Update README and 24 associated file(s)

README.md

@@ -613,6 +613,103 @@ Decoder post-processing:
 - 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)
 
+## 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 |
+| 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 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).