argus-uq 0.1.0__tar.gz

argus_uq-0.1.0/CITATION.cff

@@ -0,0 +1,38 @@
+cff-version: 1.2.0
+message: "If you use this software or data, please cite the paper below."
+title: "argus-uq"
+type: software
+version: 0.1.0
+license: MIT
+repository-code: "https://github.com/divake/argus-uq"
+authors: &authors
+  - family-names: Kumar
+    given-names: Divake
+    affiliation: University of Illinois Chicago
+  - family-names: Tayebati
+    given-names: Sina
+    affiliation: University of Illinois Chicago
+  - family-names: Naik
+    given-names: Devashri
+    affiliation: University of Illinois Chicago
+  - family-names: Rios
+    given-names: Amanda Sofie
+    affiliation: Intel Labs
+  - family-names: Ahuja
+    given-names: Nilesh
+    affiliation: Intel Labs
+  - family-names: Tickoo
+    given-names: Omesh
+    affiliation: Intel Labs
+  - family-names: Krishnan
+    given-names: Ranganath
+    affiliation: Capital One AI Labs
+  - family-names: Trivedi
+    given-names: Amit Ranjan
+    affiliation: University of Illinois Chicago
+preferred-citation:
+  type: conference-paper
+  title: "Argus: A Cross-Regime Study of Uncertainty Quantification Transferability in Single-Step GUI Grounding"
+  authors: *authors
+  collection-title: "NeurIPS 2026 Evaluations and Datasets Track"
+  year: 2026

argus_uq-0.1.0/DATA_CARD.md

@@ -0,0 +1,152 @@
+# Data Card - argus-uq
+
+Argus-UQ ships per-item uncertainty quantification scores for 28 cells
+(16 open-weight + 12 closed-API) from the paper "Argus: A Cross-Regime
+Study of Uncertainty Quantification Transferability in Single-Step GUI
+Grounding" (NeurIPS 2026 Evaluations and Datasets Track).
+
+## 1. Motivation
+
+### 1.1 Purpose
+
+Argus-UQ exists to make UQ methods comparable across vision-language
+models, datasets, and access tiers under one protocol. Users can
+load any cell, score any of the 27 (open-source) or 8 (closed-source)
+UQ methods, and reproduce every paper number from
+``data_release/derived/`` without re-running inference.
+
+### 1.2 Funding / acknowledgement
+
+Created by Divake Kumar, Sina Tayebati, Devashri Naik (University of
+Illinois Chicago); Amanda Sofie Rios, Nilesh Ahuja, Omesh Tickoo (Intel
+Labs); Ranganath Krishnan (Capital One AI Labs); Amit Ranjan Trivedi
+(University of Illinois Chicago). Funding and competing interests are
+stated in the paper.
+
+## 2. Composition
+
+### 2.1 Instances
+
+Each instance is one (model, dataset, item) triple with:
+
+- a stable item id
+- a binary correctness label (whether the agent's greedy click landed
+  in the gold bbox)
+- the normalised distance from the predicted click to the bbox
+- 27 (open-source) or 8 (closed-source) per-method UQ scores
+
+There are 28 cells (model x dataset combinations) and 19,538 instances
+across all cells (14,749 open-weight, 4,789 closed-API).
+
+### 2.2 Cell catalog
+
+**Open-source (16 cells)**: 4 models x 4 datasets
+
+| Models | Datasets |
+|---|---|
+| Q7  = Qwen2.5-VL-7B-Instruct       | V2  = ScreenSpot-v2 (1,272 items) |
+| Q72 = Qwen2.5-VL-72B-Instruct-AWQ  | SP  = ScreenSpot-Pro (1,581 items) |
+| UI  = UI-TARS-1.5-7B               | OSG = OSWorld-G non-refusal subset (564 items) |
+| PT  = POINTS-GUI-G-8B              | UIV = UI-Vision Element Grounding 501-item subset |
+
+**Closed-source (12 cells)**: 3 vendors x 4 datasets (V2 / SP / OSG /
+UIV; V2 and SP use the 300-item subsets)
+
+| Vendors |
+|---|
+| OAI = OpenAI GPT-5.4 |
+| Son = Anthropic Claude Sonnet 4.6 |
+| Gem = Google Gemini 3.1 Pro |
+
+### 2.3 Method panels
+
+Open-source cells carry **27 method scores per item** spanning seven
+families: logit (5), sampling (6), consistency (3), density (5),
+attention (3), prompt (3), VLM-native (2).
+
+Closed-source cells carry **8 OS-identical method scores per item**:
+self_consistency, semantic_entropy, lexical_similarity, ccp,
+verbalized_1s, verbalized_2s, hedge, imghedge. The other 19 methods
+are open-source-only because the closed APIs do not expose token
+logprobs, hidden states, or attention tensors.
+
+## 3. Collection process
+
+### 3.1 Open-source
+
+- Models: HuggingFace transformers; loaded onto NVIDIA RTX A6000 Ada
+  (48 GB) with bfloat16 precision (or int4-AWQ for the 72B Qwen).
+- Inference: per item, run greedy decode + N=5 stochastic samples
+  (temperature=1.0, top_p=0.95) + 3 prompt probes (P(True),
+  Verbalized-1S, Verbalized-2S) + 3 paraphrased-instruction HEDGE
+  variants + 3 image-perturbed IMG-HEDGE variants. Capture per-token
+  logprobs, per-token last-layer hidden states, and per-step
+  per-layer per-head attention tensors during greedy decoding.
+- Scoring: Phase B reads cached records, fits trainable methods
+  (SAPLMA, SEP, Mahalanobis, Mahalanobis-RMD, Mahalanobis-RDE) on a
+  calibration split, and scores the 27 methods on the test split.
+
+### 3.2 Closed-source
+
+- API surface: vendor-isolated agents under
+  ``argus_uq.models.closed_source.{openai, anthropic, google}``. The
+  released classes hold configuration only; generation methods raise
+  ``NotImplementedError``.
+- Per-item calls: 16 (greedy + 5 samples + 3 hedge_text + 3
+  hedge_image + p_true + verbalized_1s + 2 stages of verbalized_2s).
+- Per-cell budget: ~300-564 items x 16 calls = ~5k-9k API calls per
+  cell. ~$50-200 per cell at 2026 prices.
+
+## 4. Preprocessing / cleaning / labelling
+
+- Click parsing: Qwen2.5-VL tool-call output has its coordinates in
+  the model-resized image space; each agent's parser rescales to
+  original-image space using the smart_resize ratio.
+- Closed-source vendor renames (to match the open-source method names):
+  - ``coord_agreement`` -> ``ccp``
+  - ``hedge_text`` -> ``hedge``
+  - ``hedge_image`` -> ``imghedge``
+- Closed-source ``records.jsonl`` files contain no API keys. Local
+  image paths are replaced with dataset-relative paths or ``<cache>/``.
+  Model responses are verbatim and can quote text visible in the public
+  dataset screenshots, such as email addresses and file paths.
+
+## 5. Uses
+
+### 5.1 Recommended
+
+- Reproducing the paper's headline AUROC / PRR / AUSE / ECE numbers.
+- Comparing two UQ methods on a fixed cell (and confidence interval).
+- Constructing deployable conformal click-disks for selective
+  abstention in production grounding agents.
+- Ablating UQ methods on a custom set of records (via
+  ``argus_uq.manager.ScoringManager``).
+
+### 5.2 Out of scope
+
+- **Multi-step trajectory UQ.** Argus-UQ benchmarks per-step
+  grounding only.
+- **Inference-from-scratch closed-source experiments.** Vendor agents
+  in this library do not implement full inference; the bundled
+  closed-source records are the canonical artefact.
+- **Fine-tuning models.** Argus-UQ scores frozen VLMs.
+
+## 6. Distribution
+
+### 6.1 License
+
+- Library code: MIT (``LICENSE``).
+- Released per-item records, derived CSVs, and subset definitions:
+  CC-BY-4.0 (``LICENSE-DATA``).
+- Underlying datasets carry their own upstream licenses
+  (Apache 2.0 / research / MIT); see ``NOTICES.md``.
+
+### 6.2 Where it lives
+
+The code repository listed in ``pyproject.toml`` and ``CITATION.cff``
+holds the code and all of ``data_release/``.
+
+## 7. Maintenance
+
+The repository is the system of record. Issues, fixes, and method
+additions land via PRs.

argus_uq-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 The Argus-UQ Authors
+
+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.

argus_uq-0.1.0/LICENSE-DATA

@@ -0,0 +1,19 @@
+Argus-UQ released datasets and per-item records are licensed under the
+Creative Commons Attribution 4.0 International License (CC-BY-4.0).
+
+You are free to:
+  - Share -- copy and redistribute the material in any medium or format.
+  - Adapt -- remix, transform, and build upon the material for any purpose,
+    even commercially.
+
+Under the following terms:
+  - Attribution -- You must give appropriate credit, provide a link to the
+    license, and indicate if changes were made.
+
+Full license text: https://creativecommons.org/licenses/by/4.0/legalcode
+
+Note: this license covers the authors' own contributions (UQ scores, derived
+tables, subset definitions). The underlying datasets carry their own upstream
+licenses, and text or annotations taken from them (instructions, target boxes,
+item metadata) stay under those licenses. Vendor API responses stay subject to
+the vendors' terms. See NOTICES.md for the per-source attribution.

argus_uq-0.1.0/MANIFEST.in

@@ -0,0 +1,5 @@
+include CITATION.cff DATA_CARD.md setup.py
+recursive-include data_release/per_item_full *.parquet
+recursive-include data_release/derived *.csv
+recursive-include data_release/subsets *.json
+include data_release/manifest.json data_release/panel_drops.json data_release/README.md

argus_uq-0.1.0/NOTICES.md

@@ -0,0 +1,71 @@
+# NOTICES
+
+`argus-uq` builds on prior open-source work and cites the canonical
+references for every UQ method and metric implemented in the package.
+
+## Library architecture inspiration
+
+- **lm-polygraph** (Fadeeva et al. 2024). https://github.com/IINemo/lm-polygraph.
+  MIT. We adopt the StatCalculator + Estimator + UEManager design pattern.
+- **torch-uncertainty** (Lafage & Laurent 2025). https://github.com/ENSTA-U2IS-AI/torch-uncertainty.
+  Apache 2.0. Packaging convention (`pyproject.toml` extras) inspired this repo.
+
+## UQ method canonical references (also cited in each estimator's docstring)
+
+- MSP, perplexity, mean token entropy, sequence probability: standard
+  logit baselines (Hendrycks & Gimpel 2017; Malinin & Gales 2018).
+- Self-consistency: Wang et al. 2023 (arXiv 2203.11171).
+- Semantic entropy: Kuhn et al. 2023; Farquhar et al. *Nature* 2024.
+- MCSE / MCNSE: Malinin & Gales 2021.
+- Lexical similarity: Fomicheva et al. 2020.
+- CCP / CoCoA / CoCoA-1MCA: Vashurin et al. 2025 (AAAI).
+- Mahalanobis: Lee et al. 2018.
+- Mahalanobis-RMD: Ren et al. 2023 (arXiv 2209.15558).
+- Mahalanobis-RDE: KernelPCA + MinCovDet (Rousseeuw 1984) on the
+  positive class.
+- SAPLMA: Azaria & Mitchell 2023 (EMNLP).
+- SEP (Semantic Entropy Probes): Kossen et al. 2024 (arXiv 2406.15927).
+- FOCUS: Zhang et al. 2023.
+- RAUQ: Vazhentsev et al. 2023.
+- UQAC: attention-aggregation method (2025).
+- P(True): Kadavath et al. 2022 (arXiv 2207.05221).
+- Verbalized 1S: Tian et al. 2023 (arXiv 2305.14975).
+- Verbalized 2S: Xiong et al. 2024 (arXiv 2306.13063).
+- HEDGE / IMG-HEDGE: Gautam et al. 2025.
+- Inductive split CP: Vovk et al. 2005; Angelopoulos & Bates 2023.
+- Conformalized quantile regression (CQR): Romano et al. 2019.
+- Conformal Risk Control: Bates et al. 2024.
+
+## Metric references
+
+- AUROC: Hendrycks & Gimpel 2017 (incorrectness-prediction usage).
+- PRR (Prediction-Rejection Ratio): Malinin & Gales 2018.
+- AUSE: Ilg et al. 2018 (ECCV; per-item-error sparsification).
+- ECE / MCE: Guo et al. 2017 (ICML).
+- Brier score: Brier 1950.
+- Selective accuracy at coverage: Geifman & El-Yaniv 2017.
+- Risk-Coverage AUC: standard selective-prediction usage.
+
+## Datasets
+
+Licenses below are as declared by the upstream hosts.
+
+- **ScreenSpot-v2** (Apache 2.0). https://huggingface.co/datasets/OS-Copilot/ScreenSpot-v2
+- **ScreenSpot-Pro** (MIT). https://huggingface.co/datasets/likaixin/ScreenSpot-Pro
+- **OSWorld-G** non-refusal subset. https://github.com/xlang-ai/OSWorld-G
+  (repository: Apache 2.0; the Hugging Face copy `MMInstruction/OSWorld-G`
+  declares no license).
+- **UI-Vision Element Grounding** 501-item subset (MIT). https://huggingface.co/datasets/ServiceNow/ui-vision
+
+This repository does not redistribute any screenshots. It does redistribute
+item identifiers, and `data_release/closed_source_records/` also contains the
+instruction text, target boxes and item metadata from ScreenSpot-v2,
+ScreenSpot-Pro, OSWorld-G and UI-Vision, under the upstream terms above. To run anything
+that needs images, obtain the datasets from their upstream sources.
+
+## Closed-source APIs
+
+The closed-source matrix uses commercial APIs from OpenAI, Anthropic,
+and Google. Vendor SDK usage is governed by their respective Terms of
+Service. We release request / response records; we do not redistribute
+model weights.

argus_uq-0.1.0/PKG-INFO

@@ -0,0 +1,233 @@
+Metadata-Version: 2.4
+Name: argus-uq
+Version: 0.1.0
+Summary: Uncertainty quantification benchmark for VLM GUI grounding agents.
+Author: Divake Kumar, Sina Tayebati, Devashri Naik, Amanda Sofie Rios, Nilesh Ahuja, Omesh Tickoo, Ranganath Krishnan, Amit Ranjan Trivedi
+License: MIT License
+        
+        Copyright (c) 2026 The Argus-UQ Authors
+        
+        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.
+        
+Project-URL: Homepage, https://github.com/divake/argus-uq
+Keywords: uncertainty quantification,vision-language models,GUI grounding,computer-use agents,conformal prediction,benchmark
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: <3.13,>=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: LICENSE-DATA
+License-File: NOTICES.md
+Requires-Dist: numpy<2.0,>=1.26
+Requires-Dist: pandas<3.0,>=2.2
+Requires-Dist: scipy<2.0,>=1.13
+Requires-Dist: scikit-learn<2.0,>=1.5
+Requires-Dist: pyarrow<26,>=16.0
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: omegaconf>=2.3
+Requires-Dist: click<9.0,>=8.1
+Requires-Dist: tqdm>=4.66
+Requires-Dist: pillow>=10.0
+Provides-Extra: whitebox
+Requires-Dist: torch<3.0,>=2.2; extra == "whitebox"
+Requires-Dist: torchvision<1.0,>=0.17; extra == "whitebox"
+Requires-Dist: transformers>=4.45; extra == "whitebox"
+Requires-Dist: accelerate>=1.0; extra == "whitebox"
+Requires-Dist: huggingface-hub>=0.25; extra == "whitebox"
+Requires-Dist: opencv-python-headless>=4.9; extra == "whitebox"
+Provides-Extra: closed
+Requires-Dist: openai>=1.50; extra == "closed"
+Requires-Dist: anthropic>=0.40; extra == "closed"
+Requires-Dist: google-genai>=0.3; extra == "closed"
+Provides-Extra: paper
+Requires-Dist: matplotlib<4.0,>=3.9; extra == "paper"
+Requires-Dist: seaborn>=0.13; extra == "paper"
+Requires-Dist: jinja2>=3.1; extra == "paper"
+Provides-Extra: hydra
+Requires-Dist: hydra-core>=1.3; extra == "hydra"
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0; extra == "dev"
+Requires-Dist: pytest-cov>=5.0; extra == "dev"
+Requires-Dist: ruff>=0.6; extra == "dev"
+Requires-Dist: mypy>=1.10; extra == "dev"
+Requires-Dist: pre-commit>=3.5; extra == "dev"
+Provides-Extra: docs
+Requires-Dist: mkdocs>=1.6; extra == "docs"
+Requires-Dist: mkdocs-material>=9.5; extra == "docs"
+Requires-Dist: mkdocstrings[python]>=0.26; extra == "docs"
+Provides-Extra: all
+Requires-Dist: argus-uq[closed,dev,docs,hydra,paper,whitebox]; extra == "all"
+Dynamic: license-file
+
+# argus-uq
+
+A Python library for uncertainty quantification on vision-language-model
+GUI grounding agents.
+
+`argus-uq` is the first reproducible benchmark library that unifies
+**twenty-seven UQ methods across seven families** (logit, sampling,
+consistency, density, attention-trace, verbalised, VLM-native) on the
+four canonical single-step grounding datasets (ScreenSpot-v2,
+ScreenSpot-Pro, OSWorld-G, UI-Vision) for both **open-weight VLMs**
+(Qwen2.5-VL family, UI-TARS-1.5-7B, POINTS-GUI-G-8B) and
+**closed-API frontier vendors** (OpenAI GPT, Anthropic Claude, Google
+Gemini, paneled across the eight vendor-API-feasible methods that share
+formula identity with the open-weight panel).
+
+It loads per-item records and 50-seed metric matrices for **16
+open-weight cells and 12 closed-API cells**, re-scores any of the 27 UQ
+methods on user-supplied grounding records, and emits **deployable
+inductive-conformal click-disks** with formal marginal-coverage
+guarantees in three variants (Disk-Fixed, Disk-Normalized, Disk-CQR).
+
+The package accompanies the paper *Argus: A Cross-Regime Study of
+Uncertainty Quantification Transferability in Single-Step GUI Grounding*
+(**NeurIPS 2026 Evaluations and Datasets Track**).
+
+## Install
+
+Python 3.10 to 3.12.
+
+```bash
+pip install argus-uq
+```
+
+The wheel bundles the per-item scores, derived tables and subset
+definitions, so `load`, `score`, `conformal_disk` and the CLIs work
+without this repository. The raw closed-source API records
+(`data_release/closed_source_records/`) and the master long CSVs are in
+this repository only; no public call needs them.
+
+To work from source:
+
+```bash
+git clone https://github.com/divake/argus-uq
+cd argus-uq
+python -m venv .venv && source .venv/bin/activate
+pip install -e .                   # core library
+pip install -e .[whitebox]         # + HF transformers VLM stack
+pip install -e .[closed]           # + closed-source vendor SDKs
+pip install -e .[paper]            # + paper-reproducing scripts
+pip install -e .[dev]              # + linting and testing
+pip install -e .[all]              # everything
+```
+
+To use your own copy of the data, set
+`export ARGUS_DATA_DIR=/path/to/data_release`.
+
+## Quickstart
+
+```python
+import argus_uq
+
+# Load a cell - per-item records and per-seed scores from the released
+# artifacts.
+cell = argus_uq.load("PTxOSG")  # POINTS-GUI-G x OSWorld-G
+
+# Score any of the 27 UQ methods. Returns AUROC, PRR_normalized@0.5, AUSE,
+# AURC, ECE_iso, Brier_iso with bootstrap CIs (multi-seed if seeds=50).
+result = argus_uq.score(cell, method="saplma", seeds=50)
+print(result.auroc, result.auroc_std)   # ~0.802, ~0.011
+
+# Build a deployable conformal click-disk at marginal coverage 1 - alpha.
+disk = argus_uq.conformal_disk(
+    cell,
+    plug_in="saplma",             # any of the 27 UQ methods or "none"
+    alpha=0.10,
+    variant="normalized",         # "fixed" | "normalized" | "cqr"
+)
+print(disk.radius_pixels, disk.coverage)
+```
+
+## Reproducing the paper
+
+```bash
+argus_reproduce --list                     # every table name
+argus_reproduce --table cells              # the 16+12 cell catalog
+argus_reproduce --table leaderboard        # open-weight mean AUROC across cells
+argus_reproduce --table leaderboard_closed # closed-API mean AUROC across cells
+argus_reproduce --table rho                # 16x16 cross-cell Spearman matrix
+argus_reproduce --all                      # every table
+```
+
+These commands print the released CSVs in `data_release/derived/`, which
+hold the numbers reported in the paper. `argus_uq.score` and
+`argus_uq.conformal_disk` recompute from the per-item scores with the
+package's own random splits. Their AUROC means differ from the CSVs by
+up to about 0.006, the top method of every closed-API cell is unchanged,
+and their AUSE and conformal radii use different conventions
+from the CSVs, so use the CSVs to compare against the paper.
+
+`argus_capture` scores raw per-item signals (logprobs, hidden states,
+sample clicks; see `argus_uq.records.Record`). The released parquets hold
+finished scores, not raw signals, so it returns NaN or chance-level
+metrics on them. Use `argus_score` for released cells.
+
+## Layout
+
+```
+src/argus_uq/        importable library
+├── api.py           the 3-entry-point Listing 1 surface
+├── cell.py          Cell dataclass + per-item record loader
+├── factory.py       name -> class registry
+├── manager.py       ScoringManager (topo-orders stat calculators + estimators)
+├── estimators/      27 UQ methods (one file per family)
+├── stat_calculators/ inputs to estimators (greedy, samples, hidden, attention)
+├── models/          agent abstractions (whitebox HF + closed-source APIs)
+├── envs/            ScreenSpot-v2, ScreenSpot-Pro, OSWorld-G, UI-Vision
+├── ue_metrics/      AUROC, PRR, AUSE, AURC, ECE, Brier, isotonic
+├── conformal/       Disk-Fixed, Disk-Normalized, Disk-CQR, CRC
+├── analysis/        leaderboard, cross-cell rho, family aggregates
+├── data/            released-artifact loaders (parquets + CSVs)
+└── utils/           paths, seeds, logging
+
+examples/            quickstart.py and two cross-cell examples
+tests/               225 unit + integration tests + 7 live smoke tests
+data_release/        16+12 per-item parquets, 29 derived CSVs, 4 subsets
+```
+
+## License
+
+- Code: MIT (`LICENSE`).
+- Released datasets and per-item records: CC-BY-4.0 (`LICENSE-DATA`).
+- Third-party attributions: `NOTICES.md`.
+
+## Citation
+
+If you use `argus-uq`, please cite the paper (`CITATION.cff` has the same
+information):
+
+```bibtex
+@inproceedings{kumar2026argus,
+  title     = {Argus: A Cross-Regime Study of Uncertainty Quantification
+               Transferability in Single-Step GUI Grounding},
+  author    = {Kumar, Divake and Tayebati, Sina and Naik, Devashri and
+               Rios, Amanda Sofie and Ahuja, Nilesh and Tickoo, Omesh and
+               Krishnan, Ranganath and Trivedi, Amit Ranjan},
+  booktitle = {NeurIPS 2026 Evaluations and Datasets Track},
+  year      = {2026}
+}
+```