Update README and 24 associated file(s)

README.md

@@ -613,6 +613,38 @@ 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)
 
+## 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.
@@ -622,6 +654,7 @@ EditLens used as a detector. The threshold is calibrated on the whole human pool
 | 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 |
@@ -685,6 +718,38 @@ TPR at 0.1% FPR per encoder and encoding family:
 | 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 |