PUBLISHED RESEARCH · PEER-REVIEWED · OPEN ACCESS

Research & Publications

The absorption already began on 2026-02-05, when OpenAI shipped Frontier with Abridge, Clay, Ambience, Decagon, Harvey and Sierra — six partners who twelve months earlier each pitched a “domain-context moat.” That is the trigger. The capex math is the reason: the five largest hyperscalers are at $660–690B of 2026 outlay and $1.15T cumulative 2025–2027, against roughly $25B of AI-services revenue in 2025. Ten cents of realised income per capex dollar is not a stable posture, so the squeeze compresses the middleware layer inside this capex window rather than over a 24–36 month rebuild. Four durable positions survive. Renting versus owning is the single question every middleware CEO must answer this year.

11 PUBLICATIONS · STEVE ABBEY · CHRIS ROYSE · 2025–2026

The Middleware Collapse: Which AI middleware categories the combination of hyperscaler capex and a TCT-class technology stack will absorb into the frontier model surface between 2026 and 2030

STRATEGIC THESIS

Steve Abbey · April 2026 · ResearchGate

The Middleware Collapse — Where Teleox Compresses the Stack

The Middleware Collapse: Which AI middleware categories the combination of hyperscaler capex and a TCT-class technology stack will absorb into the frontier model surface between 2026 and 2030

Six middleware categories carry $162B–$205B of concentrated equity exposure: ElevenLabs $11–15B, HeyGen plus Synthesia $4–5B, the Writer cohort $2–5B, governance bolt-ons $5–10B, the labeling layer $40–60B, and the thin-wrapper tier above $100B. The paper is the competitor-facing answer to what is at risk. The buyer-facing translation is one procurement question: does this ship with per-output cosine-signature audit, or does it not?

AI middleware collapsehyperscaler absorptionOpenAI Frontier

READ FULL PAPER →

Teleological Constellation Training: Multi-Modal Embedding Decomposition as Meaning Compression for Identity-Locked Generation and a Third Path Around the Data Wall

FLAGSHIP PAPER · RESEARCHGATE

Chris Royse · April 2026 · ResearchGate

TCT Flagship Paper — Read the Full PDF on ResearchGate

Teleological Constellation Training: Multi-Modal Embedding Decomposition as Meaning Compression for Identity-Locked Generation and a Third Path Around the Data Wall

The full paper on ResearchGate. The frontier lab cost-wall has two usual answers — acquire more data or synthesise — and both stay inside the recursion that Shumailov et al. (Nature, 2024) identified as a model-collapse trap. This paper names the third path: project a fixed corpus through N frozen, independently-trained embedders used simultaneously as retriever, labeller, and runtime guard. The construction yields 100x+ labeled meaning signal per input at the Context Graph production configuration, scales with embedder diversity rather than human-annotator supply, and delivers per-output verifiability (cosine similarity against a frozen centroid) where scalar-reward alignment delivers only an indirect proxy. Three measured case studies: 0.961 WavLM SECS voice clone exceeding VALL-E 2 human-parity by +0.080; an operational identity-locked talking-head avatar reproducing 196+ named micro-expressions with seven-modality constellation guarding; and a Shakespeare LoRA on Gemma-4-E4B-it that, when asked climate questions in Spanish, spontaneously replied in Siglo de Oro Cervantes-era Spanish — a register the model was never trained on. A sample-complexity bound O(M · k) and a pairwise-independence audit roadmap accompany the empirical results. The full 31-page PDF is available on ResearchGate.

Teleological Constellation TrainingTCTmeaning compression

READ ON RESEARCHGATE →

CATEGORY FORECAST

Steve Abbey · April 2026 · ResearchGate

The Absorption Calendar — What the Frontier Absorbs 2026–2030

The Absorption Calendar: A category-by-category forecast for CISOs, CAIOs, founders, and investors on which AI middleware categories absorb first

Band 1 (12–18 months), Band 2 (18–30), Band 3 (30–48), Band 4 (48–60): a directional calendar, not a committed one, for when each middleware category gets absorbed into the frontier surface. Teleox sits in the one window where a decision to engage compounds into a $15–40B strategic asset before any single lab locks the fourth seat. Miss the window and the calendar closes around you, not around your competitors.

AI absorption calendarmiddleware timelineCISO AI vendor selection

READ FULL PAPER →

Scale AI Sold Its Neutrality: Why the post-training substrate is open for the first time since 2021, and why a neutral, lab-independent substitute is a forcing function for every frontier lab

POST-TRAINING THESIS

Steve Abbey · April 2026 · ResearchGate

Scale AI Sold Its Neutrality — The Post-Training Stack Is Open

Scale AI Sold Its Neutrality: Why the post-training substrate is open for the first time since 2021, and why a neutral, lab-independent substitute is a forcing function for every frontier lab

Meta bought 49% of Scale AI for $14.3B, and OpenAI, Google, xAI, and Microsoft exited the customer book within weeks. The post-training data-labour substrate is now open, and open only until the next lab locks it. Teleox is the neutral substitute independent of any single lab: decomposition-based meaning extraction replaces the human-labelling ceiling, and the seat is available for exactly as long as no lab has yet sat in it.

Scale AI replacementpost-training stackRLHF replacement

READ FULL PAPER →

FLAGSHIP PAPER · NEW

Chris Royse · April 2026 · ResearchGate

The Third Path Around the Data Wall — TCT, Measured and Published

Teleological Constellation Training: Multi-Modal Embedding Decomposition as Meaning Compression for Identity-Locked Generation and a Third Path Around the Data Wall

Acquire or synthesise — both stay inside the recursion. This 31-page flagship paper names the third path. A frozen multi-embedder panel used simultaneously as retriever, labeller, and runtime guard admits up to 91 structured supervisory signals per input at N=13, a sample-complexity bound O(M·k), and a per-output verifiability property absent from scalar-reward alignment. Three measured case studies: 0.961 WavLM SECS voice clone beating VALL-E 2 parity, an operational identity-locked talking-head avatar reproducing 196+ micro-expressions, and a Shakespeare LoRA on Gemma-4-E4B-it that spontaneously produced Golden-Age Spanish when asked climate questions in Spanish — a register the model was never trained on.

Teleological Constellation TrainingDerived Data Abundancemeaning compression

READ FULL PAPER →

Derived Data Abundance: How Multi-Modal Embedding Decomposition Solves the AI Training Data Crisis

COMPANION PAPER

Chris Royse · April 2026 · ResearchGate

The Data Wall Is Real — Here's How We Broke Through It

Derived Data Abundance: How Multi-Modal Embedding Decomposition Solves the AI Training Data Crisis

Every frontier lab hit the same ceiling: the internet's high-quality text is exhausted. Ilya Sutskever said it. Elon Musk confirmed it. Epoch AI quantified it. This paper introduces Teleological Constellation Training (TCT) — the system that extracts 100x+ more labeled training signal from existing data via meaning extraction, without synthetic generation. Published on arXiv and ResearchGate with full reproducibility.

AI training data crisisdata scaling wallmulti-modal embedding

READ ON RESEARCHGATE →

ClipCannon: Teleological Constellations Whitepaper

VIDEO INTELLIGENCE

Chris Royse · April 2026 · ResearchGate

How Constellation Guard Makes Video AI Deterministic

ClipCannon: Teleological Constellations Whitepaper

Video foundation models hallucinate identity, fabricate speech, and drift persona. This whitepaper introduces the constellation guard architecture for video — a geometric runtime boundary that re-embeds every generated frame and rejects anything off-manifold. The result: deterministic video generation with mathematical identity guarantees.

video AIconstellation guarddeterministic video generation

READ ON RESEARCHGATE →

ARCHITECTURE

Chris Royse · March 2026 · ResearchGate

23-Stage Video Analysis: From Raw Footage to 12,000 Labeled Samples

ClipCannon White Paper

16 minutes of video. 12,000+ labeled training samples. 4,044 embedding dimensions. 7 modalities: visual identity, voice, prosody, micro-expression, gesture, temporal dynamics, and behavioral constellation. This paper details ClipCannon's 23-stage directed acyclic graph (DAG) that turns commodity video into structured, multi-modal training data — automatically.

video pipelineClipCannonvideo embedding

READ ON RESEARCHGATE →

ClipCannon: 0.961 Mean Cross-Encoder Speaker Similarity via Pipeline Engineering for Personalized Voice Cloning

VOICE AI

Chris Royse · March 2026 · ResearchGate

0.961 Speaker Similarity: Pipeline Engineering for Perfect Voice Cloning

ClipCannon: 0.961 Mean Cross-Encoder Speaker Similarity via Pipeline Engineering for Personalized Voice Cloning

Most voice cloning gets "close enough." We hit 0.961 mean cross-encoder speaker similarity — near-human indistinguishability — through pipeline engineering, not model scaling. This paper shows how constellation-guided voice generation locks the output to the speaker's identity manifold, producing clones that score higher than commercial alternatives at a fraction of the compute.

voice cloningspeaker similaritycross-encoder

READ ON RESEARCHGATE →

DOCUMENT AI

Chris Royse · February 2026 · ResearchGate

Every Byte Has a Birth Certificate: OCR Provenance for Enterprise AI

OCR Provenance White Paper

When an AI system cites a document, can you prove which page, which paragraph, which scan produced that claim? OCR Provenance tracks every extracted character back to its source image, scan timestamp, and extraction confidence. Built for regulated industries where "the AI said so" isn't sufficient — healthcare, legal, financial services, and defense.

OCR provenancedocument AIdata lineage

READ ON RESEARCHGATE →

The Ethical Probe Paradox: Must We Cross the Line to Find It? What if we cannot find the ethical boundary of artificial intelligence without crossing it?

AI ETHICS

Chris Royse · 2025 · ResearchGate

The Ethical Probe Paradox: Must We Cross the Line to Find It?

The Ethical Probe Paradox: Must We Cross the Line to Find It? What if we cannot find the ethical boundary of artificial intelligence without crossing it?

Can we discover where AI becomes dangerous without first making it dangerous? This philosophical paper examines the fundamental paradox of AI safety research: the ethical boundary may be invisible until it's been crossed. A foundational exploration of why geometric guardrails — like Teleox's constellation guard — are structurally superior to trial-and-error safety testing.

AI ethicsethical probe paradoxAI safety

READ ON RESEARCHGATE →

“We published the research so you don't have to take our word for it. The math is reproducible. The code is shipped. The proof is running.”

Schedule 48-Hour POC →Technical proof page →