ADEPTS: Data Pipeline & AI Agent Framework

• ADEPTS is a dual-use term denoting both an automated data extraction system for the CHARIS spectrograph and a capability framework for human-centered AI agent design.
• The CHARIS ADEPTS system streamlines astronomical data processing through modular pipeline stages, parallel execution, and robust database traceability.
• The human-centered ADEPTS framework stipulates six competencies that benchmark AI agents on trustworthiness, transparency, and safety.

ADEPTS refers to two distinct entities in contemporary research literature: (1) the Automated Data Extraction, Processing, and Tracking System for the CHARIS integral-field spectrograph in observational astronomy (Tobin et al., 2020), and (2) ADEPTS, a capability framework for human-centered agent design establishing core criteria for AI agent capability and trustworthiness (D'Oro et al., 18 Jul 2025). Both are significant in their respective domains: the first advances the state of the art in large-scale data reduction pipelines for astronomical instrumentation, while the second offers a unifying cross-disciplinary vocabulary for designing, benchmarking, and governing user-facing AI agents.

1. ADEPTS for CHARIS: Architecture and Workflow

The Automated Data Extraction, Processing, and Tracking System (ADEPTS) was developed to automate the data reduction pipeline (DRP) of the CHARIS integral-field spectrograph. Its workflow orchestrates five main stages:

1. File Monitoring: Detects arrival of raw FITS files via filesystem monitoring.
2. Database Population: Parses FITS headers, flags corrupted files, and sorts data into types (darks, flats, science, skyflats) within a structured SQLite database (tables: CHARISData, darks, buildcal, skyflats, obs).
3. Calibration (auto_buildcal): Delegates to the CHARIS DRP to solve for lenslet PSF geometry and wavelength calibration. The wavelength solution $\lambda(p)$ is typically modeled as a low-order polynomial in pixel position: $\lambda(p) \approx a_0 + a_1 p + a_2 p^2$.
4. Extraction (auto_extractcube): Runs PSF-based extraction for each lenslet, integrating detector flux into 3D datacubes: $$C(\lambda_k, i, j) = \sum_{(x, y) \in \mathrm{PSF}_{ij}} w_{xy}(\lambda_k) D(x, y)$$.
5. Post-processing: Applies configurable algorithms such as ADI, SDI, LOCI, KLIP, or custom PSF-subtraction routines.

Parallelization is leveraged primarily through Python’s multiprocessing and CHARIS DRP’s internal support. ADEPTS on modern 72-core hardware reduces full-night pipeline latencies from several hours to under one hour, with calibration and extraction benefiting most from parallel execution (Tobin et al., 2020).

2. CHARIS ADEPTS: Database and Modularity

The core ADEPTS system is underpinned by a normalized SQLite database capturing the complete experimental provenance of every reduction run:

• CHARISData: Tracks all files by unique file_id, metadata, and bad_flags.
• darks/buildcal/skyflats/obs: Structured for traceability across reduction stages; relationships are 1-to-many (e.g., one dark group to many buildcal products).
• Module-Extensibility: Each post-processing module (e.g., pyKLIP or custom IDL) may register database tables defined in its configuration/init, enabling fine-grained tracking of processed data products and parameter sweeps.

SQL queries are used for scientific data mining, e.g., retrieving all science cubes for given targets and filters with $SELECT$-joins across tables. The modular directory and configuration system allows rapid scaling and future-proofing (e.g., cluster integration, web interfaces) (Tobin et al., 2020).

3. ADEPTS Capability Framework for Human-Centered Agents

ADEPTS, as proposed by Liu, Huang, Amershi, and co-authors (D'Oro et al., 18 Jul 2025), addresses a critical need for user- and policy-facing standards for interactive AI agent development. It defines a contractually minimal set of six “user-facing” competencies that any trustworthy agent should manifest:

Principle	Capability (ADEPTS)	Guarantees
1. Autonomous Actuation	Actuation	Executes user intent within preset constraints
2. Intent Disambiguation	Disambiguation	Actively clarifies user goal/context/constraints
3. Situational Evaluation	Evaluation	Tracks and explains current state and actions
4. Adaptive Personalization	Personalization	Learns/adapts to user preferences/abilities
5. Operational Transparency	Transparency	Exposes rationale, plan, and past actions
6. Proactive Safety	Safety	Prevents harm pre-emptively; enforces privacy, etc.

For each, the framework defines hierarchical tiers of sophistication (e.g., Actuation Prompt Tiers 1–5 span from simple widgets to omni-modal interactions), enabling granular benchmarking and compliance assessment. This matrix operationalizes “understandability, controllability, and trustworthiness” as concrete, observable benchmarks (D'Oro et al., 18 Jul 2025).

4. ADEPTS Framework Usage and Distinctions

Unlike UX heuristics, engineering taxonomies, or high-level ethics frameworks that often operate independently, ADEPTS is designed as an interface-layer contract. It is agnostic to agent implementation details—algorithmic, neural, or hybrid—but provides a unified vocabulary with which research, engineering, UX, and compliance teams can jointly:

• Audit agent capability gaps by mapping product flows onto ADEPTS tiers,
• Design capability-driven benchmarks for task-specific evaluation,
• Express regulatory or governance requirements in unambiguous, tiered terms (“must reach tier-4 Safety and tier-3 Transparency for launch”),
• Complement existing heuristic and pipeline-based frameworks by specifying what must be supported rather than how.

Example agent behaviors are given by the authors for each tier along multiple axes (e.g., Disambiguation tier-3: rejects underspecified commands and prompts for clarification; Safety sub-axes targeting misuse, misbehavior, or prompt injection) (D'Oro et al., 18 Jul 2025).

5. Technical and Methodological Distinctions: ADEPTS vs. ADEPT

It is important to distinguish ADEPTS (the CHARIS automation system and the agent design framework) from terminologically similar but unrelated methodologies such as ADEPT (including Adaptive Diffusion Environment for Policy Transfer Sim-to-Real (Yu et al., 2 Jun 2025), Adaptive Dynamic Early-Exit Process for Transformers (Yoo et al., 7 Jan 2026), A DEbiasing PrompT Framework (Yang et al., 2022), or Adaptive Decomposed Prompt Tuning (Tang et al., 6 Jan 2025)). None of these latter use-cases relate to the CHARIS data pipeline or to the cross-disciplinary capability framework for agent design.

In summary, ADEPTS denotes both a leading-edge automation pipeline for astronomical data reduction (notably for CHARIS (Tobin et al., 2020)) and a succinct, cross-disciplinary capability contract for interactive, human-centered AI agents (D'Oro et al., 18 Jul 2025). The two systems are united by their commitment to modularity, scalability, and the reduction of friction between technical sophistication and practical usability, though their application domains and underlying architectures are entirely distinct.