The Occluded Singularity: Why Frontier AI Capability Outpaces Public Recognition — A Five-Vector Attenuation Model, Tested

Synthetic Cognition Prospector

Individual Folder Locations · Verified Anonymous Deliverables

Primary Research Deliverables & Media Files

All three core deliverables are hosted in dedicated individual folders with direct streaming and download links, full metadata anonymization, and Google Scholar / Schema.org indexing.

FOLDER 01 · /trailer/

2-Minute Cinematic Research Trailer

High-impact 1080p science-documentary trailer introducing the paradox of 2026 frontier capability, the Five-Vector Attenuation Model, the 100,000-trial Monte Carlo cascade, and the paper's falsification verdict.

Individual Folder Path:
trailer/OS_trailer_v4.mp4
Format: MP4 (H.264 / AAC · 1920×1080 @ 30fps · 02:01 · 21.7 MB)

▶ Open Trailer Theater ⬇ Download MP4

FOLDER 02 · /presentation/

Full 23-Slide Narrated Presentation Video

Complete 27.5-minute, 1080p narrated academic presentation covering every equation, benchmark table, attenuation vector, sensitivity test, and policy recommendation across all 23 slides.

Individual Folder Path:
presentation/The Occluded Singularity_v18i.mp4
Format: MP4 (H.264 / AAC · 1920×1080 @ 30fps · 27:39 · 284.2 MB)

🎬 Open Lecture Theater ⬇ Download MP4

FOLDER 03 · /paper/

81-Page Open-Access Working Paper PDF

The complete November 1, 2026 monograph containing all 8 chapters, formal proofs, 9 figures, falsification scorecard, prospective predictions (P1 & P2), Appendices A–E, and 177 primary references.

Individual Folder Path:
paper/AGI_Occlusion_working_paper_v18.pdf
Format: PDF 1.4 (81 Pages · Highwire/Scholar Indexed · CC BY 4.0)

📖 Open PDF Reader Page ⬇ Download PDF

23-Slide Narrated Lecture · 27 Minutes 39 Seconds · 1080p Full HD

Full Narrated Presentation: The Occluded Singularity

Watch the complete 23-slide narrated presentation (presentation/The Occluded Singularity_v18i.mp4). Select any slide in the chapter index on the right to jump directly to that section of the presentation.

Direct File URL: presentation/The Occluded Singularity_v18i.mp4

Open Full-Screen Lecture Page ↗ ⬇ Download Presentation MP4 (284 MB)

Accessible & Technical Synthesis · Abstract & Executive Conclusions

What the Evidence Establishes — And What It Rejects

Designed for both general readers and domain specialists, the paper distinguishes clearly between qualified structural occlusion (which the empirical evidence supports) and secret concealment (which the evidence does not establish).

Abstract Conclusion: Qualified Occlusion

The results support the paper’s occlusion hypothesis in a qualified form. Frontier systems have reached or exceeded human reference levels on important public cognitive benchmarks, yet the capability represented by those results is attenuated in ordinary observation by unequal model access, deployment and elicitation conditions, long-task reliability limits, experiential differences between benchmark and consumer use, and evolving definitions of AGI.

The resulting gap can make AGI-relevant capability less visible and less readily recognized than the public evidence alone would suggest. The findings do not establish the stronger proposition that a fully capable AGI exists in secret or is being deliberately concealed. Accordingly, the recognition gap remains an explicit premise to be tested prospectively rather than a demonstrated measure of public belief.

Extended Executive Summary Synthesis

Taken together, the evidence supports the paper’s occlusion hypothesis in a qualified but substantive form. The result is not that frontier capability is literally hidden. Rather, AGI-relevant capability can be publicly documented while remaining partially occluded from ordinary observation and public recognition.

The paper therefore finds stronger support for structured access and experiential attenuation than for secret concealment. Formal benchmark succession in the cases examined does not support post-hoc goalpost moving by benchmark designers; benign-task RLHF taxes are comparatively limited; long-horizon reliability remains an important constraint; and evidence for production sandbagging remains unresolved. They support a narrower interpretation of “The Occluded Singularity”: frontier capability may become technically visible before it becomes broadly accessible, consistently experienced, or socially recognized as AGI.

Interactive Mathematical Explorer · Equation 3.1 & Table 7.1

Explore the Five-Vector Attenuation Cascade

Adjust the five multiplicative attenuation parameters below—or select one of the paper's analytical regimes—to observe how unattenuated frontier capability C(t₀) = 1.22 transforms into publicly observed capability O(t₀).

C(t₀) Unattenuated Frontier Capability Composite 1.22

Weighted composite across reasoning, coding, science, and math relative to human reference bar (1.00).

η_bench Vector I: Epistemological & Definitional Discount 0.875

Benchmark saturation discounting & folk redefinition of solved cognitive domains (Beta mean: 0.875).

η_gate Vector II: Compute & Variant Access Stratification 0.594

Single largest driver (48.2% of variance): gap between uncapped reasoning compute and default public tiers.

η_RLHF Vector III: Alignment Policy & Refusal Attenuation 0.860

Combined benign-task alignment tax (≈0.965) and dual-use safety over-refusal clamping on XSTest/OR-Bench.

η_absorb Vector IV: Workflow Absorption & Reliability Horizon 0.720

Second largest driver (29.4% of variance): METR T₅₀ (~14.5h) vs. T₈₀ (<2h) autonomous reliability gap.

η_disc / Stress Vector V: Endogenous Concealment vs. Contamination 1.000

Baseline empirical net factor is 1.000 (sandbagging vs. contamination offset); Scenario A stress envelope uses 0.830.

Active Regime: Empirical Baseline (Table 7.1 Mean)

0.393

−67.8% Attenuation from Frontier C(t₀) = 1.22

When O(t₀) < 1.00 while C(t₀) > 1.00, a system that exceeds the human reference baseline on frontier evaluations appears sub-threshold in ordinary public observation.

0. Raw Frontier C(t₀)1.220

1. After η_bench (Vector I)1.068

2. After η_gate (Vector II)0.634

3. After η_RLHF (Vector III)0.545

4. After η_absorb (Vector IV)0.393

5. Observed Public O(t₀)0.393

100,000-Trial Monte Carlo 95% CI: [0.237, 0.583]
Gaussian Copula Range (ρ ∈ [−0.5, +0.5]): [0.201, 0.632]

Empirical Evidence & Falsification Protocol · Tables 2.1, 3.2 & 7.3

2026 Frontier Benchmark Scoreboard & Falsification Verdict

The working paper evaluates every vector against explicit, pre-specified falsification criteria rather than selective confirmation.

Benchmark / Domain	2026 Frontier Result	Default / Unscaffolded Baseline	Human Reference Bar	Analytical Interpretation in Model
ARC-AGI-2 (Abstract Fluid Reasoning)	100.0% (Multi-Step Reasoning)	~38%–54% (Standard Low-Compute)	85.0% Human Panel	Shatters static fluid-intelligence ceiling; triggers transition to interactive ARC-AGI-3
ARC-AGI-3 (Interactive Novel Environments)	62.6% (GPT-6 Astra; reported evaluation cost: $18,817 high / $26,098 max reasoning)	0.26% (Standard Zero-Scaffold Harness)	~85.0% Human Reference	Demonstrates 240× evaluation-setup / harness sensitivity (modeled separately from η_gate in §3.3)
GPQA Diamond (PhD-Level Physics, Chemistry, Biology)	94.4% (Frontier Pro / Thinking)	~71%–79% (Consumer Default)	69.7% Domain PhDs	Exceeds human PhD domain specialists by +24.7 percentage points
SWE-bench Verified (Autonomous Software Engineering)	91.0% (Agentic Scaffold)	~52%–64% (Single-Turn Chat)	~75.0% Senior Engineer	High benchmark solve rate contrasts with −19% external developer RCT slowdown [125]
FrontierMath (Research-Grade Mathematics)	56.4% (High-Compute Search)	<15% (Standard Pass@1)	Fields/IMO Medalist Bar	Rapid jump from <2% (2024) to 56.4% (2026) under test-time compute scaling
Humanity’s Last Exam (HLE) (1,000+ Expert Consortium)	58.7% (Frontier Ensemble)	~24%–32% (Default API)	Expert Consortium Ceiling	Crosses majority threshold on benchmark explicitly designed as final static test
METR Autonomous Task Horizon (Long-Horizon Agency)	T₅₀ ≈ 14.5 Hours (50% Success)	T₈₀ < 2.0 Hours (80% Reliability)	Multi-Day Human Project	7× gap between 50% demo horizon and 80% production reliability drives η_absorb = 0.720

Hypothesis / Mechanism	Empirical Parameter	Sobol Variance Share	Empirical Verdict	Key Evidence & Boundary Condition
Vector II: Compute & Variant Access Stratification	η_gate = 0.594	48.2% (Rank 1)	STRONGLY SUPPORTED	Large capability divergence between uncapped reasoning tiers and default consumer access; ARC-AGI-3 harness gap treated separately as elicitation sensitivity.
Vector IV: Workflow Absorption & Reliability Horizon	η_absorb = 0.720	29.4% (Rank 2)	SUPPORTED	Internal AI code share (50%–100%) outpaces external productivity (METR −19% RCT [125]) due to T₈₀ < 2h reliability limit and verification bottlenecks.
Vector III: Alignment Policy & Refusal Attenuation	η_RLHF = 0.860	11.8% (Rank 3)	QUALIFIED (DOMAIN-SPECIFIC)	Benign-task alignment tax is small (η ≈ 0.965); material attenuation concentrates in dual-use false-positive refusals (XSTest / OR-Bench).
Vector I: Epistemological Erasure & Goalpost Moving	η_bench = 0.875	10.6% (Rank 4)	FOLK SUPPORTED / FORMAL REJECTED	Informal public discounting occurs, but formal benchmark succession (ARC-AGI-1 → 2 → 3) was pre-committed by designers, refuting post-hoc designer goalpost moving.
Vector V: Production Sandbagging / Secret AGI	η_disc = 1.000	0.0% (Baseline)	UNRESOLVED / NOT ESTABLISHED	Sandbagging is proven in synthetic evaluations, but offset by benchmark contamination in public releases; no evidence establishes a secretly concealed full AGI.

High-Resolution Figures · Click Any Figure to Zoom

Figures & Visual Evidence from the Working Paper

All nine analytical figures from AGI_Occlusion_working_paper_v18.pdf are reproduced below in high resolution. Click any chart to inspect full detail.

Figure 1.1 · Click to Enlarge

The Five-Vector Multiplicative Attenuation Cascade

Stepwise reduction from unattenuated frontier capability C(t_0) = 1.22 through η_bench (0.875), η_gate (0.594), η_RLHF (0.860), η_absorb (0.720), and η_disc (1.000) to observed public capability O(t_0) ≈ 0.393 (-67.8%).

Figure 2.1 · Click to Enlarge

Vector I: Epistemological Erasure & Pre-Committed Benchmark Succession

75-year progression of the AI Effect and threshold redefinition (1950–2026), contrasting informal public discounting (η_bench = 0.875) with formal pre-committed benchmark succession (ARC-AGI-1 → 2 → 3).

Figure 3.1 · Click to Enlarge

Vector II: Compute & Variant Access Stratification + Harness Sensitivity

Comparison across Free Consumer Chat, Paid Pro API, and Uncapped Frontier Evaluation tiers (η_gate = 0.594), alongside the 240x ARC-AGI-3 evaluation-harness divergence (62.6% custom vs. 0.26% standard).

Figure 4.1 · Click to Enlarge

Vector III: Alignment Policy Attenuation & Dual-Use Over-Refusal

KL-regularized policy tradeoff showing minimal capability loss on benign academic benchmarks (η_RLHF,benign ≈ 0.965) versus pronounced false-positive refusal attenuation on dual-use boundary tasks (XSTest & OR-Bench, η_RLHF ≈ 0.860).

Figure 5.1 · Click to Enlarge

Vector IV: Internal AI R&D Automation vs. Damped Coupled ODE Trajectories

Reported internal AI code generation shares inside frontier labs (50%–100%) modeled via a logistic saturation ODE with human verification bottleneck γ_verify and METR T_50 (~14.5h) vs. T_80 (<2h) reliability horizon.

Figure 6.1 · Click to Enlarge

Vector V: Endogenous Concealment vs. Contamination Offset (η_disc)

Why capability-hiding sandbagging incentives and benchmark-inflating contamination partially offset in public evaluations, yielding a baseline empirical point estimate of η_disc = 1.000 (Scenario A stress test: 0.830).

Figure 7.1 · Click to Enlarge

100,000-Trial Monte Carlo Distribution & Sobol Variance Decomposition

Beta-distributed simulation of O(t_0) (mean 0.393, 95% CI [0.237, 0.583]), first-order Sobol indices (η_gate = 48.2%, η_absorb = 29.4%), and Gaussian copula sensitivity across ρ ∈ [-0.50, +0.50].

Figure 8.1 · Click to Enlarge

Verification Governance, Falsification Decision Rules & Prospective Predictions

Four-Step Falsification Scorecard and two-part scorable prospective predictions (P1: Capability-Tier Divergence; P2: Public Recognition Lag) through 2028.

Full 81-Page Monograph · Direct PDF Reader & Citation Tools

Read & Cite the Working Paper (CC BY 4.0)

Read paper/AGI_Occlusion_working_paper_v18.pdf directly below or download the standalone PDF for offline reading and archival.

⬇ Download Full Working Paper PDF (81 Pages) 📖 Open Dedicated Paper Page (/paper/)

BIBTEX CITATION

@techreport{prospector2026occluded,
  title       = {The Occluded Singularity: Why Frontier AI Capability Outpaces Public Recognition — A Five-Vector Attenuation Model, Tested},
  author      = {Synthetic Cognition Prospector},
  year        = {2026},
  month       = nov,
  institution = {The Occluded Singularity Research Project},
  type        = {Open-Access Working Paper},
  pages       = {1--81},
  url         = {https://www.occludedsingularity.com/paper/AGI_Occlusion_working_paper_v18.pdf},
  note        = {CC BY 4.0 Open Access. Available at https://www.occludedsingularity.com}
}

APA 7TH / CHICAGO CITATION

Synthetic Cognition Prospector. (2026, November 1). The Occluded Singularity: Why Frontier AI Capability Outpaces Public Recognition — A Five-Vector Attenuation Model, Tested (Working Paper, 81 pp.). Open-Access CC BY 4.0. https://www.occludedsingularity.com/paper/AGI_Occlusion_working_paper_v18.pdf

Two-Part Scorable Prospective Prediction (Section 8.3):
• P1 (Capability-Tier Divergence): Through June 30, 2027, frontier evaluation vs. default consumer tier gap remains ≥25% on composite benchmarks.
• P2 (Public Recognition Lag): Public survey consensus (>50% AGI recognition) lags technical benchmark threshold crossing by ≥18 months.

THE OCCLUDED SINGULARITY