ITACA_144: Bias Audit & HPXeEL Detector

• ITACA_144 is a dual-natured framework integrating a bias audit tool for AI hiring compliance under NYC Local Law 144 and a high-pressure xenon detector concept for 0νββ searches.
• It computes key fairness metrics such as impact ratio, statistical parity difference, and effective bias while managing demographic subgroup challenges with statutory thresholds.
• The HPXeEL detector employs real-time 3D ion-track imaging and a precise MARS system to achieve ultra-low background rates for rare-event detection.

ITACA_144 refers to two distinct domains in contemporary research, each notable within its respective field: (1) the ITACA_144 bias audit tool for AI hiring compliance under New York City’s Local Law 144 (Clavell et al., 2024), and (2) the ITACA_144 high-pressure xenon (HPXeEL) detector concept for neutrinoless double beta decay ($0νββ$) searches (Gómez-Cadenas et al., 2 Apr 2026). Each “ITACA_144” instance addresses rigorous problems—automated legal compliance for algorithmic hiring on the one hand, and ultra-low background particle tracking for rare event search on the other—with unique architectural, methodological, and technical solutions.

1. ITACA_144 for Automated Bias Auditing in AI Hiring

ITACA_144, as developed in the context of NYC Local Law 144, is a software tooldistilled from the broader ITACA_OS bias auditing suite, tailored to the statutory requirements that AEDTs undergo annual independent bias audits. The core function of ITACA_144 is the computation, validation, visualization, and reporting of selection-rate fairness metrics—most notably the impact ratio (IR), statistical parity difference (SPD), and effective bias (EB)—according to pre-set legal criteria.

Workflow Overview

Stage	Description	Output/Action
Data Ingestion	Acquires 12-month applicant-level data with protected-class labels, decisions, and timestamps	Raw audit dataset
Pre-Processing	Schema validation, demographic check, subgroup (<2%) flagging	Validated, annotated data
Metric Computation	IR, SPD, EB, intersectional subgroup analysis, application of 80% rule and configurable thresholds	Bias/fairness metrics and pass/fail results
Reporting	Generation of PDF/HTML audits, tables, metric visualizations, summaries, and recommendations	Formal audit report
Audit Export	Compliance package creation for NYC Office of Technology & Innovation	Raw results, codebook, compliance attestation

The tool defaults to “impact ratio only” computation as required explicitly by Local Law 144 (the "lite" variant), embeds the 80% rule and 2% group-size minimums, and pre-populates New York City-specific census benchmarks for demographic representativity (Clavell et al., 2024).

2. Data Requirements, Demographic Handling, and Inclusiveness

For compliance, ITACA_144 mandates comprehensive applicant-level records: applicant ID, protected-class group label(s), model decision, optional ground-truth hire outcome, and temporal eligibility (within last 12 months). Missing demographic attributes are imputed as “Unknown,” excluded from IR computation but tracked separately in aggregate counts. Groups constituting less than 2% of the total are flagged; exclusion is permitted under statute, but ITACA_144 prompts for aggregation or specialized statistical methods, resisting outright omission. Intersectional analyses (e.g., race × sex) are automatically computed if subgroup counts exceed a 50-record default threshold.

3. Fairness Metrics and Thresholds

ITACA_144 implements and reports:

• Impact Ratio (IR):

$$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$

An IR < 0.8 (i.e., falling below the 80% rule) triggers a “fail” flag: $$\mathrm{IR}_{A,\mathrm{majority}} < 0.8 \implies \text{Compliance Violation}$$.

• Statistical Parity Difference (SPD):

$$\mathrm{SPD}_{A,B} = P(\text{decision}=1 \mid A) - P(\text{decision}=1 \mid B)$$

• Effective Bias (EB):

$$\mathrm{EB}_G = \left| \frac{N_{G,1}/\sum_G N_{G,1}}{N_G/\sum_G N_G} - 1 \right|$$

This quantifies divergence of subgroup hiring rates from applicant pool representation.

Additional built-in features include support for stricter-than-statutory thresholds and flexible aggregation rules for rare/underrepresented groups (Clavell et al., 2024).

4. Implementation Barriers and Best Practice Recommendations

Key technical and legal challenges include:

• Data reliability: Risk of underreporting, mislabeling, or missing class labels in audit data.
• Small/intersectional subgroup underrepresentation: The 2% group-size rule, while legally permitted, can systemically omit analysis of the most vulnerable populations.
• Proxy attribute bias: Use of non-protected but correlated features (e.g., ZIP code) can enable indirect discrimination not detectable by selection-rate calculations.
• Remediation enforcement: No statutory requirement exists for post-audit remediation if thresholds are violated.

Recommendations from audit practice include enforcing 5% on-premises ground-truth retesting, lowering/excising the 2% exclusion for groups with mandatory “other” aggregation, benchmarking against census data for representativity, and comprehensive pipeline documentation to ensure traceability. Auditors should be provided both model-level and outcome-level data to assess drift and effective bias longitudinally.

5. ITACA_144 in High-Pressure Xenon TPCs for $0νββ$

A separate instantiation of ITACA_144—articulated in the context of high-pressure xenon time-projection chambers—combines a 1 tonne, 15 bar HPXeEL TPC (energy resolution ≲1% FWHM) with real-time 3D ion-track imaging via a repositionable Topmetal CMOS ASIC detector (NAUSICA), and a Magnetically Actuated Rotor System (MARS) for precise spatial positioning of the ion sensor (Gómez-Cadenas et al., 2 Apr 2026).

Detector Parameters

Parameter	Value	Note
Xenon Mass ($M_\mathrm{fid}$)	≈1050 kg	1 tonne fiducial
Pressure	15 bar	Pure Xe
Drift Field ($E$)	200 V/cm	150 cm drift
EL Region Field ($E_\mathrm{EL}$)	≈2.1 kV cm⁻¹ bar⁻¹	FAT-GEM, 5 mm
Electron Drift Velocity ($v_e$)	≈1 mm/μs	$$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$0 ms
Ion Drift Velocity ($$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$1)	≈10 cm/s	$$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$2 s
Ion Spatial Diffusion ($$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$3)	≈1 mm	Over full drift

The NAUSICA CMOS ion detector consists of a 16 × 16 cm² active area with 2 mm pixel pitch, affording 0.58 mm transverse and 1 mm longitudinal sampling, matched to the natural diffusion scale for Xe₂⁺ ions. Readout provides time-sampled 3D imaging.

6. MARS Mechanical Positioning and Systematics

The MARS mechanism utilizes two concentric rotors, magnetically coupled through a titanium well, to enable $$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$4 actuation of the ion sensor in <1.1 s, well within the 15 s ion arrival window. The dead layer due to transit time is ≈110 mm, corresponding to ≈7.3% of the drift region. Aerodynamic and electrostatic analyses show that mechanical disturbances are confined within <4.3 mm and efficiently blocked by the ion-focusing grid (IFG), safeguarding drift integrity at high pressure. Vertical lift of 10 mm positions the ion plate to dock with minimal risk of disturbed gas interacting with the signal (Gómez-Cadenas et al., 2 Apr 2026).

7. Sensitivity to $$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$5 and Background Rejection

The combined background rate in the region of interest (ROI, 2400–2500 keV) due to $$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$6Bi and $$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$7Tl is calculated as $$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$8 ev/ton·yr after topological and energy cuts. This is substantially lower than conventional HPXeEL TPCs lacking ion-track discrimination (B ≈ 0.7 ev/ty). The projected median exclusion sensitivity at 90% CL for the $$\mathrm{IR}_{A,B} = \frac{P(\text{decision}=1 \mid \text{group}=A)}{P(\text{decision}=1 \mid \text{group}=B)} = \frac{N_{A,1} / N_A}{N_{B,1} / N_B}$$9 half-life is

$$\mathrm{IR}_{A,\mathrm{majority}} < 0.8 \implies \text{Compliance Violation}$$0

Even at 5 ton·yr exposure, sensitivity remains above $$\mathrm{IR}_{A,\mathrm{majority}} < 0.8 \implies \text{Compliance Violation}$$1 yr, directly probing the normal neutrino mass hierarchy window.

---

ITACA_144 thus designates both a legally compliant, automated bias audit framework for hiring AI (Clavell et al., 2024), and an advanced HPXeEL detector concept offering subdominant backgrounds for $$\mathrm{IR}_{A,\mathrm{majority}} < 0.8 \implies \text{Compliance Violation}$$2 decay discovery (Gómez-Cadenas et al., 2 Apr 2026). Both illustrate domain-specific, real-time, and standards-oriented measurement and analysis pipelines, setting benchmarks in their respective technical and regulatory landscapes.