Publish Draftmark browser v3 ONNX release and GPU handoff

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Draftmark contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,44 @@
+---
+license: mit
+language:
+- en
+pipeline_tag: text-classification
+base_model: Oxidane/tmr-ai-text-detector
+base_model_relation: finetune
+tags:
+- onnx
+- roberta
+- ai-text-detection
+- onnxruntime
+- wasm
+- browser
+---
+
+# Draftmark RoBERTa FT INT8
+
+Draftmark is a free, on-device writing-pattern detector for English prose. This is the exact weight-only INT8 ONNX export used by the live browser release. Its model and tokenizer assets total **131.1 MB**. [Live Draftmark](https://draftmark-virid.vercel.app/) · [Release manifest](release-manifest.json)
+
+## Inference and intended use
+
+The ONNX graph at `onnx/model.int8.onnx` accepts `input_ids` and `attention_mask` as rank-two INT64 tensors. Its two logits are ordered `human`, then `ai`. Apply a softmax per window. The shipped worker scores sentence-aligned windows of up to 512 tokens including special tokens and averages the AI-class scores weighted by content-token count. Oversized sentences are split across windows. The shipped flag threshold is **0.999146566314596** and the limited-signal cutoff is **0.25**.
+
+The evaluated app accepts English prose of 150–2,000 words and up to 20,000 characters. Scores can support manual review. Authorship decisions require independent evidence. Results depend on the generator, rewriting method, domain, and writing population.
+
+## Fresh holdout results
+
+| Measure | Previous Ettin | Draftmark FT INT8 |
+| --- | ---: | ---: |
+| Unedited AI recall | 215/300 | 277/300 |
+| Decoder rewrite recall | 18/150 | 89/150 |
+| Sequence-to-sequence rewrite recall | 132/150 | 110/150 |
+| Mean paraphrase recall | 50.0% | 66.3% |
+| General-human false flags | 10/300 | 1/300 |
+| Learner-essay false flags | 80/220 | 0/220 |
+
+The paired 95% confidence interval for the mean paraphrase gain was **+10.7 to +22.0 percentage points**. Sequence-to-sequence recall declined. The learner false-flag count applies to the 220 ASAP 2.0 essays evaluated. The results are a comparison against Draftmark’s previous detector on these samples. [Aggregate benchmark](finetune-v3.json)
+
+## Training and attribution
+
+Draftmark fine-tuned the last four encoder layers and the classification head of Oxidane’s MIT-licensed TMR RoBERTa at revision `0ceddea903015ef99cbaa040a4d8a216aed9c683`. Training used source-grouped DetectRL-X pairs, original AI and human prose from DetectionAI, and learner essays from the ASAP 2.0 corpus and the Shijaku–Canhasi TOEFL collection. The report records the source pins, split sizes, and overlap screening. The ASAP 2.0 corpus is CC BY 4.0 and is credited to Scott Crossley and contributors, Vanderbilt University, and The Learning Agency Lab. [TMR](https://huggingface.co/Oxidane/tmr-ai-text-detector) · [DetectRL-X](https://aclanthology.org/2026.acl-long.1773/) · [DetectionAI](https://github.com/brianjabarian/DetectionAI) · [ASAP 2.0](https://github.com/scrosseye/ASAP_2.0) · [Shijaku–Canhasi corpus](https://github.com/rexshijaku/chatgpt-generated-text-detection-corpus)
+
+The [GPU handoff](handoff.md) documents how to reconstruct a Transformers checkpoint from the pinned base and the [replacement tensors on the training branch](https://huggingface.co/saksornr/draftmark-roberta-int8/blob/training/training/trainable.safetensors). The main branch is the browser distribution.

config.json

@@ -0,0 +1,34 @@
+{
+  "architectures": [
+    "RobertaForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "dtype": "float32",
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "human",
+    "1": "ai"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "label2id": {
+    "ai": 1,
+    "human": 0
+  },
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 514,
+  "model_type": "roberta",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "transformers_version": "4.57.3",
+  "type_vocab_size": 1,
+  "use_cache": true,
+  "vocab_size": 50265
+}

finetune-v3.json

@@ -0,0 +1,823 @@
+{
+  "baseline": {
+    "model": "ettin",
+    "version": "browser-v1",
+    "upper": 0.9998051638246666,
+    "lower": 0.25,
+    "experimental": false
+  },
+  "selected": "finetuned/full",
+  "policy": {
+    "model": "finetuned",
+    "scoring": "full",
+    "version": "browser-v3",
+    "experimental": false,
+    "upper": 0.999146566314596,
+    "lower": 0.25
+  },
+  "limits": {
+    "general_fpr": 0.02,
+    "learner_fpr": 0.05,
+    "minimum_paraphrase_gain": 0.1,
+    "maximum_clean_recall_loss": 0.05,
+    "maximum_model_bytes": 160000000,
+    "maximum_warm_p95_ms": 5000
+  },
+  "manifest": {
+    "heldout_sha256": "c305977e14a03b9ae6c07ac35ac95ac4afc09f9fd1f8df2f9348f9d69afb7f0e",
+    "development_sha256": "a81c7b90b5d1dd5a1ebc3cb4711ada93858f22f31604db3e3f408e5a46ee8878",
+    "data_sha256": "cd200cc87a23ba552e6182d9479a0bdfaa9826e80022ab4d09cd4e4b571485ee",
+    "model_manifest_sha256": "330d445dd29b9f81995cdb20bf935d28b518cca4a850a1a807f42c6601b0c527",
+    "model_sha256": "9104295a0903256724f44b86523ad32da0b800efc1ff6a3fc3d2f8e2d2339418",
+    "worker_sha256": "e6eba6365b39e878221f4d8e6b274b5ffa4092c6002a253c235a3240a21a434e",
+    "wasm_sha256": "4b884c203b0fa48d057be3332c65513ff85aa3c5bddd6946acc0fab7ecda001d",
+    "decision_sha256": "1416ece45b413cb5dc8616e634baa0bef7ae0d790186d5df6a155454b064a3a5",
+    "performance_sha256": "24de2e1f9cbffcde48612ad248e1f5716e94abc39f493b8ca6c4f999e544a4a7"
+  },
+  "calibration": {
+    "live": {
+      "cohorts": {
+        "asap_learner": {
+          "flags": {
+            "hits": 41,
+            "n": 101,
+            "rate": 0.40594059405940597,
+            "ci95": [
+              0.31532231224182006,
+              0.5034519147825671
+            ]
+          },
+          "abstention": {
+            "hits": 39,
+            "n": 101,
+            "rate": 0.38613861386138615,
+            "ci95": [
+              0.2970216442209859,
+              0.48359978849274593
+            ]
+          },
+          "outside_input_limits": 0
+        },
+        "clean_ai": {
+          "flags": {
+            "hits": 152,
+            "n": 220,
+            "rate": 0.6909090909090909,
+            "ci95": [
+              0.6270045309504402,
+              0.7482608352728177
+            ]
+          },
+          "abstention": {
+            "hits": 51,
+            "n": 220,
+            "rate": 0.2318181818181818,
+            "ci95": [
+              0.18094651434879364,
+              0.29189499500424887
+            ]
+          },
+          "outside_input_limits": 0
+        },
+        "decoder": {
+          "flags": {
+            "hits": 6,
+            "n": 110,
+            "rate": 0.05454545454545454,
+            "ci95": [
+              0.025235483062071407,
+              0.11391927231733298
+            ]
+          },
+          "abstention": {
+            "hits": 58,
+            "n": 110,
+            "rate": 0.5272727272727272,
+            "ci95": [
+              0.4346352821640567,
+              0.6180695287310817
+            ]
+          },
+          "outside_input_limits": 1
+        },
+        "general_human": {
+          "flags": {
+            "hits": 4,
+            "n": 220,
+            "rate": 0.01818181818181818,
+            "ci95": [
+              0.007092586915565129,
+              0.04580910785430784
+            ]
+          },
+          "abstention": {
+            "hits": 107,
+            "n": 220,
+            "rate": 0.4863636363636364,
+            "ci95": [
+              0.42111944404533,
+              0.552075886938723
+            ]
+          },
+          "outside_input_limits": 1
+        },
+        "seq2seq": {
+          "flags": {
+            "hits": 102,
+            "n": 110,
+            "rate": 0.9272727272727272,
+            "ci95": [
+              0.8630186778224797,
+              0.962690026201357
+            ]
+          },
+          "abstention": {
+            "hits": 8,
+            "n": 110,
+            "rate": 0.07272727272727272,
+            "ci95": [
+              0.03730997379864303,
+              0.13698132217752032
+            ]
+          },
+          "outside_input_limits": 0
+        }
+      },
+      "paraphrase_discrimination": {
+        "decoder": {
+          "auroc": 0.541996564869507,
+          "recall_at_1pct_fpr": 0.01834862385321101,
+          "recall_at_5pct_fpr": 0.12844036697247707
+        },
+        "seq2seq": {
+          "auroc": 0.986384391863844,
+          "recall_at_1pct_fpr": 0.9,
+          "recall_at_5pct_fpr": 0.9454545454545454
+        }
+      }
+    },
+    "finetuned/full": {
+      "cohorts": {
+        "asap_learner": {
+          "flags": {
+            "hits": 0,
+            "n": 101,
+            "rate": 0.0,
+            "ci95": [
+              0,
+              0.03664194365595336
+            ]
+          },
+          "abstention": {
+            "hits": 1,
+            "n": 101,
+            "rate": 0.009900990099009901,
+            "ci95": [
+              0.001749866201761846,
+              0.05396847460951917
+            ]
+          },
+          "outside_input_limits": 0
+        },
+        "clean_ai": {
+          "flags": {
+            "hits": 204,
+            "n": 220,
+            "rate": 0.9272727272727272,
+            "ci95": [
+              0.8851382050617915,
+              0.9547414241045473
+            ]
+          },
+          "abstention": {
+            "hits": 16,
+            "n": 220,
+            "rate": 0.07272727272727272,
+            "ci95": [
+              0.04525857589545253,
+              0.1148617949382084
+            ]
+          },
+          "outside_input_limits": 0
+        },
+        "decoder": {
+          "flags": {
+            "hits": 57,
+            "n": 110,
+            "rate": 0.5181818181818182,
+            "ci95": [
+              0.4257777238231319,
+              0.6093588167736271
+            ]
+          },
+          "abstention": {
+            "hits": 49,
+            "n": 110,
+            "rate": 0.44545454545454544,
+            "ci95": [
+              0.3559755711590631,
+              0.53861480705066
+            ]
+          },
+          "outside_input_limits": 1
+        },
+        "general_human": {
+          "flags": {
+            "hits": 2,
+            "n": 220,
+            "rate": 0.00909090909090909,
+            "ci95": [
+              0.00249655343526377,
+              0.03253536199064455
+            ]
+          },
+          "abstention": {
+            "hits": 23,
+            "n": 220,
+            "rate": 0.10454545454545454,
+            "ci95": [
+              0.0706788038882292,
+              0.15198579464930803
+            ]
+          },
+          "outside_input_limits": 1
+        },
+        "seq2seq": {
+          "flags": {
+            "hits": 78,
+            "n": 110,
+            "rate": 0.7090909090909091,
+            "ci95": [
+              0.6183048036071215,
+              0.7857654132556071
+            ]
+          },
+          "abstention": {
+            "hits": 32,
+            "n": 110,
+            "rate": 0.2909090909090909,
+            "ci95": [
+              0.214234586744393,
+              0.3816951963928785
+            ]
+          },
+          "outside_input_limits": 0
+        }
+      },
+      "paraphrase_discrimination": {
+        "decoder": {
+          "auroc": 0.9780486783125969,
+          "recall_at_1pct_fpr": 0.5229357798165137,
+          "recall_at_5pct_fpr": 0.8990825688073395
+        },
+        "seq2seq": {
+          "auroc": 0.9839352428393525,
+          "recall_at_1pct_fpr": 0.7090909090909091,
+          "recall_at_5pct_fpr": 0.8272727272727273
+        }
+      }
+    }
+  },
+  "release": true,
+  "gates": {
+    "general_fpr": true,
+    "learner_fpr": true,
+    "paraphrase_gain": true,
+    "clean_ai_recall": true,
+    "native_cpu_parity": true,
+    "model_size": true,
+    "warm_latency": true,
+    "general_human_support": true,
+    "asap_learner_support": true,
+    "decoder_support": true,
+    "seq2seq_support": true,
+    "clean_ai_support": true
+  },
+  "model": {
+    "repo": "Oxidane/tmr-ai-text-detector",
+    "revision": "draftmark-ft-9104295a09032567",
+    "baseRevision": "0ceddea903015ef99cbaa040a4d8a216aed9c683",
+    "name": "Draftmark RoBERTa FT INT8",
+    "license": "MIT",
+    "licenseUrl": "https://huggingface.co/Oxidane/tmr-ai-text-detector#license",
+    "limit": 512,
+    "tokenizerType": "roberta",
+    "source": "static",
+    "quantization": "weights",
+    "files": {
+      "model": [
+        "model.int8.onnx",
+        127550037,
+        "9104295a0903256724f44b86523ad32da0b800efc1ff6a3fc3d2f8e2d2339418"
+      ],
+      "tokenizer": [
+        "tokenizer.json",
+        3558741,
+        "1f33749d010b4d63908e5c174c341622cb45039dd73a139dcd95bd74cc7e304b"
+      ],
+      "config": [
+        "tokenizer_config.json",
+        1354,
+        "288b4077af1ffb3beead6d96fccfc93beb2df9b689cbb038c4eb329165efc43a"
+      ]
+    }
+  },
+  "model_bytes": 131110132,
+  "training": {
+    "base": "Oxidane/tmr-ai-text-detector@0ceddea903015ef99cbaa040a4d8a216aed9c683",
+    "device": "mps",
+    "torch": "2.8.0",
+    "seed": 36513,
+    "max_training_tokens": 384,
+    "trainable_parameters": 28943618,
+    "top_encoder_layers": 4,
+    "batch_size": 16,
+    "learning_rate": 2e-05,
+    "cohort_weights": {
+      "learner": 1.45,
+      "decoder": 1.3,
+      "seq2seq": 0.9,
+      "encoder": 1.0
+    },
+    "epochs": [
+      {
+        "upper": 0.975287437438965,
+        "rates": {
+          "asap_learner": 0.009900990099009901,
+          "seq2seq": 0.8636363636363636,
+          "clean_ai": 0.9590909090909091,
+          "general_human": 0.00909090909090909,
+          "decoder": 0.4090909090909091
+        },
+        "selection": 0.9720454545454545,
+        "epoch": 1,
+        "train_loss": 0.33598193104093854
+      },
+      {
+        "upper": 0.9825209975242616,
+        "rates": {
+          "asap_learner": 0.009900990099009901,
+          "seq2seq": 0.8818181818181818,
+          "clean_ai": 0.9363636363636364,
+          "general_human": 0.00909090909090909,
+          "decoder": 0.6818181818181818
+        },
+        "selection": 1.1095454545454544,
+        "epoch": 2,
+        "train_loss": 0.1008723359008851
+      },
+      {
+        "upper": 0.9987732768058778,
+        "rates": {
+          "asap_learner": 0.009900990099009901,
+          "seq2seq": 0.9727272727272728,
+          "clean_ai": 0.9772727272727273,
+          "general_human": 0.00909090909090909,
+          "decoder": 0.6272727272727273
+        },
+        "selection": 1.1420454545454546,
+        "epoch": 3,
+        "train_loss": 0.061243358406307756
+      }
+    ],
+    "selected_epoch": 3,
+    "checkpoint_sha256": "4c10be4f8da2d2f75ee14c2b1d8fb143f549edec478e0d9121877a0977fa1ea4"
+  },
+  "data": {
+    "sources": {
+      "DetectRL-X": "https://huggingface.co/datasets/WUJUNCHAO/DetectRL-X/tree/ccc9a0a74793dd3448e0da1650b430f5df1b0402",
+      "ASAP 2.0": "https://github.com/scrosseye/ASAP_2.0/tree/060ab225ca0984c0bf43ff222bf6bb2bc6195b6f",
+      "DetectionAI": "https://github.com/brianjabarian/DetectionAI/tree/5610637e3df52e107e642ed151e7f842b15f79d2",
+      "MAGE": "https://huggingface.co/datasets/yaful/MAGE/tree/342663f0a2b775455c023f5d36a1341ff0ec5402",
+      "APT-Eval": "https://github.com/ShoumikSaha/ai-polished-text/tree/53b1d2619a633e025998d6c7d0a12b0228b589da",
+      "Heterogeneous AI Spans": "https://huggingface.co/datasets/open-text-detector/heterogeneous-ai-spans/tree/c8a59b129c34cfd9c67f739ae049abfff26fc2d3",
+      "Ettin public training subset": "https://huggingface.co/datasets/nakas/gpttrace-text-train/tree/3f31325869bfc5414c1e8c713a6a4fd75b3eebb7",
+      "Shijaku\u2013Canhasi TOEFL": "https://github.com/rexshijaku/chatgpt-generated-text-detection-corpus/tree/c0eb34cfd5c96cc02f7109a561eb753128cae9e6",
+      "RAID train_none.csv": "sha256:c5467bca6fc7f5c728c676450c7f84ce401df6c6ccc6d82c47e3b5f3c6d6fce4"
+    },
+    "asap_train_sha256": "aec01e98eb16cc3bc3828010ca053a09daecc2821dd2105f8a7333089271beb4",
+    "counts": {
+      "train": {
+        "decoder": 680,
+        "clean_ai": 1755,
+        "general_human": 1754,
+        "seq2seq": 430,
+        "encoder": 420,
+        "learner": 1300
+      },
+      "calibration": {
+        "decoder": 110,
+        "clean_ai": 220,
+        "general_human": 220,
+        "seq2seq": 110,
+        "asap_learner": 101
+      },
+      "holdout": {
+        "decoder": 150,
+        "clean_ai": 300,
+        "general_human": 300,
+        "seq2seq": 150,
+        "asap_learner": 220
+      }
+    },
+    "sha256": {
+      "train": "bb4793d1a71c8c9470196eaf19810605e01433e7854fc20d678ee6ad3a035707",
+      "calibration": "a81c7b90b5d1dd5a1ebc3cb4711ada93858f22f31604db3e3f408e5a46ee8878",
+      "holdout": "c305977e14a03b9ae6c07ac35ac95ac4afc09f9fd1f8df2f9348f9d69afb7f0e"
+    },
+    "groups_excluded_for_overlap": 272,
+    "groups_excluded_for_pretrained_overlap": 265,
+    "training_overlap_limit": "Screened against the previous benchmark, the new training/calibration data, RAID train without attacks, MAGE train, and the public GPTTrace training subset. Unpublished training rows are unverifiable."
+  },
+  "performance": {
+    "browser": "Chromium 131.0.6778.205",
+    "device": "darwin 25.6.0 Apple M4 Pro",
+    "engine": "ONNX Runtime Web WASM CPU, one thread per worker",
+    "worker_sha256": "e6eba6365b39e878221f4d8e6b274b5ffa4092c6002a253c235a3240a21a434e",
+    "parity": true,
+    "label_order": [
+      "human",
+      "ai"
+    ],
+    "cases": [
+      {
+        "fixture": "example",
+        "expected": 0.9987629118304412,
+        "score": 0.9987629265593501,
+        "delta": 1.4728908870331736e-08,
+        "token_ids_match": true,
+        "pass_token_counts": [
+          204
+        ],
+        "chunks": 1,
+        "tokens": 204,
+        "expected_chunks": 1,
+        "expected_tokens": 204
+      },
+      {
+        "fixture": "300-word",
+        "expected": 0.9992861080622527,
+        "score": 0.9992861089126702,
+        "delta": 8.504175141865744e-10,
+        "token_ids_match": true,
+        "pass_token_counts": [
+          336
+        ],
+        "chunks": 1,
+        "tokens": 336,
+        "expected_chunks": 1,
+        "expected_tokens": 336
+      },
+      {
+        "fixture": "multi-window",
+        "expected": 0.9687787378573235,
+        "score": 0.9687795552453534,
+        "delta": 8.173880299278125e-07,
+        "token_ids_match": true,
+        "pass_token_counts": [
+          643
+        ],
+        "chunks": 2,
+        "tokens": 643,
+        "expected_chunks": 2,
+        "expected_tokens": 643
+      },
+      {
+        "fixture": "oversized-sentence",
+        "expected": 0.006943256878519529,
+        "score": 0.0069432706767937865,
+        "delta": 1.3798274257556531e-08,
+        "token_ids_match": true,
+        "pass_token_counts": [
+          1281
+        ],
+        "chunks": 3,
+        "tokens": 1281,
+        "expected_chunks": 3,
+        "expected_tokens": 1281
+      },
+      {
+        "fixture": "unicode",
+        "expected": 0.04916038580631243,
+        "score": 0.04916000132066757,
+        "delta": 3.8448564485676906e-07,
+        "token_ids_match": true,
+        "pass_token_counts": [
+          246
+        ],
+        "chunks": 1,
+        "tokens": 246,
+        "expected_chunks": 1,
+        "expected_tokens": 246
+      }
+    ],
+    "warm_p50_ms": 1247,
+    "warm_p95_ms": 1262,
+    "warm_samples": 20,
+    "reference_words": 300,
+    "first_scan_ms": 1364,
+    "first_scan_download_observed": true
+  },
+  "holdout_runtime": {
+    "browser": "Chromium 131.0.6778.205",
+    "device": "darwin 25.6.0 Apple M4 Pro",
+    "engine": "ONNX Runtime Web WASM CPU, one thread per worker",
+    "worker_sha256": "e6eba6365b39e878221f4d8e6b274b5ffa4092c6002a253c235a3240a21a434e",
+    "signature": "80da49b4ca2df09f57c5d3efc50407d6e8ce7f1ef5a91250ee5130035dc4056a",
+    "specs": {
+      "live": {
+        "model": "ettin",
+        "version": "browser-v1",
+        "upper": 0.9998051638246666,
+        "lower": 0.25,
+        "experimental": false
+      },
+      "finetuned/full": {
+        "model": "finetuned",
+        "scoring": "full",
+        "version": "browser-v3",
+        "experimental": false,
+        "upper": 0.999146566314596,
+        "lower": 0.25
+      }
+    }
+  },
+  "holdout": {
+    "live": {
+      "cohorts": {
+        "asap_learner": {
+          "flags": {
+            "hits": 80,
+            "n": 220,
+            "rate": 0.36363636363636365,
+            "ci95": [
+              0.3029141434403334,
+              0.42903916640019657
+            ]
+          },
+          "abstention": {
+            "hits": 98,
+            "n": 220,
+            "rate": 0.44545454545454544,
+            "ci95": [
+              0.3812726964449632,
+              0.5115086274912487
+            ]
+          },
+          "outside_input_limits": 0
+        },
+        "clean_ai": {
+          "flags": {
+            "hits": 215,
+            "n": 300,
+            "rate": 0.7166666666666667,
+            "ci95": [
+              0.6631845423956875,
+              0.7646699712719605
+            ]
+          },
+          "abstention": {
+            "hits": 63,
+            "n": 300,
+            "rate": 0.21,
+            "ci95": [
+              0.16772109488365303,
+              0.25961209451503364
+            ]
+          },
+          "outside_input_limits": 0
+        },
+        "decoder": {
+          "flags": {
+            "hits": 18,
+            "n": 150,
+            "rate": 0.12,
+            "ci95": [
+              0.0772684043967938,
+              0.18170966135395242
+            ]
+          },
+          "abstention": {
+            "hits": 71,
+            "n": 150,
+            "rate": 0.47333333333333333,
+            "ci95": [
+              0.39509757671394596,
+              0.5529008840404923
+            ]
+          },
+          "outside_input_limits": 1
+        },
+        "general_human": {
+          "flags": {
+            "hits": 10,
+            "n": 300,
+            "rate": 0.03333333333333333,
+            "ci95": [
+              0.018204749873650852,
+              0.0602624514575691
+            ]
+          },
+          "abstention": {
+            "hits": 150,
+            "n": 300,
+            "rate": 0.5,
+            "ci95": [
+              0.44377849611195097,
+              0.556221503888049
+            ]
+          },
+          "outside_input_limits": 1
+        },
+        "seq2seq": {
+          "flags": {
+            "hits": 132,
+            "n": 150,
+            "rate": 0.88,
+            "ci95": [
+              0.8182903386460476,
+              0.9227315956032063
+            ]
+          },
+          "abstention": {
+            "hits": 18,
+            "n": 150,
+            "rate": 0.12,
+            "ci95": [
+              0.0772684043967938,
+              0.18170966135395242
+            ]
+          },
+          "outside_input_limits": 1
+        }
+      },
+      "paraphrase_discrimination": {
+        "decoder": {
+          "auroc": 0.553657605889879,
+          "recall_at_1pct_fpr": 0.03355704697986577,
+          "recall_at_5pct_fpr": 0.12751677852348994
+        },
+        "seq2seq": {
+          "auroc": 0.9725258692285246,
+          "recall_at_1pct_fpr": 0.825503355704698,
+          "recall_at_5pct_fpr": 0.8859060402684564
+        }
+      }
+    },
+    "finetuned/full": {
+      "cohorts": {
+        "asap_learner": {
+          "flags": {
+            "hits": 0,
+            "n": 220,
+            "rate": 0.0,
+            "ci95": [
+              0,
+              0.017162136081943655
+            ]
+          },
+          "abstention": {
+            "hits": 3,
+            "n": 220,
+            "rate": 0.013636363636363636,
+            "ci95": [
+              0.004648152891035449,
+              0.039318652206855195
+            ]
+          },
+          "outside_input_limits": 0
+        },
+        "clean_ai": {
+          "flags": {
+            "hits": 277,
+            "n": 300,
+            "rate": 0.9233333333333333,
+            "ci95": [
+              0.8875891334382622,
+              0.9483727624969882
+            ]
+          },
+          "abstention": {
+            "hits": 23,
+            "n": 300,
+            "rate": 0.07666666666666666,
+            "ci95": [
+              0.05162723750301175,
+              0.11241086656173778
+            ]
+          },
+          "outside_input_limits": 0
+        },
+        "decoder": {
+          "flags": {
+            "hits": 89,
+            "n": 150,
+            "rate": 0.5933333333333334,
+            "ci95": [
+              0.51334517440468,
+              0.6686602129547861
+            ]
+          },
+          "abstention": {
+            "hits": 59,
+            "n": 150,
+            "rate": 0.3933333333333333,
+            "ci95": [
+              0.3187585669410804,
+              0.47323527607667293
+            ]
+          },
+          "outside_input_limits": 1
+        },
+        "general_human": {
+          "flags": {
+            "hits": 1,
+            "n": 300,
+            "rate": 0.0033333333333333335,
+            "ci95": [
+              0.0005886425761407379,
+              0.018637164554943353
+            ]
+          },
+          "abstention": {
+            "hits": 39,
+            "n": 300,
+            "rate": 0.13,
+            "ci95": [
+              0.09657483743516235,
+              0.17278130076316203
+            ]
+          },
+          "outside_input_limits": 1
+        },
+        "seq2seq": {
+          "flags": {
+            "hits": 110,
+            "n": 150,
+            "rate": 0.7333333333333333,
+            "ci95": [
+              0.6573840257093969,
+              0.7976294426892675
+            ]
+          },
+          "abstention": {
+            "hits": 39,
+            "n": 150,
+            "rate": 0.26,
+            "ci95": [
+              0.19642034813317322,
+              0.33556579865677183
+            ]
+          },
+          "outside_input_limits": 1
+        }
+      },
+      "paraphrase_discrimination": {
+        "decoder": {
+          "auroc": 0.9850957329801799,
+          "recall_at_1pct_fpr": 0.7449664429530202,
+          "recall_at_5pct_fpr": 0.9261744966442953
+        },
+        "seq2seq": {
+          "auroc": 0.9826715449709322,
+          "recall_at_1pct_fpr": 0.7651006711409396,
+          "recall_at_5pct_fpr": 0.8523489932885906
+        }
+      }
+    }
+  },
+  "paraphrase_gain": {
+    "delta": 0.16333333333333333,
+    "by_method": {
+      "decoder": 0.47333333333333333,
+      "seq2seq": -0.14666666666666667
+    },
+    "ci95": [
+      0.10662015139065152,
+      0.22040486624644678
+    ],
+    "bootstrap_source_groups": 300,
+    "bootstrap_samples": 10000
+  },
+  "clean_ai": {
+    "live": {
+      "hits": 215,
+      "n": 300,
+      "rate": 0.7166666666666667,
+      "ci95": [
+        0.6631845423956875,
+        0.7646699712719605
+      ]
+    },
+    "candidate": {
+      "hits": 277,
+      "n": 300,
+      "rate": 0.9233333333333333,
+      "ci95": [
+        0.8875891334382622,
+        0.9483727624969882
+      ]
+    },
+    "delta": 0.20666666666666667
+  }
+}

handoff.md

@@ -0,0 +1,101 @@
+# Draftmark GPU training handoff
+
+Draftmark v3 is a binary English writing-pattern detector fine-tuned from TMR RoBERTa. Production runs the weight-only INT8 ONNX artifact in a browser WASM worker. The next training run should target decoder rewrites without giving back sequence-to-sequence recall or increasing false flags on human prose and learner essays.
+
+Paths to scripts below refer to the Draftmark source workspace. Transfer the source checkout or a source-only archive to the cluster, then rebuild the benchmark cache from the pinned sources. Keep cluster credentials in the scheduler's secret store.
+
+## Artifacts and starting point
+
+| Artifact | Location |
+| --- | --- |
+| Public model release | [saksornr/draftmark-roberta-int8](https://huggingface.co/saksornr/draftmark-roberta-int8) |
+| Trainable weight delta | [training/trainable.safetensors](https://huggingface.co/saksornr/draftmark-roberta-int8/blob/training/training/trainable.safetensors) |
+| Pinned TMR base | [Oxidane/tmr-ai-text-detector at 0ceddea](https://huggingface.co/Oxidane/tmr-ai-text-detector/tree/0ceddea903015ef99cbaa040a4d8a216aed9c683) |
+| Browser export | [onnx/model.int8.onnx](https://huggingface.co/saksornr/draftmark-roberta-int8/blob/main/onnx/model.int8.onnx) |
+| Frozen evaluation | [finetune-v3.json](https://draftmark-virid.vercel.app/detector-finetune-benchmark.json) |
+| Original workspace recipe | `README.md`, section “Train and evaluate the local RoBERTa candidate” |
+
+The training branch stores the **replacement tensors** for the last four encoder layers and classification head. Load them over the pinned base, with no addition to the base weights. The checkpoint contains 28,943,618 trained parameters. Start a new optimizer and scheduler for continued training.
+
+Use Python 3.12 with the cluster's CUDA-enabled PyTorch build and the compatible dependencies in `detector/requirements-finetune.txt`. The following commands create a standard Transformers checkpoint in scratch storage.
+
+~~~sh
+export DRAFTMARK_SCRATCH="$PWD/draftmark-scratch"
+mkdir -p "$DRAFTMARK_SCRATCH"
+hf download saksornr/draftmark-roberta-int8 training/trainable.safetensors \
+  --revision training --local-dir "$DRAFTMARK_SCRATCH"
+
+python - <<'PY'
+import hashlib
+import os
+from pathlib import Path
+
+from safetensors.torch import load_file
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+
+scratch = Path(os.environ["DRAFTMARK_SCRATCH"])
+delta = scratch / "training/trainable.safetensors"
+expected_hash = "4c10be4f8da2d2f75ee14c2b1d8fb143f549edec478e0d9121877a0977fa1ea4"
+with delta.open("rb") as stream:
+    assert hashlib.file_digest(stream, "sha256").hexdigest() == expected_hash
+
+base = "Oxidane/tmr-ai-text-detector"
+revision = "0ceddea903015ef99cbaa040a4d8a216aed9c683"
+model = AutoModelForSequenceClassification.from_pretrained(base, revision=revision)
+tokenizer = AutoTokenizer.from_pretrained(base, revision=revision)
+tensors = load_file(delta)
+prefixes = tuple(f"roberta.encoder.layer.{index}." for index in range(8, 12)) + ("classifier.",)
+expected = {name for name, _ in model.named_parameters() if name.startswith(prefixes)}
+assert set(tensors) == expected
+assert sum(value.numel() for value in tensors.values()) == 28_943_618
+missing, unexpected = model.load_state_dict(tensors, strict=False)
+assert not unexpected and set(missing) == set(model.state_dict()) - expected
+assert model.config.id2label == {0: "human", 1: "ai"}
+
+destination = scratch / "warm-start"
+model.save_pretrained(destination, safe_serialization=True)
+tokenizer.save_pretrained(destination)
+print(destination)
+PY
+~~~
+
+The resulting directory is the GPU training input. The browser graph and tokenizer total 131,110,132 bytes. The exact ONNX SHA-256 is `9104295a0903256724f44b86523ad32da0b800efc1ff6a3fc3d2f8e2d2339418`.
+
+## Data and experiment plan
+
+1. Reproduce the v3 control with `detector/finetune_data.py` and `detector/finetune_model.py`. The run used 6,339 training drafts, 761 calibration drafts, and 1,120 held-out drafts. The aggregate report contains source pins, cohort counts, split hashes, and overlap exclusions.
+2. Prepare fresh source-grouped splits before looking at predictions. Keep human originals, paired AI outputs, rewrites, and near duplicates in the same split. Screen across splits and against the previous Draftmark evaluation manifests and available pretrained manifests. Treat the v3 holdout as a regression suite and reserve an untouched final test for the next release.
+3. Add decoder rewrites from new generators and methods, along with their matched clean AI drafts and human originals. Retain sequence-to-sequence examples in each training run. Add learner essays from a separately licensed source as a fairness check, ideally from a corpus distinct from ASAP 2.0. Retain learner negatives from multiple prompts and proficiency levels. Track method, generator, domain, prompt, original source, and corpus revision.
+4. In DetectRL-X, join the attack file's changed `llm_generated_text` to its clean counterpart in the general file. Preserve the separate category for human-authored drafts that AI polished. Use the author's original label for human controls and exclude ambiguous pairs. [DetectRL-X paper](https://aclanthology.org/2026.acl-long.1773/) · [ASAP 2.0 corpus](https://github.com/scrosseye/ASAP_2.0)
+
+Keep source essays and per-document predictions in access-controlled storage. Share aggregate metrics, source manifests or hashes, the release policy, and the model artifacts.
+
+## CUDA cluster implementation
+
+The checked-in trainer supports CPU and Apple MPS. Add a CUDA/DDP entry point before launching a multi-GPU job. Reuse the existing model, tokenization, cohort weights, and checkpoint selection logic so the first cluster run has a comparable control.
+
+- **Control settings** Keep encoder layers 0–7 and embeddings frozen, train layers 8–11 and the classifier, and use a 384-token training cap, global batch 16, AdamW at 2e-5, weight decay 0.01, gradient clip 1.0, and three epochs. Warm up over the first twelfth of optimizer steps and decay to 25% of the peak learning rate. The original seed was 36513. Learner, decoder, seq2seq, and encoder example weights were 1.45, 1.30, 0.90, and 1.0.
+- **Distributed execution** Bind each process to `LOCAL_RANK`, initialize NCCL, and wrap the classifier with `DistributedDataParallel`. Use one process per GPU, BF16 autocast on supported hardware, and length-bucketed rank shards with the same number of updates. Derive gradient accumulation from the global batch target. Normalize weighted loss across ranks and accumulation steps so the effective cohort weights match the control.
+- **Validation and resume** Gather raw scores with stable IDs and deduplicate padding rows from distributed sampling. The v3 selection criterion was mean paraphrase recall plus 0.35 times unedited AI recall on calibration, with false-flag targets of 1% on general prose and 3% on learner essays. Save model, optimizer, scheduler, scaler if applicable, epoch, step, RNG state, dataset hashes, and hyperparameters on rank zero. Verify that a resumed smoke run matches the expected progress before requesting a longer allocation.
+- **Launch** After implementing the CUDA entry point, start with `torchrun --standalone --nproc-per-node=4 <entrypoint> <arguments>` on one node. For multiple nodes, use one `torchrun` launcher per node with the same rendezvous endpoint and job-specific rendezvous ID. Record world size, effective batch, CUDA/PyTorch versions, GPU type, and job configuration. [PyTorch torchrun documentation](https://docs.pytorch.org/docs/stable/elastic/run.html)
+
+Run a matched warm-start experiment before changing model size or training all layers. Useful ablations are the prior cohort mix versus added decoder and learner hard negatives, then a 384- versus 512-token training window. Compare runs at frozen false-positive budgets. Watch decoder and seq2seq recall separately so their mean cannot hide a large regression.
+
+## Export and release gates
+
+1. Save the chosen Transformers checkpoint and export dynamic batch and sequence axes with ONNX opset 17. Reuse the weight-only INT8 export in `detector/finetune_model.py`. Float activations avoided the native/Chrome score drift observed with the dynamic INT8 trial. Verify the model plus tokenizer are **under 160 MB**.
+2. Preserve production semantics from `src/detector/detector.worker.ts` and `src/detector/decision.ts`. Inputs use English prose of 150–2,000 words and at most 20,000 characters. Score sentence-aligned chunks up to 512 tokens including special tokens, split oversized sentences without truncating the draft, and weight the score by content-token count. Class order is human, then AI. The current upper and lower cutoffs are 0.999146566314596 and 0.25. Fit new cutoffs on the exported WASM calibration scores.
+3. Freeze the candidate, threshold, holdout, and worker checksums before evaluating the final set. Reuse `scripts/paraphrase-browser.mjs` and `detector/finetune_eval.py`, updating their baseline loader from the archived Ettin policy to the published v3 artifact and policy. Report both paraphrase families, clean AI recall, general and learner false flags, abstentions, counts, confidence intervals, and browser latency. Count every visible AI flag as a positive and every AI abstention as a miss.
+4. Require **<2%** false flags on general human prose and **<5%** on human-written learner essays. Use the shipped v3 as the new comparison baseline. The last release required at least a **10-point** mean paraphrase recall gain with a positive lower bound on the paired 95% interval, at most a **5-point** clean-AI recall loss, native/WASM score parity, and **≤5 seconds p95** for a warm 300-word browser scan. Evaluate on both desktop and mobile before promoting a new static build.
+
+| v3 result on the 1,120-draft holdout | Measured result |
+| --- | --- |
+| Mean paraphrase recall | 66.3% vs 50.0% for Ettin |
+| Decoder rewrite recall | 89/150 vs 18/150 |
+| Sequence-to-sequence rewrite recall | 110/150 vs 132/150 |
+| Unedited AI recall | 277/300 vs 215/300 |
+| General-human false flags | 1/300 |
+| Learner false flags | 0/220 |
+| Warm 300-word scan | 1.26 s p95 on Apple M4 Pro in Chromium |
+
+These results establish performance on the listed cohorts. A state-of-the-art comparison requires evaluating the exact new artifact against other detectors on the same independent benchmark at matched false-positive rates. [RAID](https://github.com/liamdugan/raid)

onnx/model.int8.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9104295a0903256724f44b86523ad32da0b800efc1ff6a3fc3d2f8e2d2339418
+size 127550037

release-manifest.json

@@ -0,0 +1,29 @@
+{
+  "version": "browser-v3",
+  "base_model": "Oxidane/tmr-ai-text-detector",
+  "base_revision": "0ceddea903015ef99cbaa040a4d8a216aed9c683",
+  "policy": {
+    "model": "finetuned",
+    "scoring": "full",
+    "version": "browser-v3",
+    "experimental": false,
+    "upper": 0.999146566314596,
+    "lower": 0.25
+  },
+  "weight_only_int8": true,
+  "browser_bytes": 131110132,
+  "files": {
+    "onnx/model.int8.onnx": {
+      "size": 127550037,
+      "sha256": "9104295a0903256724f44b86523ad32da0b800efc1ff6a3fc3d2f8e2d2339418"
+    },
+    "tokenizer.json": {
+      "size": 3558741,
+      "sha256": "1f33749d010b4d63908e5c174c341622cb45039dd73a139dcd95bd74cc7e304b"
+    },
+    "tokenizer_config.json": {
+      "size": 1354,
+      "sha256": "288b4077af1ffb3beead6d96fccfc93beb2df9b689cbb038c4eb329165efc43a"
+    }
+  }
+}

tokenizer_config.json

@@ -0,0 +1,62 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50264": {
+      "content": "<mask>",
+      "lstrip": true,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "mask_token": "<mask>",
+  "max_length": 512,
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "stride": 0,
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "truncation_side": "right",
+  "truncation_strategy": "longest_first",
+  "unk_token": "<unk>"
+}