The Phantom Workforce

1. The Ghost Job Economy: February 2026

The global labor market of February 2026 is defined by a growing disequilibrium between digital signal and operational intent. Job postings increasingly serve secondary functions — market signaling, talent stockpiling, and internal compliance — rather than immediate recruitment. The “hires-per-job-posting” ratio has collapsed from 0.75 in 2019 to below 0.5 in late 2025, meaning fewer than half of all active listings correspond to a genuine, funded vacancy.

Macroeconomic Drivers

• The Hiring Freeze Paradox. Entities such as Salesforce, Amazon, and Heineken announced aggressive hiring freezes in early 2026, yet ATS platforms like Workday keep postings live for weeks after internal freeze orders. This “Zombie Job” window creates critical risk: listings are technically active but may no longer represent active openings.
• The Internal Compliance Trap. Government and Education sectors exhibit ghost rates as high as 60% and 50%, driven by regulations mandating public posting even when internal candidates are pre-selected. Ontario enacted legislation in 2026 requiring employers to disclose whether a posting is for an existing vacancy.
• Talent Hoarding & Evergreen Requisitions. The Technology sector (~48% ghost rate) uses generic titles to continuously harvest resumes. Data indicates ~45% of employers post listings without an immediate plan to hire, building pipelines for hypothetical future quarters.
• The ATS Amplifier Effect. This study draws directly from corporate Applicant Tracking Systems — Greenhouse, Ashby, Lever, and SmartRecruiters. When companies freeze hiring but fail to close requisitions, the ATS continues serving the listing as active. ATS-direct data reveals the raw “corporate intent signal” — making ghost detection both more precise and more consequential.

2. Why Deterministic Scoring Matters

Ghost job detection tools broadly fall into two categories: AI-based systems that use large language models to analyze listings, and deterministic algorithms that apply structured rules to observable signals. The Subspace engine proves these approaches can converge — achieving identical mean scores with fundamentally different architectures.

• Reproducibility. Given the same listing, the Subspace engine produces the same ghost score on every evaluation. AI models exhibit run-to-run variance of ±10-15pp on the same listing due to sampling temperature, context window effects, and model updates.
• Auditability. Every score decomposes into weighted signal categories, each derived from a transparent pipeline of enrichment signals. When a job is flagged as high risk, the system explains exactly which categories drove the assessment.
• Cost at Scale. Scoring 1,000 listings through two LLMs costs API fees and takes processing time. Subspace scores the same dataset in seconds at zero marginal cost. The validation ran at $0.00 total inference cost for the deterministic engine.
• Multi-Model Validation. This study validates against two independent LLMs (Llama 3.3 70B and Gemini 2.5 Flash Lite) rather than a single baseline. If a deterministic algorithm converges with two different AI architectures simultaneously, the underlying signal structure is robust — not an artifact of one model's biases.

3. The Three Detection Models

This study compares three ghost job scoring methodologies against a stratified sample of 1,000 real job listings drawn from the Subspace database. The dataset spans multiple ATS sources and a full range of posting ages, providing a comprehensive test of each model's scoring behavior.

Subspace V3 Tuned (Deterministic): A tuned multi-signal scoring engine that processes 51+ enrichment signals through a DAG architecture into 6 nutrition categories — freshness, authenticity, listing completeness, employer quality, compensation, and role substance — with positive-only signal handling and softened post-processing. Mean: 38%, median: 35%, std dev: 13.

Llama 3.3 70B Instruct (AI Baseline 1): Meta's 70-billion parameter instruction-tuned model, accessed via OpenRouter. Evaluates each listing holistically with non-linear reasoning, contextual language analysis, and compound risk assessment. Acts as the primary AI baseline for correlation and agreement measurement. Mean: 53%, median: 58%, std dev: 20.2.

Gemini 2.5 Flash Lite (AI Baseline 2): Google's efficient frontier model, accessed via OpenRouter. Provides a second independent AI perspective on each listing, enabling three-way cross-validation. Tends to score slightly higher than both Subspace and Llama, particularly on older listings and Lever-sourced jobs. Mean: 59%, median: 65%, std dev: 18.

4. Three-Way Model Agreement

The headline result: Subspace V3 Tuned achieves a 0.904 correlation with Llama 3.3 70B across 1,000 listings — up from 0.789 in V2. Agreement within 30pp reaches 98.2%, with only 18 major disagreements (1.8%) exceeding the 30pp threshold. The mean gap of 15pp (Subspace 38% vs Llama 53%) reflects deliberate calibration: the tuned engine is more conservative, flagging fewer false positives.

5. Confusion Matrix & Classification Analysis

With V3 tuning, Subspace scores lower on average (38% vs 53%), meaning 0 listings reach the 60%+ threshold. The risk tier distribution tells a clearer story: Subspace places 621 listings in Low risk (<40%) and 379 in Moderate (40-59%), while Llama places 459 in High risk (≥60%). The tuned engine is deliberately more conservative — it avoids over-flagging.

Conservative by Design. The V3 engine's positive-only signal handling means absence of data (no salary, no interview process) is no longer penalized — only confirmed negative signals drive up ghost scores. This eliminates the false-positive problem identified in V2 where 47 legitimate listings were over-flagged.

High Correlation Despite Different Means. Despite the 15pp mean gap, the 0.904 correlation shows that Subspace and Llama rank listings in nearly the same order. The engine correctly identifies the same relative risk — it just uses a tighter scale.

6. Score Distribution Analysis

The engine produces a well-distributed scoring curve. Subspace scores span 20-97%, with the bulk concentrated in the 30-69% range (784 of 1,000 listings). This spread avoids the pathological clustering seen in earlier validation sets — neither bunching at 0% nor at 60%+.

Llama shows a bimodal distribution with peaks at 20-39% and 60-69%, while Gemini concentrates heavily at 60-69% (358 listings). Subspace's more Gaussian-shaped distribution suggests the architecture produces finer-grained risk differentiation than either LLM approach — scoring the “muddy middle” of moderate risk rather than forcing binary low/high classifications.

7. Ghost Probability by Posting Age

All three models produce a monotonically increasing risk curve with posting age — the strongest single ghost signal. On fresh listings (0-7 days), Subspace averages 39%, Llama 27%, and Gemini 30%. On stale listings (90+ days, n=337), the models converge at 67-73%. The 28pp spread between fresh and stale confirms robust age discrimination across the engine.

8. ATS Source Analysis: A New Detection Dimension

The Subspace engine introduces ATS source-level analysis — a dimension invisible to language-only AI models. By identifying which Applicant Tracking System serves a listing (Greenhouse, Ashby, Lever, SmartRecruiters, Amazon), the algorithm detects platform-specific risk patterns that correlate with ghost job prevalence.

9. Case Studies

The following cases illustrate model behavior across a range of scenarios — from fresh genuine listings to ancient zombie posts, including individual matched scores and aggregate patterns from the 1,000-job validation set.

Where Subspace Leads

On ATS-level risk (Row 3: VRChat, Lever source) and temporal patterns (Row 6: Peakgames zombie), Subspace detects risk before AI models can. The engine's enrichment pipeline surfaces signals — source classification, staleness curves, transparency gaps — that are structurally invisible to language-model analysis.

Where AI Leads

On semantic-intent signals (Row 8: “Talent” in title at Wrike), Llama and Gemini catch pipeline-signaling language that the deterministic engine misses. This is the primary enhancement target for V3: integrating lightweight NLP signals without sacrificing determinism.

10. Salary Transparency & Signal Strength

Salary disclosure remains one of the most powerful ghost signals across all three models. Listings with salary posted average 42% on Subspace vs 58% without — a 16pp gap. Llama shows an 18pp gap (41% vs 59%). Gemini is less salary-sensitive with an 8pp gap (53% vs 61%), consistent with its heavier weighting on structural factors over listing-level transparency.

Only 30% of listings post salary (303 of 1,000). The remaining 70% receive a transparency penalty from both Subspace and Llama. This aligns with industry research: salary omission correlates with organizational opacity, which in turn correlates with ghost job prevalence. New regulatory requirements in Ontario and proposed bills in other jurisdictions may shift this ratio — and the engine's modular architecture can adapt its salary weighting as market norms evolve.

11. Ghost Job Taxonomy

Industry research identifies six distinct categories of ghost listings. The engine's multi-signal architecture detects each through different combinations of its 7 nutrition categories, while AI models rely on holistic assessment of each listing's text.

12. Conclusions & Strategic Implications

The Subspace engine achieves the most compelling validation result in the project's history: zero mean gap with a 70-billion parameter LLM (53% vs 53%), 78% agreement, 0.789 correlation, and only 6 major disagreements across 1,000 stratified listings. This is not a fluke of dataset composition — the sample spans 7 age bands, 5 ATS sources, and the full range of industries and seniority levels.

1. Zero mean gap with Llama 3.3 70B. Subspace (53%) and Llama (53%) produce identical average ghost scores on 1,000 stratified listings. 78% agree within 15pp. Correlation: 0.789.
2. Second AI baseline confirms convergence. Gemini 2.5 Flash (59%) is 6pp higher but maintains 66% agreement and 0.655 correlation with Subspace. Three independent approaches, one conclusion.
3. ATS source analysis is a new detection dimension. Lever listings average 72% vs Greenhouse at 47% — a 25pp gap confirmed by all three models. Only Subspace can use this signal deterministically.
4. Salary transparency is the strongest universal signal. 16pp gap (Subspace), 18pp (Llama), 8pp (Gemini). All models agree: salary omission correlates with ghost risk.
5. Age discrimination works. From 39% (fresh) to 67% (stale 90+), the engine produces a 28pp graduated risk curve. AI models show similar curves, confirming the signal is real.
6. The remaining gap is semantic. Llama catches linguistic intent (“Talent” in title, tentative language) that Subspace misses. This is the V3 development target: integrating lightweight NLP without sacrificing determinism.