237820682 num_examples: 17591 download_size: 143241072 dataset_size: 237820682 - config_name: shard_5 features: - name: original dtype: string - name: prompt struct: - name: chat_turns list: string - name: use_multiturn dtype: bool - name: examples list: 'null' - name: metadata struct: - name: PROMPT_TYPE dtype: string - name: response_0 dtype: string - name: response_1 dtype: string - name: final_response dtype: string - name: source_row dtype: int64 - name: encoder_index dtype: int64 - name: encoder_model dtype: string - name: encoder_generation_params dtype: string - name: generator_model dtype: string - name: generation_params dtype: string - name: decoder_index dtype: int64 - name: decoder_model dtype: string - name: source_shard dtype: int64 - name: jaccard_1 dtype: float64 - name: jaccard_2 dtype: float64 - name: levenshtein dtype: float64 - name: softngram dtype: float64 - name: cosdist dtype: float64 - name: bertscore_precision dtype: float64 - name: bertscore_recall dtype: float64 - name: bertscore dtype: float64 - name: moverscore dtype: float64 - name: reranker dtype: float64 - name: original_editlens_bucket_roberta_large dtype: int64 - name: original_editlens_score_roberta_large dtype: float64 - name: final_response_editlens_bucket_roberta_large dtype: int64 - name: final_response_editlens_score_roberta_large dtype: float64 splits: - name: train num_bytes: 223237348 num_examples: 17630 download_size: 132460074 dataset_size: 223237348 - config_name: shard_6 features: - name: original dtype: string - name: prompt struct: - name: chat_turns list: string - name: use_multiturn dtype: bool - name: examples list: 'null' - name: metadata struct: - name: PROMPT_TYPE dtype: string - name: response_0 dtype: string - name: response_1 dtype: string - name: final_response dtype: string - name: source_row dtype: int64 - name: encoder_index dtype: int64 - name: encoder_model dtype: string - name: encoder_generation_params dtype: string - name: generator_model dtype: string - name: generation_params dtype: string - name: decoder_index dtype: int64 - name: decoder_model dtype: string - name: source_shard dtype: int64 - name: jaccard_1 dtype: float64 - name: jaccard_2 dtype: float64 - name: levenshtein dtype: float64 - name: softngram dtype: float64 - name: cosdist dtype: float64 - name: bertscore_precision dtype: float64 - name: bertscore_recall dtype: float64 - name: bertscore dtype: float64 - name: moverscore dtype: float64 - name: reranker dtype: float64 - name: original_editlens_bucket_roberta_large dtype: int64 - name: original_editlens_score_roberta_large dtype: float64 - name: final_response_editlens_bucket_roberta_large dtype: int64 - name: final_response_editlens_score_roberta_large dtype: float64 splits: - name: train num_bytes: 240613087 num_examples: 17411 download_size: 143866401 dataset_size: 240613087 configs: - config_name: shard_0 data_files: - split: train path: shard_0/train-* - config_name: shard_1 data_files: - split: train path: shard_1/train-* - config_name: shard_2 data_files: - split: train path: shard_2/train-* - config_name: shard_3 data_files: - split: train path: shard_3/train-* - config_name: shard_4 data_files: - split: train path: shard_4/train-* - config_name: shard_5 data_files: - split: train path: shard_5/train-* - config_name: shard_6 data_files: - split: train path: shard_6/train-* --- # Encoder/decoder trial: encoder-marginal report - Dataset: `G-reen/encoder-decoder-trial-stat` - Rows analysed: 122,933 (every kept (encoder, decoder, source row) triple; source `G-reen/cc-re-2021-filtered` shard 0, 2000 rows of at most 4000 words) - Prompt file: `prompts/indirect_reference_dataset_train.json` (turn 0 encodes the document, turn 1 reconstructs it from the encoding alone) - Encoders: 9 (granite-4.2-30b-nvfp4 [0], Ornith-1.5-35B-A3B-NVFP4 [1], Llama-3.3-70B-Instruct-NVFP4 [2], Qwen3.8-27B-AWQ-INT4 [3], Mistral-Small-4-119B-2603-NVFP4 [4], gemma-4-31B-it-AWQ-4bit [5], Laguna-S-2.1-NVFP4 [6], claude-sonnet-5 [8], gpt-5.6-terra [9]) - Decoders: 7 (granite-4.2-30b-nvfp4 [0], Ornith-1.5-35B-A3B-NVFP4 [1], Llama-3.3-70B-Instruct-NVFP4 [2], Qwen3.8-27B-AWQ-INT4 [3], Mistral-Small-4-119B-2603-NVFP4 [4], gemma-4-31B-it-AWQ-4bit [5], Laguna-S-2.1-NVFP4 [6]) Models (trial index: config): - 0: `config/gen/train/shard_0.toml` - 1: `config/gen/train/shard_1.toml` - 2: `config/gen/train/shard_2.toml` - 3: `config/gen/train/shard_3.toml` - 4: `config/gen/train/shard_4.toml` - 5: `config/gen/train/shard_5.toml` - 6: `config/gen/train/shard_6.toml` - 7: `config/gen/train/shard_7.toml` (decode only) - 8: `config/encdec/claude_sonnet_5.toml` (encode only) - 9: `config/encdec/gpt_5_6_terra.toml` (encode only) Statistics: - every configured statistic was present. ## Summary: per-encoder means, marginalised over decoders Every encoder encoded the same 2000 rows and every decoder decoded all of every encoder's encodings (the same source rows for every encoder), so each encoder's row is an average over the same decoders and source texts. Values are the decoder-balanced means (mean of the per-decoder means); `Rows` counts the kept rows behind each. For reference, the human source texts score 0.0615 (std 0.1016) on the same EditLens model. | Encoder | Editlens Score | Editlens Bucket | Cosdist | Jaccard 1 | Levenshtein | Softngram | Bertscore | Moverscore | Reranker | Rows | |---|---|---|---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5099 | 1.4517 | 0.1582 | 0.6346 | 2105.3807 | 0.5588 | 0.1327 | 0.5235 | -5.1181 | 13,601 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.4636 | 1.2975 | 0.1559 | 0.6168 | 2042.6838 | 0.5054 | 0.1265 | 0.5114 | -5.1840 | 13,615 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.5696 | 1.6533 | 0.2060 | 0.6611 | 2132.4408 | 0.5877 | 0.1448 | 0.5472 | -3.3607 | 13,723 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.5102 | 1.4585 | 0.1580 | 0.6467 | 2065.3474 | 0.5613 | 0.1336 | 0.5289 | -5.2792 | 13,639 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5443 | 1.5684 | 0.1761 | 0.6657 | 2237.7226 | 0.5832 | 0.1415 | 0.5456 | -4.6743 | 13,731 | | gemma-4-31B-it-AWQ-4bit [5] | 0.5160 | 1.4680 | 0.1720 | 0.6694 | 2073.2000 | 0.5851 | 0.1410 | 0.5459 | -4.9291 | 13,509 | | Laguna-S-2.1-NVFP4 [6] | 0.5377 | 1.5442 | 0.1788 | 0.6693 | 2309.0983 | 0.6054 | 0.1424 | 0.5484 | -4.5894 | 13,662 | | ❗ claude-sonnet-5 [8] | 0.4449 | 1.2363 | 0.1315 | 0.6124 | 2086.9193 | 0.5250 | 0.1214 | 0.5002 | -5.8408 | 13,732 | | gpt-5.6-terra [9] | 0.4816 | 1.3664 | 0.1365 | 0.6467 | 2231.6794 | 0.5662 | 0.1312 | 0.5254 | -5.7724 | 13,721 | ✔️ marks the highest EditLens score (most AI-like reconstructions), ❗ the lowest. Kept rows per encoder x decoder (after the decoders' post-processing): | Encoder \ Decoder | Granite-4.2-30B-Nvfp4 [0] | Ornith-1.5-35B-A3B-Nvfp4 [1] | Llama-3.3-70B-Instruct-Nvfp4 [2] | Qwen3.8-27B-Awq-Int4 [3] | Mistral-Small-4-119B-2603-Nvfp4 [4] | Gemma-4-31B-It-Awq-4Bit [5] | Laguna-S-2.1-Nvfp4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 1,936 | 1,940 | 1,979 | 1,929 | 1,946 | 1,945 | 1,926 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 1,920 | 1,947 | 1,970 | 1,946 | 1,945 | 1,959 | 1,928 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 1,935 | 1,964 | 1,984 | 1,969 | 1,952 | 1,974 | 1,945 | | Qwen3.8-27B-AWQ-INT4 [3] | 1,934 | 1,959 | 1,974 | 1,921 | 1,958 | 1,960 | 1,933 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 1,939 | 1,960 | 1,982 | 1,963 | 1,971 | 1,970 | 1,946 | | gemma-4-31B-it-AWQ-4bit [5] | 1,901 | 1,939 | 1,966 | 1,924 | 1,939 | 1,925 | 1,915 | | Laguna-S-2.1-NVFP4 [6] | 1,936 | 1,948 | 1,981 | 1,954 | 1,953 | 1,963 | 1,927 | | claude-sonnet-5 [8] | 1,942 | 1,959 | 1,987 | 1,965 | 1,957 | 1,970 | 1,952 | | gpt-5.6-terra [9] | 1,950 | 1,957 | 1,982 | 1,959 | 1,970 | 1,964 | 1,939 | Decoder post-processing: - Decoder 0 (`granite-4.2-30b-nvfp4`): 17,393 kept / 603 trashed of 17,996; last pass decoded 9,000 rows; failed requests 6; runtime 4.2 h (rejection reasons, last pass: empty or too short: 83, refusal: 12, filler output: 4, unfilled placeholder: 93, task meta-commentary: 14, echoed instruction: 10, identical to source: 18) - Decoder 1 (`Ornith-1.5-35B-A3B-NVFP4`): 17,573 kept / 423 trashed of 17,996; last pass decoded 15,296 rows; failed requests 2; runtime 2.2 h (rejection reasons, last pass: empty or too short: 23, refusal: 71, filler output: 25, unfilled placeholder: 85, task meta-commentary: 101, echoed instruction: 2, identical to source: 62) - Decoder 2 (`Llama-3.3-70B-Instruct-NVFP4`): 17,805 kept / 191 trashed of 17,996; last pass decoded 15,296 rows; failed requests 2; runtime 9.7 h (rejection reasons, last pass: empty or too short: 45, refusal: 19, filler output: 3, unfilled placeholder: 53, task meta-commentary: 8, echoed instruction: 14, identical to source: 36) - Decoder 3 (`Qwen3.8-27B-AWQ-INT4`): 17,530 kept / 466 trashed of 17,996; last pass decoded 9,000 rows; failed requests 4; runtime 5.9 h (rejection reasons, last pass: empty or too short: 85, refusal: 29, filler output: 4, unfilled placeholder: 69, task meta-commentary: 26, echoed instruction: 3, identical to source: 23) - Decoder 4 (`Mistral-Small-4-119B-2603-NVFP4`): 17,591 kept / 405 trashed of 17,996; last pass decoded 9,000 rows; failed requests 0; runtime 1.6 h (rejection reasons, last pass: empty or too short: 50, refusal: 2, filler output: 7, unfilled placeholder: 151, task meta-commentary: 7, echoed instruction: 2, identical to source: 21) - Decoder 5 (`gemma-4-31B-it-AWQ-4bit`): 17,630 kept / 366 trashed of 17,996; last pass decoded 9,000 rows; failed requests 0; runtime 6.3 h (rejection reasons, last pass: empty or too short: 37, refusal: 104, filler output: 2, unfilled placeholder: 79, task meta-commentary: 24, echoed instruction: 1, identical to source: 20) - Decoder 6 (`Laguna-S-2.1-NVFP4`): 17,411 kept / 585 trashed of 17,996; last pass decoded 9,000 rows; failed requests 5; runtime 2.4 h (rejection reasons, last pass: empty or too short: 35, refusal: 36, filler output: 5, unfilled placeholder: 117, task meta-commentary: 27, echoed instruction: 5, identical to source: 19) ## Encoding length per encoder Length of each encoder's turn-0 output over all of its encodings, in whitespace-separated words and in characters. Empty encodings (failed requests) are counted in `Empty` and left out of the means. A run of emoji without spaces counts as one word, so the character column is given as well. For reference, the source texts average 569 words. | Encoder | Encodings | Mean Words | Median Words | Mean Characters | Empty | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 2,000 | 417.1670 | 248.0000 | 2776.8305 | 0 | | ✔️ Ornith-1.5-35B-A3B-NVFP4 [1] | 2,000 | 509.1720 | 287.0000 | 3291.3920 | 0 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 2,000 | 326.9910 | 208.0000 | 2079.2785 | 0 | | Qwen3.8-27B-AWQ-INT4 [3] | 2,000 | 386.0980 | 247.0000 | 2522.8660 | 0 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 2,000 | 364.5440 | 220.5000 | 2368.4520 | 0 | | ❗ gemma-4-31B-it-AWQ-4bit [5] | 2,000 | 314.0585 | 201.0000 | 1972.3145 | 0 | | Laguna-S-2.1-NVFP4 [6] | 2,000 | 396.7795 | 239.0000 | 2507.4320 | 0 | | claude-sonnet-5 [8] | 1,996 | 446.6914 | 359.0000 | 2927.3612 | 4 | | gpt-5.6-terra [9] | 2,000 | 488.0585 | 254.5000 | 3363.1065 | 0 | ✔️ marks the longest encodings on average, ❗ the shortest. Mean words per encoding family: | Encoder | Descriptive | Partial | Translation | Prompt | |---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 521.5380 | 433.8340 | 571.5000 | 141.7960 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 656.6380 | 498.4680 | 708.1940 | 173.3880 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 332.6500 | 328.9360 | 589.1860 | 57.1920 | | Qwen3.8-27B-AWQ-INT4 [3] | 422.1640 | 373.2620 | 576.4660 | 172.5000 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 434.8280 | 359.7780 | 550.2180 | 113.3520 | | gemma-4-31B-it-AWQ-4bit [5] | 251.1440 | 297.6640 | 590.8420 | 116.5840 | | Laguna-S-2.1-NVFP4 [6] | 388.5400 | 468.5040 | 563.9320 | 166.1420 | | claude-sonnet-5 [8] | 501.8998 | 421.5714 | 595.3340 | 267.9200 | | gpt-5.6-terra [9] | 895.3940 | 325.9700 | 583.7060 | 147.1640 | ## Detection: TPR at 0.1% FPR and AUROC EditLens used as a detector. The threshold is calibrated on the whole human pool of `G-reen/cc-re-2021-filtered`: 387,327 human texts (20 shards, checkpoint pangram/editlens_roberta-large). At a 0.1% false positive budget the threshold is **0.9570**: a text counts as AI when its score is above it, which flags 0.100% of the human pool. As a check, 0.10% of the trial's own 2,000 source texts score above it. `TPR` is the share of decoded texts above the threshold; `AUROC` ranks each slice of decoded texts against the same human pool. ### Detection by encoding family | Slice | AUROC | TPR | N | |---|---|---|---| | All decoded texts | 0.9155 | 0.2572 | 122,933 | | All except translation | 0.9312 | 0.3433 | 91,515 | | descriptive | 0.9646 | 0.4018 | 30,480 | | partial | 0.8614 | 0.1825 | 30,909 | | translation | 0.8701 | 0.0064 | 31,418 | | prompt | 0.9688 | 0.4492 | 30,126 | ### Detection by encoder `TPR Decoder Balanced` is the mean of the per-decoder TPRs. | Encoder | AUROC | TPR | TPR Decoder Balanced | N | |---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.9162 | 0.2635 | 0.2636 | 13,601 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.9048 | 0.2174 | 0.2176 | 13,615 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.9179 | 0.3580 | 0.3582 | 13,723 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.9218 | 0.2362 | 0.2363 | 13,639 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.9219 | 0.3055 | 0.3056 | 13,731 | | gemma-4-31B-it-AWQ-4bit [5] | 0.9176 | 0.2593 | 0.2594 | 13,509 | | Laguna-S-2.1-NVFP4 [6] | 0.9213 | 0.2890 | 0.2891 | 13,662 | | ❗ claude-sonnet-5 [8] | 0.9042 | 0.1681 | 0.1683 | 13,732 | | gpt-5.6-terra [9] | 0.9142 | 0.2178 | 0.2179 | 13,721 | ✔️ marks the highest TPR (easiest to detect), ❗ the lowest. AUROC per encoder and encoding family: | Encoder | Descriptive | Partial | Translation | Prompt | |---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.9638 | 0.8716 | 0.8653 | 0.9665 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.9401 | 0.8680 | 0.8642 | 0.9490 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.9823 | 0.8504 | 0.8505 | 0.9921 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.9702 | 0.8794 | 0.8796 | 0.9596 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.9777 | 0.8473 | 0.8879 | 0.9777 | | gemma-4-31B-it-AWQ-4bit [5] | 0.9778 | 0.8604 | 0.8709 | 0.9675 | | Laguna-S-2.1-NVFP4 [6] | 0.9684 | 0.8754 | 0.8649 | 0.9787 | | claude-sonnet-5 [8] | 0.9500 | 0.8378 | 0.8659 | 0.9653 | | gpt-5.6-terra [9] | 0.9512 | 0.8632 | 0.8815 | 0.9626 | TPR at 0.1% FPR per encoder and encoding family: | Encoder | Descriptive | Partial | Translation | Prompt | |---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.3871 | 0.1555 | 0.0057 | 0.5176 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.3324 | 0.1452 | 0.0069 | 0.3937 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.5504 | 0.2008 | 0.0060 | 0.6897 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.3961 | 0.1889 | 0.0057 | 0.3617 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.4726 | 0.1888 | 0.0046 | 0.5684 | | gemma-4-31B-it-AWQ-4bit [5] | 0.4395 | 0.2285 | 0.0072 | 0.3788 | | Laguna-S-2.1-NVFP4 [6] | 0.4598 | 0.2183 | 0.0049 | 0.4830 | | claude-sonnet-5 [8] | 0.2921 | 0.1025 | 0.0114 | 0.2721 | | gpt-5.6-terra [9] | 0.2879 | 0.2131 | 0.0054 | 0.3716 | ### Detection by decoder | Decoder | AUROC | TPR | N | |---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.9263 | 0.2656 | 17,393 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.8948 | 0.2068 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.8946 | 0.2494 | 17,805 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.9051 | 0.2400 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.9412 | 0.2901 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.9185 | 0.2303 | 17,630 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.9288 | 0.3193 | 17,411 | ### Detection excluding translation The same threshold (0.9570) and human pool, with the decoded texts of the translation family left out (`detection_excluded_families` in the trial config). 91,515 of 122,933 decoded texts remain. By encoder: | Encoder | AUROC | TPR | TPR Decoder Balanced | N | |---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.9337 | 0.3524 | 0.3525 | 10,114 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.9188 | 0.2897 | 0.2900 | 10,133 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.9409 | 0.4783 | 0.4787 | 10,227 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.9364 | 0.3155 | 0.3156 | 10,149 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.9336 | 0.4083 | 0.4084 | 10,235 | | gemma-4-31B-it-AWQ-4bit [5] | 0.9339 | 0.3473 | 0.3476 | 10,013 | | Laguna-S-2.1-NVFP4 [6] | 0.9405 | 0.3862 | 0.3864 | 10,181 | | ❗ claude-sonnet-5 [8] | 0.9174 | 0.2217 | 0.2219 | 10,236 | | gpt-5.6-terra [9] | 0.9253 | 0.2903 | 0.2904 | 10,227 | ✔️ marks the highest TPR (easiest to detect), ❗ the lowest. By decoder: | Decoder | AUROC | TPR | N | |---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.9452 | 0.3556 | 12,906 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.9097 | 0.2768 | 13,081 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.9260 | 0.3325 | 13,312 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.9077 | 0.3201 | 13,043 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.9491 | 0.3848 | 13,096 | | gemma-4-31B-it-AWQ-4bit [5] | 0.9340 | 0.3077 | 13,137 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.9467 | 0.4272 | 12,940 | ### Detection by encoding instruction | [Family] Encoding instruction | AUROC | TPR | N | |---|---|---|---| | [descriptive] Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.9287 | 0.1925 | 5,227 | | [descriptive] Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.9887 | 0.5303 | 5,029 | | [descriptive] Reformat this text into a sensible, structured JSON object. | 0.9542 | 0.2970 | 5,192 | | [descriptive] Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.9876 | 0.6584 | 4,787 | | [descriptive] Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.9429 | 0.2326 | 5,181 | | [descriptive] Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.9890 | 0.5282 | 5,064 | | [partial] Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.9812 | 0.3075 | 6,283 | | [partial] Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.7169 | 0.0333 | 6,073 | | [partial] Summarize this text, attempting to preserve as much of the original language of the text a... | 0.9347 | 0.2352 | 6,196 | | [partial] Take this text and replace one word in every four with a single underscore (use one unders... | 0.7318 | 0.0443 | 6,096 | | [partial] Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.9351 | 0.2843 | 6,261 | | [prompt] Create a prompt that might cause an LLM to generate an output resembling this text. | 0.9916 | 0.6094 | 6,167 | | [prompt] Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.9787 | 0.3712 | 6,148 | | [prompt] Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.9541 | 0.5091 | 5,806 | | [prompt] This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.9836 | 0.4198 | 6,091 | | [prompt] Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.9341 | 0.3348 | 5,914 | | [translation] Translate the given text to French. | 0.8611 | 0.0141 | 7,849 | | [translation] Translate the given text to German. | 0.8799 | 0.0081 | 7,866 | | [translation] Translate the given text to Hindi. | 0.8852 | 0.0032 | 7,843 | | [translation] Translate the given text to Spanish. | 0.8543 | 0.0003 | 7,860 | ## Rows filtered out per encoder Decoded rows that the decoders' post-processing rejected, per encoder (summed over decoders). A rejected row may carry several reasons, so the reason columns can add up to more than `Trashed`. `Empty Encodings At Encode` counts the encoder's own failed requests (those rows were never sent to a decoder). | Encoder | Decoded | Kept | Trashed | Trashed Rate | Echoed Instruction | Empty Or Too Short | Filler Output | Identical To Source | Refusal | Task Meta-Commentary | Unfilled Placeholder | Empty Encodings At Encode | |---|---|---|---|---|---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 14,000 | 13,601 | 399 | 0.0285 | 10 | 38 | 9 | 63 | 68 | 91 | 149 | 0 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 14,000 | 13,615 | 385 | 0.0275 | 17 | 82 | 12 | 70 | 59 | 72 | 98 | 0 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 14,000 | 13,723 | 277 | 0.0198 | 4 | 21 | 2 | 18 | 38 | 63 | 147 | 0 | | Qwen3.8-27B-AWQ-INT4 [3] | 14,000 | 13,639 | 361 | 0.0258 | 3 | 31 | 24 | 144 | 37 | 66 | 79 | 0 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 14,000 | 13,731 | 269 | 0.0192 | 6 | 13 | 10 | 9 | 54 | 73 | 127 | 0 | | gemma-4-31B-it-AWQ-4bit [5] | 14,000 | 13,509 | 491 | 0.0351 | 9 | 184 | 20 | 4 | 88 | 74 | 190 | 0 | | Laguna-S-2.1-NVFP4 [6] | 14,000 | 13,662 | 338 | 0.0241 | 1 | 30 | 21 | 30 | 51 | 88 | 138 | 0 | | claude-sonnet-5 [8] | 13,972 | 13,732 | 240 | 0.0172 | 4 | 15 | 20 | 4 | 24 | 58 | 129 | 4 | | gpt-5.6-terra [9] | 14,000 | 13,721 | 279 | 0.0199 | 9 | 26 | 14 | 39 | 62 | 56 | 105 | 0 | Trashed rows per encoder x decoder: | Encoder \ Decoder | Granite-4.2-30B-Nvfp4 [0] | Ornith-1.5-35B-A3B-Nvfp4 [1] | Llama-3.3-70B-Instruct-Nvfp4 [2] | Qwen3.8-27B-Awq-Int4 [3] | Mistral-Small-4-119B-2603-Nvfp4 [4] | Gemma-4-31B-It-Awq-4Bit [5] | Laguna-S-2.1-Nvfp4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 64 | 60 | 21 | 71 | 54 | 55 | 74 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 80 | 53 | 30 | 54 | 55 | 41 | 72 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 65 | 36 | 16 | 31 | 48 | 26 | 55 | | Qwen3.8-27B-AWQ-INT4 [3] | 66 | 41 | 26 | 79 | 42 | 40 | 67 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 61 | 40 | 18 | 37 | 29 | 30 | 54 | | gemma-4-31B-it-AWQ-4bit [5] | 99 | 61 | 34 | 76 | 61 | 75 | 85 | | Laguna-S-2.1-NVFP4 [6] | 64 | 52 | 19 | 46 | 47 | 37 | 73 | | claude-sonnet-5 [8] | 54 | 37 | 9 | 31 | 39 | 26 | 44 | | gpt-5.6-terra [9] | 50 | 43 | 18 | 41 | 30 | 36 | 61 | ## Contents 1. [EditLens score (higher = more AI-like) (`final_response_editlens_score_roberta_large`)](#editlens-score-higher-more-ai-like-final-response-editlens-score-roberta-large) 2. [EditLens bucket (higher = more AI-like) (`final_response_editlens_bucket_roberta_large`)](#editlens-bucket-higher-more-ai-like-final-response-editlens-bucket-roberta-large) 3. [Jaccard-1 distance to the source (`jaccard_1`)](#jaccard-1-distance-to-the-source-jaccard-1) 4. [Jaccard-2 distance to the source (`jaccard_2`)](#jaccard-2-distance-to-the-source-jaccard-2) 5. [Levenshtein distance to the source (`levenshtein`)](#levenshtein-distance-to-the-source-levenshtein) 6. [Soft n-gram distance to the source (`softngram`)](#soft-n-gram-distance-to-the-source-softngram) 7. [Embedding cosine distance to the source (`cosdist`)](#embedding-cosine-distance-to-the-source-cosdist) 8. [BERTScore distance to the source (`bertscore`)](#bertscore-distance-to-the-source-bertscore) 9. [BERTScore precision distance (`bertscore_precision`)](#bertscore-precision-distance-bertscore-precision) 10. [BERTScore recall distance (`bertscore_recall`)](#bertscore-recall-distance-bertscore-recall) 11. [MoverScore distance to the source (`moverscore`)](#moverscore-distance-to-the-source-moverscore) 12. [Reranker distance to the source (`reranker`)](#reranker-distance-to-the-source-reranker) ## Statistics ### EditLens score (higher = more AI-like) (`final_response_editlens_score_roberta_large`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5097 | 0.3686 | 0.5099 | 0.0442 | 13,601 | 7 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.4634 | 0.3616 | 0.4636 | 0.0513 | 13,615 | 7 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.5693 | 0.3910 | 0.5696 | 0.0428 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.5100 | 0.3573 | 0.5102 | 0.0510 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5442 | 0.3744 | 0.5443 | 0.0352 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | 0.5158 | 0.3681 | 0.5160 | 0.0416 | 13,509 | 7 | | Laguna-S-2.1-NVFP4 [6] | 0.5375 | 0.3731 | 0.5377 | 0.0396 | 13,662 | 7 | | ❗ claude-sonnet-5 [8] | 0.4447 | 0.3417 | 0.4449 | 0.0549 | 13,732 | 7 | | gpt-5.6-terra [9] | 0.4815 | 0.3562 | 0.4816 | 0.0477 | 13,721 | 7 | ![EditLens score (higher = more AI-like) per encoder](ENC_final_response_editlens_score_roberta_large.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5246 | 0.4331 | 0.4760 | 0.4990 | 0.5683 | 0.5047 | 0.5634 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.4805 | 0.3882 | 0.4221 | 0.4360 | 0.5375 | 0.4512 | 0.5297 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.5895 | 0.5133 | 0.5158 | 0.5558 | 0.6326 | 0.5639 | 0.6160 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.5341 | 0.4387 | 0.4735 | 0.4758 | 0.5860 | 0.4908 | 0.5722 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5550 | 0.4919 | 0.5115 | 0.5350 | 0.5877 | 0.5332 | 0.5959 | | gemma-4-31B-it-AWQ-4bit [5] | 0.5474 | 0.4542 | 0.4839 | 0.4957 | 0.5731 | 0.4953 | 0.5623 | | Laguna-S-2.1-NVFP4 [6] | 0.5591 | 0.4763 | 0.5057 | 0.5270 | 0.5976 | 0.5194 | 0.5785 | | claude-sonnet-5 [8] | 0.4777 | 0.3662 | 0.3898 | 0.4359 | 0.5195 | 0.4152 | 0.5100 | | gpt-5.6-terra [9] | 0.5083 | 0.4054 | 0.4386 | 0.4691 | 0.5402 | 0.4666 | 0.5429 | ![EditLens score (higher = more AI-like) encoder x decoder](MATRIX_final_response_editlens_score_roberta_large.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5307 | 0.3627 | 0.5307 | 0.0350 | 17,393 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.4408 | 0.3623 | 0.4408 | 0.0459 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.4685 | 0.3819 | 0.4685 | 0.0409 | 17,805 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.4922 | 0.3652 | 0.4921 | 0.0398 | 17,530 | | ✔️ Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5714 | 0.3511 | 0.5714 | 0.0329 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.4934 | 0.3584 | 0.4934 | 0.0422 | 17,630 | | Laguna-S-2.1-NVFP4 [6] | 0.5634 | 0.3738 | 0.5634 | 0.0308 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.6573 | 0.3009 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.5166 | 0.3330 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 0.8406 | 0.2435 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.6947 | 0.2984 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.7360 | 0.3348 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.7971 | 0.2697 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 0.6119 | 0.3335 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.1625 | 0.2208 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 0.5316 | 0.3415 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.1916 | 0.2520 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.5736 | 0.3555 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.7343 | 0.2903 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.8412 | 0.2699 | 4,787 | | Translate the given text to French. | 0.2129 | 0.1872 | 7,849 | | Translate the given text to German. | 0.2175 | 0.1652 | 7,866 | | Translate the given text to Hindi. | 0.2248 | 0.1603 | 7,843 | | Translate the given text to Spanish. | 0.1742 | 0.1259 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.6100 | 0.3620 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.5776 | 0.3326 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.8114 | 0.2575 | 5,064 | ### EditLens bucket (higher = more AI-like) (`final_response_editlens_bucket_roberta_large`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 1.4513 | 1.2834 | 1.4517 | 0.1459 | 13,601 | 7 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 1.2968 | 1.2665 | 1.2975 | 0.1688 | 13,615 | 7 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 1.6525 | 1.3439 | 1.6533 | 0.1401 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 1.4582 | 1.2665 | 1.4585 | 0.1634 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 1.5681 | 1.2994 | 1.5684 | 0.1158 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | 1.4674 | 1.2950 | 1.4680 | 0.1366 | 13,509 | 7 | | Laguna-S-2.1-NVFP4 [6] | 1.5437 | 1.3019 | 1.5442 | 0.1286 | 13,662 | 7 | | ❗ claude-sonnet-5 [8] | 1.2356 | 1.2144 | 1.2363 | 0.1761 | 13,732 | 7 | | gpt-5.6-terra [9] | 1.3660 | 1.2578 | 1.3664 | 0.1548 | 13,721 | 7 | ![EditLens bucket (higher = more AI-like) per encoder](ENC_final_response_editlens_bucket_roberta_large.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 1.5052 | 1.1995 | 1.3380 | 1.4142 | 1.6434 | 1.4334 | 1.6282 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 1.3490 | 1.0508 | 1.1503 | 1.2122 | 1.5429 | 1.2660 | 1.5114 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 1.7235 | 1.4638 | 1.4788 | 1.6074 | 1.8637 | 1.6424 | 1.7933 | | Qwen3.8-27B-AWQ-INT4 [3] | 1.5476 | 1.2297 | 1.3440 | 1.3514 | 1.7079 | 1.3852 | 1.6441 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 1.6055 | 1.4010 | 1.4470 | 1.5420 | 1.7164 | 1.5365 | 1.7302 | | gemma-4-31B-it-AWQ-4bit [5] | 1.5786 | 1.2702 | 1.3535 | 1.4023 | 1.6596 | 1.4005 | 1.6115 | | Laguna-S-2.1-NVFP4 [6] | 1.6105 | 1.3393 | 1.4523 | 1.5005 | 1.7373 | 1.4885 | 1.6809 | | claude-sonnet-5 [8] | 1.3404 | 0.9837 | 1.0564 | 1.2300 | 1.4798 | 1.1294 | 1.4344 | | gpt-5.6-terra [9] | 1.4554 | 1.1185 | 1.2200 | 1.3313 | 1.5553 | 1.3228 | 1.5616 | ![EditLens bucket (higher = more AI-like) encoder x decoder](MATRIX_final_response_editlens_bucket_roberta_large.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 1.5239 | 1.2686 | 1.5240 | 0.1187 | 17,393 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 1.2286 | 1.2696 | 1.2285 | 0.1504 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 1.3156 | 1.3528 | 1.3156 | 0.1370 | 17,805 | | Qwen3.8-27B-AWQ-INT4 [3] | 1.3993 | 1.2700 | 1.3990 | 0.1268 | 17,530 | | ✔️ Mistral-Small-4-119B-2603-NVFP4 [4] | 1.6562 | 1.2271 | 1.6563 | 0.1107 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 1.4006 | 1.2622 | 1.4005 | 0.1429 | 17,630 | | Laguna-S-2.1-NVFP4 [6] | 1.6217 | 1.2995 | 1.6217 | 0.1034 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 1.9452 | 1.1293 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 1.4791 | 1.2074 | 5,227 | | ✔️ Create a prompt that might cause an LLM to generate an output resembling this text. | 2.5465 | 0.8898 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 2.0597 | 1.1038 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 2.1988 | 1.1656 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 2.3941 | 0.9896 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 1.7987 | 1.2107 | 5,192 | | Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.3359 | 0.7415 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 1.5350 | 1.2135 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.4221 | 0.8448 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 1.6790 | 1.2669 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 2.2033 | 1.0782 | 6,091 | | Translate the entirety of this text into a sequence of emojis that captures the literal me... | 2.5402 | 0.9567 | 4,787 | | Translate the given text to French. | 0.4640 | 0.7240 | 7,849 | | Translate the given text to German. | 0.4652 | 0.6533 | 7,866 | | Translate the given text to Hindi. | 0.4867 | 0.6452 | 7,843 | | ❗ Translate the given text to Spanish. | 0.3095 | 0.4888 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 1.7973 | 1.2894 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 1.6889 | 1.2322 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 2.4690 | 0.9505 | 5,064 | ### Jaccard-1 distance to the source (`jaccard_1`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.6346 | 0.2601 | 0.6346 | 0.0258 | 13,601 | 7 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.6167 | 0.2521 | 0.6168 | 0.0279 | 13,615 | 7 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.6610 | 0.2578 | 0.6611 | 0.0161 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.6467 | 0.2441 | 0.6467 | 0.0256 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.6656 | 0.2437 | 0.6657 | 0.0157 | 13,731 | 7 | | ✔️ gemma-4-31B-it-AWQ-4bit [5] | 0.6694 | 0.2292 | 0.6694 | 0.0160 | 13,509 | 7 | | Laguna-S-2.1-NVFP4 [6] | 0.6693 | 0.2403 | 0.6693 | 0.0164 | 13,662 | 7 | | ❗ claude-sonnet-5 [8] | 0.6123 | 0.2314 | 0.6124 | 0.0245 | 13,732 | 7 | | gpt-5.6-terra [9] | 0.6466 | 0.2298 | 0.6467 | 0.0181 | 13,721 | 7 | ![Jaccard-1 distance to the source per encoder](ENC_jaccard_1.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.6463 | 0.6076 | 0.6239 | 0.6010 | 0.6606 | 0.6260 | 0.6772 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.6278 | 0.5895 | 0.6027 | 0.5818 | 0.6492 | 0.6055 | 0.6612 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.6699 | 0.6500 | 0.6458 | 0.6459 | 0.6817 | 0.6491 | 0.6853 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.6606 | 0.6232 | 0.6317 | 0.6163 | 0.6767 | 0.6321 | 0.6867 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.6693 | 0.6566 | 0.6538 | 0.6475 | 0.6784 | 0.6580 | 0.6961 | | gemma-4-31B-it-AWQ-4bit [5] | 0.6770 | 0.6637 | 0.6540 | 0.6560 | 0.6861 | 0.6528 | 0.6964 | | Laguna-S-2.1-NVFP4 [6] | 0.6762 | 0.6566 | 0.6545 | 0.6526 | 0.6890 | 0.6605 | 0.6960 | | claude-sonnet-5 [8] | 0.6257 | 0.5912 | 0.5939 | 0.5895 | 0.6392 | 0.5939 | 0.6530 | | gpt-5.6-terra [9] | 0.6586 | 0.6366 | 0.6353 | 0.6298 | 0.6601 | 0.6270 | 0.6793 | ![Jaccard-1 distance to the source encoder x decoder](MATRIX_jaccard_1.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.6568 | 0.2384 | 0.6568 | 0.0184 | 17,393 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.6305 | 0.2512 | 0.6305 | 0.0272 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.6328 | 0.2425 | 0.6328 | 0.0212 | 17,805 | | ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.6245 | 0.2560 | 0.6245 | 0.0268 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.6690 | 0.2330 | 0.6690 | 0.0164 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.6338 | 0.2510 | 0.6339 | 0.0221 | 17,630 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.6812 | 0.2307 | 0.6813 | 0.0146 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.7801 | 0.0888 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.5981 | 0.2342 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 0.8516 | 0.0511 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.7967 | 0.0909 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.8225 | 0.1740 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.8534 | 0.0621 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 0.6797 | 0.1892 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.2187 | 0.1756 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 0.6119 | 0.2350 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.2510 | 0.2141 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.7016 | 0.1656 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.8229 | 0.0692 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.9088 | 0.0530 | 4,787 | | Translate the given text to French. | 0.4678 | 0.0828 | 7,849 | | Translate the given text to German. | 0.4749 | 0.0869 | 7,866 | | Translate the given text to Hindi. | 0.5467 | 0.0988 | 7,843 | | Translate the given text to Spanish. | 0.4175 | 0.0952 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.8199 | 0.1645 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.7838 | 0.0908 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.8804 | 0.0349 | 5,064 | ### Jaccard-2 distance to the source (`jaccard_2`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.7686 | 0.2655 | 0.7686 | 0.0261 | 13,601 | 7 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.7522 | 0.2543 | 0.7523 | 0.0270 | 13,615 | 7 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.7888 | 0.2464 | 0.7889 | 0.0160 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.7862 | 0.2492 | 0.7862 | 0.0247 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.8011 | 0.2310 | 0.8012 | 0.0141 | 13,731 | 7 | | ✔️ gemma-4-31B-it-AWQ-4bit [5] | 0.8119 | 0.2044 | 0.8120 | 0.0145 | 13,509 | 7 | | Laguna-S-2.1-NVFP4 [6] | 0.8055 | 0.2323 | 0.8056 | 0.0149 | 13,662 | 7 | | claude-sonnet-5 [8] | 0.7615 | 0.2245 | 0.7616 | 0.0229 | 13,732 | 7 | | gpt-5.6-terra [9] | 0.7943 | 0.2228 | 0.7944 | 0.0159 | 13,721 | 7 | ![Jaccard-2 distance to the source per encoder](ENC_jaccard_2.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.7805 | 0.7394 | 0.7586 | 0.7356 | 0.7982 | 0.7589 | 0.8088 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.7624 | 0.7233 | 0.7415 | 0.7183 | 0.7869 | 0.7416 | 0.7920 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.7990 | 0.7783 | 0.7744 | 0.7760 | 0.8106 | 0.7730 | 0.8107 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.8004 | 0.7623 | 0.7725 | 0.7585 | 0.8171 | 0.7700 | 0.8226 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.8056 | 0.7942 | 0.7907 | 0.7864 | 0.8126 | 0.7904 | 0.8284 | | gemma-4-31B-it-AWQ-4bit [5] | 0.8201 | 0.8091 | 0.7968 | 0.8040 | 0.8258 | 0.7927 | 0.8353 | | Laguna-S-2.1-NVFP4 [6] | 0.8115 | 0.7939 | 0.7923 | 0.7923 | 0.8253 | 0.7949 | 0.8287 | | claude-sonnet-5 [8] | 0.7745 | 0.7422 | 0.7436 | 0.7423 | 0.7890 | 0.7420 | 0.7977 | | gpt-5.6-terra [9] | 0.8053 | 0.7868 | 0.7846 | 0.7812 | 0.8078 | 0.7738 | 0.8212 | ![Jaccard-2 distance to the source encoder x decoder](MATRIX_jaccard_2.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.7955 | 0.2293 | 0.7955 | 0.0178 | 17,393 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.7699 | 0.2493 | 0.7699 | 0.0279 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.7728 | 0.2373 | 0.7728 | 0.0196 | 17,805 | | ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.7661 | 0.2594 | 0.7661 | 0.0273 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.8081 | 0.2196 | 0.8081 | 0.0135 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.7708 | 0.2469 | 0.7708 | 0.0191 | 17,630 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.8161 | 0.2168 | 0.8162 | 0.0139 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.9255 | 0.0667 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.7426 | 0.2499 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 0.9673 | 0.0266 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.9395 | 0.0595 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.9245 | 0.1811 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.9635 | 0.0383 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 0.8286 | 0.2053 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.3124 | 0.2106 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 0.7485 | 0.2589 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.3533 | 0.2535 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.8228 | 0.1761 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.9530 | 0.0421 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.9851 | 0.0361 | 4,787 | | Translate the given text to French. | 0.6669 | 0.0862 | 7,849 | | Translate the given text to German. | 0.6758 | 0.0859 | 7,866 | | Translate the given text to Hindi. | 0.7480 | 0.0904 | 7,843 | | Translate the given text to Spanish. | 0.6132 | 0.1066 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.9063 | 0.1566 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.9089 | 0.0773 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.9775 | 0.0155 | 5,064 | ### Levenshtein distance to the source (`levenshtein`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 2105.0217 | 2445.1394 | 2105.3807 | 145.5901 | 13,601 | 7 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 2041.9071 | 2797.1727 | 2042.6838 | 141.2548 | 13,615 | 7 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 2131.9093 | 2515.1605 | 2132.4408 | 105.1155 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 2065.1839 | 2758.4968 | 2065.3474 | 121.7832 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 2237.2207 | 2992.4775 | 2237.7226 | 140.1776 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | 2073.1535 | 2457.9207 | 2073.2000 | 96.0188 | 13,509 | 7 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 2308.2526 | 3898.7931 | 2309.0983 | 148.3083 | 13,662 | 7 | | claude-sonnet-5 [8] | 2085.9876 | 4508.4155 | 2086.9193 | 226.0749 | 13,732 | 7 | | gpt-5.6-terra [9] | 2231.2725 | 4426.4117 | 2231.6794 | 182.3787 | 13,721 | 7 | ![Levenshtein distance to the source per encoder](ENC_levenshtein.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 2142.1364 | 2149.0258 | 2017.1107 | 1949.2789 | 2266.4157 | 1898.2540 | 2315.4434 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 2092.6406 | 2028.9692 | 1936.7614 | 1885.9527 | 2146.8653 | 1900.2323 | 2307.3651 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 2130.6997 | 2194.5234 | 2063.4793 | 2048.6953 | 2279.5835 | 1966.1864 | 2243.9177 | | Qwen3.8-27B-AWQ-INT4 [3] | 2106.0796 | 2079.4421 | 2021.3582 | 1928.7715 | 2151.6813 | 1893.8138 | 2276.2856 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 2298.4471 | 2221.7383 | 2154.4147 | 2111.7040 | 2388.3876 | 2034.9360 | 2454.4301 | | gemma-4-31B-it-AWQ-4bit [5] | 2073.0352 | 2081.1078 | 2048.4593 | 2002.4683 | 2186.3847 | 1909.2691 | 2211.6757 | | Laguna-S-2.1-NVFP4 [6] | 2430.5517 | 2425.7295 | 2221.5962 | 2134.4831 | 2412.0691 | 2077.6032 | 2461.6554 | | claude-sonnet-5 [8] | 2173.4439 | 2217.3492 | 1941.5279 | 1888.4061 | 2049.8426 | 1809.8386 | 2528.0272 | | gpt-5.6-terra [9] | 2389.4651 | 2290.8809 | 2351.7992 | 2047.5268 | 2169.4046 | 1914.2164 | 2458.4631 | ![Levenshtein distance to the source encoder x decoder](MATRIX_levenshtein.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 2204.5921 | 3731.8047 | 2204.0555 | 126.4551 | 17,393 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 2187.7551 | 3970.7840 | 2187.6407 | 115.0659 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 2084.1770 | 3389.4024 | 2084.0563 | 128.0916 | 17,805 | | Qwen3.8-27B-AWQ-INT4 [3] | 1999.9944 | 2464.4671 | 1999.6985 | 87.1174 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 2227.9455 | 2532.5306 | 2227.8483 | 112.0787 | 17,591 | | ❗ gemma-4-31B-it-AWQ-4bit [5] | 1933.9303 | 2608.9887 | 1933.8166 | 76.2289 | 17,630 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 2362.2757 | 3968.4034 | 2361.9182 | 107.6955 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 2708.4195 | 2321.3652 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 2036.5969 | 2627.1115 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 3316.2526 | 2860.5408 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 2835.4305 | 2569.2781 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 2806.7494 | 2901.9416 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 3323.9753 | 3848.3830 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 2593.6308 | 3057.3767 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 430.3017 | 2691.7751 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 2146.6162 | 4226.0453 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 559.8406 | 1643.9632 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 2577.7528 | 2548.4190 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 3104.4745 | 4701.8958 | 6,091 | | Translate the entirety of this text into a sequence of emojis that captures the literal me... | 3359.8018 | 5816.8765 | 4,787 | | Translate the given text to French. | 946.0595 | 1025.3997 | 7,849 | | Translate the given text to German. | 989.9259 | 1223.0919 | 7,866 | | Translate the given text to Hindi. | 1085.5447 | 1783.5637 | 7,843 | | Translate the given text to Spanish. | 752.6126 | 643.9184 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 2950.4307 | 4422.2077 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 2858.1764 | 5017.5288 | 5,181 | | ✔️ Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 3750.6929 | 3652.7911 | 5,064 | ### Soft n-gram distance to the source (`softngram`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5588 | 0.3671 | 0.5588 | 0.0321 | 13,601 | 7 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.5053 | 0.3597 | 0.5054 | 0.0369 | 13,615 | 7 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.5875 | 0.3806 | 0.5877 | 0.0236 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.5612 | 0.3529 | 0.5613 | 0.0336 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5831 | 0.3566 | 0.5832 | 0.0219 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | 0.5850 | 0.3570 | 0.5851 | 0.0245 | 13,509 | 7 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.6052 | 0.3523 | 0.6054 | 0.0232 | 13,662 | 7 | | claude-sonnet-5 [8] | 0.5249 | 0.3538 | 0.5250 | 0.0339 | 13,732 | 7 | | gpt-5.6-terra [9] | 0.5661 | 0.3447 | 0.5662 | 0.0280 | 13,721 | 7 | ![Soft n-gram distance to the source per encoder](ENC_softngram.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5715 | 0.5290 | 0.5451 | 0.5177 | 0.5862 | 0.5455 | 0.6167 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.5126 | 0.4800 | 0.4925 | 0.4518 | 0.5372 | 0.4909 | 0.5729 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.6011 | 0.5743 | 0.5717 | 0.5551 | 0.6130 | 0.5735 | 0.6250 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.5745 | 0.5373 | 0.5422 | 0.5143 | 0.5924 | 0.5482 | 0.6201 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5851 | 0.5703 | 0.5694 | 0.5560 | 0.5976 | 0.5760 | 0.6278 | | gemma-4-31B-it-AWQ-4bit [5] | 0.5943 | 0.5784 | 0.5678 | 0.5507 | 0.6087 | 0.5686 | 0.6270 | | Laguna-S-2.1-NVFP4 [6] | 0.6179 | 0.5907 | 0.5855 | 0.5775 | 0.6292 | 0.5929 | 0.6441 | | claude-sonnet-5 [8] | 0.5416 | 0.4958 | 0.4988 | 0.4959 | 0.5601 | 0.4991 | 0.5838 | | gpt-5.6-terra [9] | 0.5854 | 0.5493 | 0.5494 | 0.5377 | 0.5856 | 0.5386 | 0.6173 | ![Soft n-gram distance to the source encoder x decoder](MATRIX_softngram.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5760 | 0.3595 | 0.5760 | 0.0300 | 17,393 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.5450 | 0.3682 | 0.5450 | 0.0360 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.5469 | 0.3481 | 0.5469 | 0.0305 | 17,805 | | ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.5286 | 0.3670 | 0.5285 | 0.0360 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5900 | 0.3569 | 0.5900 | 0.0262 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.5481 | 0.3576 | 0.5482 | 0.0327 | 17,630 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.6149 | 0.3524 | 0.6150 | 0.0212 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.7411 | 0.1765 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.5552 | 0.3115 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 0.9115 | 0.1060 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.8202 | 0.1903 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.7770 | 0.2809 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.9276 | 0.0950 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 0.6376 | 0.2604 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.1051 | 0.1730 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 0.5004 | 0.3040 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.1732 | 0.2321 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.5652 | 0.2894 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.8686 | 0.1397 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.9783 | 0.0735 | 4,787 | | Translate the given text to French. | 0.2061 | 0.1110 | 7,849 | | Translate the given text to German. | 0.2029 | 0.1118 | 7,866 | | Translate the given text to Hindi. | 0.3047 | 0.1719 | 7,843 | | Translate the given text to Spanish. | 0.1621 | 0.0921 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.7272 | 0.3211 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.8208 | 0.1449 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.9684 | 0.0474 | 5,064 | ### Embedding cosine distance to the source (`cosdist`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1582 | 0.1651 | 0.1582 | 0.0086 | 13,601 | 7 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1559 | 0.1760 | 0.1559 | 0.0090 | 13,615 | 7 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.2061 | 0.2024 | 0.2060 | 0.0102 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1580 | 0.1646 | 0.1580 | 0.0090 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1762 | 0.1806 | 0.1761 | 0.0085 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1720 | 0.1695 | 0.1720 | 0.0096 | 13,509 | 7 | | Laguna-S-2.1-NVFP4 [6] | 0.1788 | 0.1791 | 0.1788 | 0.0082 | 13,662 | 7 | | ❗ claude-sonnet-5 [8] | 0.1315 | 0.1443 | 0.1315 | 0.0094 | 13,732 | 7 | | gpt-5.6-terra [9] | 0.1365 | 0.1465 | 0.1365 | 0.0096 | 13,721 | 7 | ![Embedding cosine distance to the source per encoder](ENC_cosdist.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1562 | 0.1426 | 0.1717 | 0.1587 | 0.1532 | 0.1590 | 0.1661 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1524 | 0.1436 | 0.1692 | 0.1500 | 0.1588 | 0.1495 | 0.1676 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1997 | 0.1879 | 0.2225 | 0.2091 | 0.2024 | 0.2067 | 0.2139 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1571 | 0.1415 | 0.1706 | 0.1559 | 0.1615 | 0.1521 | 0.1672 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1680 | 0.1619 | 0.1890 | 0.1763 | 0.1772 | 0.1762 | 0.1845 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1721 | 0.1560 | 0.1887 | 0.1682 | 0.1740 | 0.1656 | 0.1795 | | Laguna-S-2.1-NVFP4 [6] | 0.1751 | 0.1625 | 0.1896 | 0.1809 | 0.1802 | 0.1766 | 0.1866 | | claude-sonnet-5 [8] | 0.1332 | 0.1157 | 0.1420 | 0.1266 | 0.1381 | 0.1226 | 0.1422 | | gpt-5.6-terra [9] | 0.1377 | 0.1238 | 0.1501 | 0.1329 | 0.1402 | 0.1240 | 0.1470 | ![Embedding cosine distance to the source encoder x decoder](MATRIX_cosdist.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1612 | 0.1634 | 0.1613 | 0.0191 | 17,393 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1484 | 0.1683 | 0.1484 | 0.0205 | 17,573 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1770 | 0.1841 | 0.1770 | 0.0226 | 17,805 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1621 | 0.1730 | 0.1621 | 0.0238 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1650 | 0.1715 | 0.1651 | 0.0194 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1591 | 0.1694 | 0.1591 | 0.0250 | 17,630 | | Laguna-S-2.1-NVFP4 [6] | 0.1727 | 0.1716 | 0.1727 | 0.0205 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.1574 | 0.0848 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.1183 | 0.0921 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 0.2321 | 0.1096 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.1822 | 0.1116 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.2315 | 0.1545 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.2712 | 0.1404 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 0.1178 | 0.0733 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.0215 | 0.0535 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 0.0988 | 0.0787 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.0418 | 0.0735 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.1571 | 0.0973 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.1941 | 0.0977 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.6000 | 0.1837 | 4,787 | | Translate the given text to French. | 0.0333 | 0.0372 | 7,849 | | Translate the given text to German. | 0.0306 | 0.0363 | 7,866 | | Translate the given text to Hindi. | 0.0550 | 0.0635 | 7,843 | | Translate the given text to Spanish. | 0.0319 | 0.0330 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.3234 | 0.1773 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.2848 | 0.1359 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.4019 | 0.1472 | 5,064 | ### BERTScore distance to the source (`bertscore`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1327 | 0.0748 | 0.1327 | 0.0057 | 13,601 | 7 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1265 | 0.0736 | 0.1265 | 0.0066 | 13,615 | 7 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1448 | 0.0786 | 0.1448 | 0.0042 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1335 | 0.0702 | 0.1336 | 0.0060 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1415 | 0.0736 | 0.1415 | 0.0042 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1410 | 0.0712 | 0.1410 | 0.0042 | 13,509 | 7 | | Laguna-S-2.1-NVFP4 [6] | 0.1424 | 0.0726 | 0.1424 | 0.0042 | 13,662 | 7 | | ❗ claude-sonnet-5 [8] | 0.1213 | 0.0663 | 0.1214 | 0.0068 | 13,732 | 7 | | gpt-5.6-terra [9] | 0.1312 | 0.0672 | 0.1312 | 0.0054 | 13,721 | 7 | ![BERTScore distance to the source per encoder](ENC_bertscore.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1343 | 0.1259 | 0.1318 | 0.1259 | 0.1377 | 0.1305 | 0.1429 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1271 | 0.1196 | 0.1251 | 0.1186 | 0.1337 | 0.1236 | 0.1380 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1447 | 0.1402 | 0.1459 | 0.1396 | 0.1495 | 0.1421 | 0.1514 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1349 | 0.1275 | 0.1309 | 0.1270 | 0.1405 | 0.1303 | 0.1437 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1401 | 0.1373 | 0.1414 | 0.1367 | 0.1448 | 0.1406 | 0.1498 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1411 | 0.1379 | 0.1399 | 0.1367 | 0.1460 | 0.1370 | 0.1483 | | Laguna-S-2.1-NVFP4 [6] | 0.1431 | 0.1379 | 0.1415 | 0.1377 | 0.1474 | 0.1398 | 0.1494 | | claude-sonnet-5 [8] | 0.1237 | 0.1147 | 0.1189 | 0.1153 | 0.1283 | 0.1152 | 0.1333 | | gpt-5.6-terra [9] | 0.1331 | 0.1274 | 0.1309 | 0.1262 | 0.1349 | 0.1247 | 0.1415 | ![BERTScore distance to the source encoder x decoder](MATRIX_bertscore.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1358 | 0.0706 | 0.1358 | 0.0067 | 17,393 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1298 | 0.0725 | 0.1298 | 0.0086 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1340 | 0.0735 | 0.1340 | 0.0083 | 17,805 | | ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.1293 | 0.0730 | 0.1293 | 0.0083 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1403 | 0.0702 | 0.1403 | 0.0068 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1315 | 0.0731 | 0.1315 | 0.0086 | 17,630 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.1443 | 0.0731 | 0.1443 | 0.0057 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.1661 | 0.0365 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.1156 | 0.0562 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 0.1916 | 0.0285 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.1683 | 0.0382 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.1906 | 0.0575 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.1954 | 0.0325 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 0.1343 | 0.0454 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.0344 | 0.0385 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 0.1162 | 0.0549 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.0452 | 0.0461 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.1438 | 0.0473 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.1814 | 0.0321 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.2518 | 0.0558 | 4,787 | | Translate the given text to French. | 0.0727 | 0.0219 | 7,849 | | Translate the given text to German. | 0.0730 | 0.0234 | 7,866 | | Translate the given text to Hindi. | 0.0900 | 0.0315 | 7,843 | | Translate the given text to Spanish. | 0.0642 | 0.0218 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.1984 | 0.0629 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.1752 | 0.0336 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.2160 | 0.0249 | 5,064 | ### BERTScore precision distance (`bertscore_precision`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1325 | 0.0768 | 0.1325 | 0.0079 | 13,601 | 7 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1235 | 0.0741 | 0.1236 | 0.0087 | 13,615 | 7 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1414 | 0.0803 | 0.1414 | 0.0065 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1315 | 0.0699 | 0.1315 | 0.0084 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1392 | 0.0744 | 0.1392 | 0.0062 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1370 | 0.0708 | 0.1370 | 0.0064 | 13,509 | 7 | | Laguna-S-2.1-NVFP4 [6] | 0.1414 | 0.0746 | 0.1414 | 0.0063 | 13,662 | 7 | | ❗ claude-sonnet-5 [8] | 0.1226 | 0.0699 | 0.1226 | 0.0088 | 13,732 | 7 | | gpt-5.6-terra [9] | 0.1314 | 0.0693 | 0.1314 | 0.0077 | 13,721 | 7 | ![BERTScore precision distance per encoder](ENC_bertscore_precision.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1342 | 0.1240 | 0.1309 | 0.1233 | 0.1401 | 0.1284 | 0.1466 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1238 | 0.1155 | 0.1222 | 0.1127 | 0.1320 | 0.1195 | 0.1394 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1416 | 0.1357 | 0.1418 | 0.1330 | 0.1495 | 0.1366 | 0.1516 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1331 | 0.1240 | 0.1287 | 0.1212 | 0.1402 | 0.1270 | 0.1464 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1375 | 0.1339 | 0.1381 | 0.1324 | 0.1446 | 0.1366 | 0.1514 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1368 | 0.1329 | 0.1350 | 0.1298 | 0.1444 | 0.1316 | 0.1486 | | Laguna-S-2.1-NVFP4 [6] | 0.1423 | 0.1358 | 0.1398 | 0.1345 | 0.1486 | 0.1365 | 0.1522 | | claude-sonnet-5 [8] | 0.1251 | 0.1140 | 0.1184 | 0.1165 | 0.1308 | 0.1145 | 0.1391 | | gpt-5.6-terra [9] | 0.1335 | 0.1257 | 0.1300 | 0.1246 | 0.1368 | 0.1227 | 0.1464 | ![BERTScore precision distance encoder x decoder](MATRIX_bertscore_precision.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1342 | 0.0721 | 0.1342 | 0.0061 | 17,393 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1268 | 0.0735 | 0.1268 | 0.0078 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1316 | 0.0740 | 0.1316 | 0.0074 | 17,805 | | ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.1253 | 0.0732 | 0.1253 | 0.0073 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1408 | 0.0730 | 0.1408 | 0.0063 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1282 | 0.0713 | 0.1282 | 0.0076 | 17,630 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.1468 | 0.0765 | 0.1469 | 0.0046 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.1661 | 0.0385 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.1205 | 0.0605 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 0.1980 | 0.0336 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.1741 | 0.0403 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.1771 | 0.0587 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.2033 | 0.0377 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 0.1409 | 0.0515 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.0371 | 0.0452 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 0.1159 | 0.0608 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.0485 | 0.0527 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.1303 | 0.0563 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.1883 | 0.0367 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.2467 | 0.0614 | 4,787 | | Translate the given text to French. | 0.0711 | 0.0211 | 7,849 | | Translate the given text to German. | 0.0715 | 0.0226 | 7,866 | | Translate the given text to Hindi. | 0.0867 | 0.0297 | 7,843 | | Translate the given text to Spanish. | 0.0621 | 0.0213 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.1751 | 0.0666 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.1651 | 0.0366 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.2148 | 0.0295 | 5,064 | ### BERTScore recall distance (`bertscore_recall`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1321 | 0.0765 | 0.1321 | 0.0036 | 13,601 | 7 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1283 | 0.0783 | 0.1283 | 0.0046 | 13,615 | 7 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1474 | 0.0802 | 0.1474 | 0.0021 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1346 | 0.0749 | 0.1346 | 0.0037 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1430 | 0.0767 | 0.1430 | 0.0024 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1440 | 0.0755 | 0.1440 | 0.0021 | 13,509 | 7 | | Laguna-S-2.1-NVFP4 [6] | 0.1426 | 0.0743 | 0.1426 | 0.0022 | 13,662 | 7 | | ❗ claude-sonnet-5 [8] | 0.1195 | 0.0657 | 0.1195 | 0.0049 | 13,732 | 7 | | gpt-5.6-terra [9] | 0.1304 | 0.0686 | 0.1304 | 0.0032 | 13,721 | 7 | ![BERTScore recall distance per encoder](ENC_bertscore_recall.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1336 | 0.1269 | 0.1318 | 0.1279 | 0.1346 | 0.1317 | 0.1383 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1292 | 0.1226 | 0.1270 | 0.1233 | 0.1341 | 0.1266 | 0.1354 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1471 | 0.1440 | 0.1494 | 0.1452 | 0.1488 | 0.1470 | 0.1505 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1357 | 0.1302 | 0.1322 | 0.1317 | 0.1398 | 0.1327 | 0.1400 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1419 | 0.1400 | 0.1438 | 0.1403 | 0.1441 | 0.1439 | 0.1474 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1445 | 0.1421 | 0.1440 | 0.1424 | 0.1467 | 0.1416 | 0.1471 | | Laguna-S-2.1-NVFP4 [6] | 0.1432 | 0.1393 | 0.1423 | 0.1401 | 0.1454 | 0.1423 | 0.1456 | | claude-sonnet-5 [8] | 0.1218 | 0.1147 | 0.1189 | 0.1136 | 0.1253 | 0.1154 | 0.1268 | | gpt-5.6-terra [9] | 0.1319 | 0.1283 | 0.1313 | 0.1270 | 0.1325 | 0.1261 | 0.1358 | ![BERTScore recall distance encoder x decoder](MATRIX_bertscore_recall.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.1365 | 0.0733 | 0.1365 | 0.0078 | 17,393 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | 0.1320 | 0.0750 | 0.1320 | 0.0094 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.1356 | 0.0768 | 0.1356 | 0.0093 | 17,805 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.1324 | 0.0768 | 0.1324 | 0.0098 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.1390 | 0.0715 | 0.1390 | 0.0074 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.1341 | 0.0780 | 0.1341 | 0.0098 | 17,630 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.1407 | 0.0738 | 0.1408 | 0.0071 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.1653 | 0.0420 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.1100 | 0.0565 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 0.1846 | 0.0300 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.1621 | 0.0399 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.2018 | 0.0669 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.1865 | 0.0357 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 0.1270 | 0.0452 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.0314 | 0.0339 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 0.1154 | 0.0563 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.0413 | 0.0428 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.1555 | 0.0498 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.1737 | 0.0332 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.2558 | 0.0558 | 4,787 | | Translate the given text to French. | 0.0742 | 0.0235 | 7,849 | | Translate the given text to German. | 0.0744 | 0.0252 | 7,866 | | Translate the given text to Hindi. | 0.0930 | 0.0345 | 7,843 | | Translate the given text to Spanish. | 0.0662 | 0.0234 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.2188 | 0.0693 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.1839 | 0.0418 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.2165 | 0.0312 | 5,064 | ### MoverScore distance to the source (`moverscore`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5235 | 0.1911 | 0.5235 | 0.0171 | 13,601 | 7 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.5114 | 0.1890 | 0.5114 | 0.0182 | 13,615 | 7 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.5472 | 0.1958 | 0.5472 | 0.0102 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.5289 | 0.1821 | 0.5289 | 0.0164 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5456 | 0.1828 | 0.5456 | 0.0108 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | 0.5459 | 0.1768 | 0.5459 | 0.0106 | 13,509 | 7 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.5484 | 0.1805 | 0.5484 | 0.0110 | 13,662 | 7 | | ❗ claude-sonnet-5 [8] | 0.5001 | 0.1718 | 0.5002 | 0.0173 | 13,732 | 7 | | gpt-5.6-terra [9] | 0.5254 | 0.1715 | 0.5254 | 0.0136 | 13,721 | 7 | ![MoverScore distance to the source per encoder](ENC_moverscore.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5294 | 0.5031 | 0.5236 | 0.5010 | 0.5393 | 0.5162 | 0.5517 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.5166 | 0.4907 | 0.5099 | 0.4887 | 0.5320 | 0.5019 | 0.5402 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.5490 | 0.5373 | 0.5481 | 0.5364 | 0.5600 | 0.5368 | 0.5628 | | Qwen3.8-27B-AWQ-INT4 [3] | 0.5353 | 0.5112 | 0.5249 | 0.5104 | 0.5473 | 0.5179 | 0.5554 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5446 | 0.5363 | 0.5450 | 0.5329 | 0.5543 | 0.5394 | 0.5668 | | gemma-4-31B-it-AWQ-4bit [5] | 0.5478 | 0.5399 | 0.5418 | 0.5382 | 0.5573 | 0.5320 | 0.5644 | | Laguna-S-2.1-NVFP4 [6] | 0.5506 | 0.5377 | 0.5462 | 0.5365 | 0.5615 | 0.5398 | 0.5668 | | claude-sonnet-5 [8] | 0.5075 | 0.4838 | 0.4949 | 0.4846 | 0.5175 | 0.4828 | 0.5303 | | gpt-5.6-terra [9] | 0.5317 | 0.5162 | 0.5244 | 0.5132 | 0.5346 | 0.5073 | 0.5502 | ![MoverScore distance to the source encoder x decoder](MATRIX_moverscore.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | 0.5347 | 0.1771 | 0.5347 | 0.0143 | 17,393 | | Ornith-1.5-35B-A3B-NVFP4 [1] | 0.5174 | 0.1872 | 0.5174 | 0.0204 | 17,573 | | Llama-3.3-70B-Instruct-NVFP4 [2] | 0.5287 | 0.1862 | 0.5287 | 0.0172 | 17,805 | | ❗ Qwen3.8-27B-AWQ-INT4 [3] | 0.5158 | 0.1901 | 0.5158 | 0.0200 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | 0.5449 | 0.1748 | 0.5449 | 0.0142 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | 0.5193 | 0.1873 | 0.5193 | 0.0185 | 17,630 | | ✔️ Laguna-S-2.1-NVFP4 [6] | 0.5543 | 0.1759 | 0.5543 | 0.0120 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | 0.6253 | 0.0766 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | 0.5046 | 0.1561 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | 0.6719 | 0.0537 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | 0.6267 | 0.0789 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | 0.6659 | 0.1300 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | 0.6826 | 0.0622 | 5,029 | | Reformat this text into a sensible, structured JSON object. | 0.5547 | 0.1215 | 5,192 | | ❗ Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | 0.2354 | 0.1318 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | 0.5046 | 0.1571 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | 0.2773 | 0.1537 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | 0.5809 | 0.1105 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | 0.6546 | 0.0643 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 0.7642 | 0.0852 | 4,787 | | Translate the given text to French. | 0.3731 | 0.0635 | 7,849 | | Translate the given text to German. | 0.3737 | 0.0680 | 7,866 | | Translate the given text to Hindi. | 0.4237 | 0.0813 | 7,843 | | Translate the given text to Spanish. | 0.3472 | 0.0691 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | 0.6847 | 0.1246 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | 0.6491 | 0.0708 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 0.7177 | 0.0410 | 5,064 | ### Reranker distance to the source (`reranker`) Per encoder, marginalised over every decoder. `Pooled Mean` weights every kept row equally; `Decoder Balanced Mean` is the mean of the per-decoder means, and `Spread Across Decoders` their standard deviation. | Encoder | Pooled Mean | Pooled Std | Decoder Balanced Mean | Spread Across Decoders | N | Decoders | |---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | -5.1087 | 4.6634 | -5.1181 | 1.2963 | 13,601 | 7 | | Ornith-1.5-35B-A3B-NVFP4 [1] | -5.1772 | 4.8973 | -5.1840 | 1.2832 | 13,615 | 7 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | -3.3532 | 5.9677 | -3.3607 | 1.4214 | 13,723 | 7 | | Qwen3.8-27B-AWQ-INT4 [3] | -5.2728 | 4.5851 | -5.2792 | 1.2749 | 13,639 | 7 | | Mistral-Small-4-119B-2603-NVFP4 [4] | -4.6685 | 5.1148 | -4.6743 | 1.2902 | 13,731 | 7 | | gemma-4-31B-it-AWQ-4bit [5] | -4.9193 | 4.8794 | -4.9291 | 1.3546 | 13,509 | 7 | | Laguna-S-2.1-NVFP4 [6] | -4.5815 | 5.0423 | -4.5894 | 1.3084 | 13,662 | 7 | | ❗ claude-sonnet-5 [8] | -5.8345 | 4.2569 | -5.8408 | 1.2062 | 13,732 | 7 | | gpt-5.6-terra [9] | -5.7665 | 4.3953 | -5.7724 | 1.3443 | 13,721 | 7 | ![Reranker distance to the source per encoder](ENC_reranker.png) Encoder x decoder cell means: | Encoder \ Decoder | granite-4.2-30b-nvfp4 [0] | Ornith-1.5-35B-A3B-NVFP4 [1] | Llama-3.3-70B-Instruct-NVFP4 [2] | Qwen3.8-27B-AWQ-INT4 [3] | Mistral-Small-4-119B-2603-NVFP4 [4] | gemma-4-31B-it-AWQ-4bit [5] | Laguna-S-2.1-NVFP4 [6] | |---|---|---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | -5.7992 | -6.0496 | -1.9920 | -5.5104 | -5.7292 | -5.3601 | -5.3862 | | Ornith-1.5-35B-A3B-NVFP4 [1] | -6.0300 | -6.0421 | -2.1075 | -5.7409 | -5.6393 | -5.4786 | -5.2499 | | Llama-3.3-70B-Instruct-NVFP4 [2] | -4.2751 | -4.3410 | 0.0502 | -3.8787 | -3.9640 | -3.5304 | -3.5856 | | Qwen3.8-27B-AWQ-INT4 [3] | -6.0635 | -6.1672 | -2.2127 | -5.7549 | -5.7552 | -5.5994 | -5.4017 | | Mistral-Small-4-119B-2603-NVFP4 [4] | -5.5409 | -5.5674 | -1.5758 | -5.1226 | -5.1078 | -4.9699 | -4.8356 | | gemma-4-31B-it-AWQ-4bit [5] | -5.5996 | -5.9167 | -1.6619 | -5.5304 | -5.4200 | -5.2505 | -5.1244 | | Laguna-S-2.1-NVFP4 [6] | -5.3840 | -5.5235 | -1.4411 | -5.0479 | -5.0816 | -4.8792 | -4.7685 | | claude-sonnet-5 [8] | -6.3986 | -6.8100 | -2.9369 | -6.2911 | -6.2062 | -6.1842 | -6.0587 | | gpt-5.6-terra [9] | -6.4315 | -6.6533 | -2.5106 | -6.3451 | -6.1760 | -6.2698 | -6.0205 | ![Reranker distance to the source encoder x decoder](MATRIX_reranker.png) Per decoder, marginalised over every encoder (for contrast): | Decoder | Pooled Mean | Pooled Std | Encoder Balanced Mean | Spread Across Encoders | N | |---|---|---|---|---|---| | granite-4.2-30b-nvfp4 [0] | -5.7255 | 4.2002 | -5.7247 | 0.6177 | 17,393 | | ❗ Ornith-1.5-35B-A3B-NVFP4 [1] | -5.8962 | 4.1339 | -5.8968 | 0.6825 | 17,573 | | ✔️ Llama-3.3-70B-Instruct-NVFP4 [2] | -1.8208 | 6.4574 | -1.8209 | 0.7970 | 17,805 | | Qwen3.8-27B-AWQ-INT4 [3] | -5.4675 | 4.4800 | -5.4691 | 0.7017 | 17,530 | | Mistral-Small-4-119B-2603-NVFP4 [4] | -5.4538 | 4.5753 | -5.4533 | 0.6470 | 17,591 | | gemma-4-31B-it-AWQ-4bit [5] | -5.2793 | 4.4861 | -5.2802 | 0.7648 | 17,630 | | Laguna-S-2.1-NVFP4 [6] | -5.1590 | 4.5684 | -5.1590 | 0.6996 | 17,411 | Per encoding instruction, marginalised over every encoder and decoder: | Encoding instruction | Mean | Std | N | |---|---|---|---| | Convert this text into a series of logical propositions or syllogisms that represent the c... | -3.9055 | 3.9910 | 6,283 | | Create a detailed dictionary that fully describes all the meaningful phrases and words thi... | -5.1357 | 3.7171 | 5,227 | | Create a prompt that might cause an LLM to generate an output resembling this text. | -3.5881 | 4.1917 | 6,167 | | Create a prompt that would allow an LLM to generate this text as accurately/verbatim as po... | -4.9329 | 3.7351 | 6,148 | | Create the shortest possible prompt that can be sent to an LLM to reproduce this text. The... | -4.1475 | 4.6113 | 5,806 | | Envision the scenario in which the author wrote this text. Describe that scenario in exhau... | -2.4133 | 4.8027 | 5,029 | | Reformat this text into a sensible, structured JSON object. | -5.6481 | 3.4504 | 5,192 | | Replace every verb in this text with the tag [VERB], keeping all nouns, adjectives, articl... | -8.0528 | 1.7414 | 6,073 | | Summarize this text, attempting to preserve as much of the original language of the text a... | -6.1211 | 3.0805 | 6,196 | | Take this text and replace one word in every four with a single underscore (use one unders... | -7.4950 | 2.2661 | 6,096 | | Take this text, but extract only the most meaningful sentences out of it to create a new t... | -5.0779 | 3.7985 | 6,261 | | This piece of text was cleverely generated by an LLM with a human supervising it so that t... | -4.1690 | 3.9916 | 6,091 | | ✔️ Translate the entirety of this text into a sequence of emojis that captures the literal me... | 6.9514 | 6.1475 | 4,787 | | Translate the given text to French. | -8.3883 | 1.3209 | 7,849 | | Translate the given text to German. | -8.2925 | 1.5709 | 7,866 | | Translate the given text to Hindi. | -7.8838 | 1.6850 | 7,843 | | ❗ Translate the given text to Spanish. | -8.4516 | 1.0648 | 7,860 | | Write a 'few-shot' prompt using minimal, fragmented excerpts from this text as examples. T... | -3.2339 | 4.5885 | 5,914 | | Write a detailed descriptor for how this text is stylistically differentiated from other t... | -2.3529 | 4.5233 | 5,181 | | Write a detailed descriptor for the (imaginary) personality of the author who wrote this t... | 1.2162 | 4.9882 | 5,064 |