IRAC Annotated Data Overview

Updated 19 April 2026

IRAC Annotated Data is a comprehensive term covering both astrophysics imaging products and legal reasoning datasets with detailed, traceable annotations.
In astrophysics, these datasets include science mosaics, exposure maps, and PSF models, which enable precise photometric assessments and multi-wavelength studies.
In the legal domain, IRAC-annotated data structures decisions into Issue, Rule, Application, and Conclusion labels, benchmarking legal NLP and structured reasoning models.

The term "IRAC Annotated Data" encompasses a range of high-level, systematically curated datasets and annotation protocols centered on data collected with the Spitzer Infrared Array Camera (IRAC), as well as corpora where the term "IRAC" refers to the legal reasoning framework (Issue, Rule, Application, Conclusion). In scientific contexts (primarily astrophysics and cosmology), "IRAC Annotated Data" refers to imaging products, catalogs, ancillary maps, PSF models, and metadata necessary for precise photometric, reliability, and verification analyses. In legal contexts, it designates rigorously labeled datasets for benchmarking reasoning—particularly legal reasoning aligned to the IRAC method—providing structured ground truth for supervised models and evaluation. The following exposition surveys principal IRAC-annotated datasets, methodologies for annotation and verification, and their astrophysical or legal scientific significance.

1. Definition and Scope of IRAC Annotated Data

"IRAC Annotated Data" in astrophysics denotes public releases derived from deep or wide-area IRAC imaging surveys. Such datasets integrate science mosaics, exposure-time maps, PSF models, per-exposure subproducts, provenance metadata, and catalog-level annotations, supporting both detection and assessment workflows at the pixel and source levels. These releases target scientific exploitation, enabling studies of high-redshift galaxies, variability, and systematic effects, and are indispensable for deep multi-wavelength analyses.

In the legal domain, "IRAC Annotated Data" refers to datasets in which legal scenarios or court decisions are labeled according to the IRAC reasoning paradigm. This annotation includes explicit demarcation of Issues, cited Rules, structured Applications (reasoning steps), and Conclusions—supporting information extraction, legal reasoning, and machine learning evaluation tasks in legal NLP (Kang et al., 2023, Kang et al., 2024, Jang et al., 8 Jan 2026).

2. Major IRAC-Annotated Data Releases: Survey Products and Data Models

Astrophysical Context

The "Ultradeep IRAC Imaging Over The HUDF And GOODS-South" program defines the contemporary state of the art in IRAC data annotation (Labbe et al., 2015). The release comprises:

Science Mosaics: FITS images at 0.3″/pix, WCS-aligned to CANDELS/GOODS-South, with coverage up to 200 hr in the EGS, HUDF, and associated parallel fields.
Exposure-Time Maps: Per-pixel integration time annotations (hour units), essential for S/N assessment and reliable depth characterization.
Per-AOR Mosaics: 706 submosaics—one per Astronomical Observation Request (AOR) and channel—enabling cross-epoch verification, artifact tracing, and variability assessment.
Ancillary Maps: Tier masks for integration depth, position-angle maps, and weight maps for structured reliability analysis.
PSF Maps: Empirically constructed, spatially varying PSF maps from a two-stage process involving >2000 stellar cutouts, median stacking, rotation, and exposure-weighted coaddition, yielding continuously sampled PSFs crucial for deblending, prior-based photometry, and comparison with HST/WFC3 data.
Data Provenance: Each step, from background subtraction to artifact rejection and astrometric registration, is traceable via pixel-level and catalog-level annotation.

Other canonical surveys hosting IRAC-annotated data include:

Spitzer-South Pole Telescope Deep Field (SSDF): Band-merged catalogs (3.6 and 4.5 μm), completeness curves, SExtractor flagging, astrometric and photometric calibration metadata (Ashby et al., 2013).
Spitzer IRAC Equatorial Survey (SpIES): Multi-million-source catalogs, extensive coverage maps, per-source reliability and contamination flags (Timlin et al., 2016).
SIMES in the South Ecliptic Pole: Structured columns for photometry, S/N curves, multiwavelength associations, and reliability indicators for cross-matching to MIPS/SPIRE/optical sources (Baronchelli et al., 2016).
EGS Multi-wavelength Catalog: Panchromatic SEDs and cross-matched catalog-level metadata facilitating spectrophotometric estimation, with completeness and deblending annotations (Barro et al., 2011).
IRAC–Herchel SPIRE Matching: Catalog-level reliability, color-magnitude cuts, and likelihood-ratio formalism for SPIRE–IRAC association (Kim et al., 2011).

Legal Context

LegalSemi: 54 Malaysian contract law scenarios, human-annotated for IRAC roles, JSON-encoded, and linked to structured legal knowledge graphs (SKG/SKE) in Neo4j. Annotations include explicit span-level legal concept identification, issue formulation, citation of statutory and precedent rules, stepwise application logic, and labeled conclusions (Kang et al., 2024).
SIRAC: 50-case corpus annotated with fully serialized IRAC structure, scenario context, explicit IF…THEN application logic, and multi-level adjudication. Each scenario's annotation is provided as a JSON object under top-level keys: Issue, Rules, Application, Conclusion (Kang et al., 2023).
PILOT-Bench: 18,049 PTAB patent appeal decisions with IRAC-mapped labels for Issue, Rule, and Conclusion, LLM-segmented into appellant and examiner roles, and evaluated with metrics including micro/macro-F₁, EM, and Hamming loss (Jang et al., 8 Jan 2026).

3. Annotation Procedures, Metadata Schema, and Data Provenance

Astrophysics

Comprehensive IRAC annotation is defined by multi-level provenance and rigorous artifact/model documentation:

Pixel-level: Per-exposure coverage, MAD/statistics maps, exposure history, and systematics tracking (column pull-down, persistence masking).
Source-level: Catalog columns for position, photometry, flags (blending, saturation, edge truncation), S/N, PSF correction, local depth, and reliability.
PSF modeling: Empirical construction involves co-registering thousands of stellar cutouts, then mapping the field with local roll-angle-weighted super-PSFs, yielding spatially varying, high-dynamic-range PSF annotations.
Verification layers: Sub-exposure mosaics and position-angle maps allow end-users to isolate sources observed under particular conditions or instrument orientations for targeted reliability or variability research.

Legal

Annotation is executed via stratified, protocol-driven workflows:

Expert double-annotation: Independent markup followed by adjudication and IAA measurement (e.g., Cohen's κ).
Role-based segmentation: Automated or human labeling of fact, argument, opinion, and evidence sentences, aligned with IRAC subfields.
Schema: Uniform JSON formats with scenario metadata, span-mapped legal concepts, explicit step-logic for Application, statutorily referenced Rules, and conclusion string. Knowledge graph node mappings enable automated traversal and legal-concept grounding (LegalSemi).
Label normalization: Unified curation via lower-casing, punctuation stripping, synonym merging using LLMs (PILOT-Bench).
Inter-annotator validity: Agreement metrics (e.g., IAA>0.8, κ=0.55–0.75), human-in-the-loop QA, rubric-based human and automatic quality control (ρ≈0.89 correlation with human judgment).

4. Reliability, Completeness, and Sensitivity Characterization

In astrophysics, reliability and depth are quantified by injection-recovery statistics, Poisson noise scaling, and completeness curves:

Noise Scaling: Empirical fits to the noise-depth relation, e.g., $\sigma_{3.6}(\mathrm{nJy}) = 172\,t_\mathrm{exp}^{-0.453}$ over $t_\mathrm{exp}=20$ –$200$ hr (1σ point-source sensitivity to 15 nJy at 3.6 μm); observed scaling matches $t_\mathrm{exp}^{-1/2}$ (Labbe et al., 2015).
Confusion and Catastrophic Residuals: The use of prior-based fitting suppresses classical confusion below expectations (e.g., only ∼12% with $>$ 5σ catastrophic residuals versus the classical 0.6 μJy floor) (Labbe et al., 2015).
Completeness: Monte Carlo source injections yield completeness values (e.g., 93% at 16.0 mag, 71% at 18.0 mag, and 39% at 19.0 mag for SSDF (Ashby et al., 2013), 50% at 22.5 mag for H-ATLAS/IRAC (Kim et al., 2011)). Catalog-level reliability is further tracked by SExtractor and survey-specific flags, with astrometric accuracy benchmarked against external catalogs (e.g., 0.05″–0.07″ r.m.s. vs. 2MASS).

In legal annotation, reliability is evaluated with inter-annotator agreement, consistency scoring, recall/precision/F1, and, for automatically generated splits, determinism guarantees.

5. Applications and Scientific Use Cases

Astrophysics

IRAC annotated data empower the following:

High-z Galaxy Studies: Direct detection of galaxies at $z>7$ and deep stellar population constraints (e.g., detected to $H_\mathrm{AB}=28$ in IUDF (Labbe et al., 2015)).
Prior-based Photometry: Use of HST/WFC3 priors and spatial PSF maps to extract robust IRAC fluxes despite severe crowding.
Reliability Studies: Sub-exposure/AOR-resolved mosaics permit systematic assessment of source variability, instrumental systematics, and photometric consistency across observing conditions.
Multi-wavelength Cross-Matching: Provision of band-merged catalogs and ancillary matching metrics for robust identification of infrared counterparts to submillimeter, radio, or X-ray sources.
Statistical Studies: Integral number counts, mass and SFR function estimation, and color selection of high-redshift clusters and rare source populations.
JWST Planning: Deep IRAC fields function as preparatory datasets for JWST high-redshift target selection and survey design.

Legal

IRAC-annotated legal datasets are central for:

Benchmarking LLM Reasoning: PILOT-Bench enables zero/few-shot and supervised evaluation of structured legal reasoning and classification (Jang et al., 8 Jan 2026).
Neuro-symbolic Reasoners: LegalSemi and SIRAC form test-beds for hybrid extraction/generation pipelines, with explicit annotation facilitating RAG or chain-of-thought paradigms (Kang et al., 2023, Kang et al., 2024).
Retrieval/natural language inference: Granular IRAC markup supports information retrieval, rule extraction, and concept disambiguation in complex legal texts.
Assessment of Model Alignment: Rigorously annotated corpora furnish gold standards for measuring the fidelity of machine-generated legal reasoning, from issue decomposition to conclusion, supporting model selection and prompt engineering.

6. Data Access, Formats, and Community Resources

All major IRAC-annotated datasets are released in transparent, machine-readable formats with extensive documentation and public code availability.

Astrophysics: FITS mosaics, per-exposure stacks, ancillary maps, and band-merged catalogs (ASCII/FITS) disseminated via dedicated archives (e.g., CANDELS/GOODS/IRSA for IUDF, SSDF, SpIES, SIMES) (Labbe et al., 2015, Ashby et al., 2013, Timlin et al., 2016, Baronchelli et al., 2016).
Metadata: Full SExtractor parameter dumps, calibration tables, completeness and reliability curves, and per-source flags/annotations are typically included.
Legal: JSON, TXT, CSV, and Neo4j/CSV graph formats are standard. Example: PILOT-Bench provides complete code, prompts, label dictionaries, and task utilities on GitHub; LegalSemi supplies bundled scenario annotations, SKG graph dumps, and REST APIs for query access (Jang et al., 8 Jan 2026, Kang et al., 2024, Kang et al., 2023).

Open-access repositories, versioned DOIs, and value-added web interfaces (e.g., Rainbow Navigator for EGS) allow for advanced querying, visualization, and integration into new scientific workflows.

These IRAC-annotated datasets, spanning astrophysical and legal contexts, are characterized by precise multi-level annotation, rigorous quality controls, and a design philosophy focused on traceability, reproducibility, and extensibility—making them exemplary for high-impact data-driven research in their respective domains (Labbe et al., 2015, Jang et al., 8 Jan 2026, Kang et al., 2024, Kang et al., 2023).