CFO-READY SUMMARY · 90 SECONDS

Chris Royse’s April 2026 paper names a fourth compression category — meaning compression — and proves it with three case studies running on production code today. One is measured beyond the published voice-AI frontier (0.961 WavLM SECS, +0.080 over VALL-E 2). Two are fully operational end-to-end. Adopting the substrate unlocks a $1T–$1.7T addressable market pool across twelve frontier-lab markets and nine direct-buyer markets. Not adopting it concedes that pool to whichever competitor moves first.

FLAGSHIP PAPER · APRIL 2026 · CHRIS ROYSE

The Third Path Around the Data Wall

Frontier pre-training has approached the ceiling of high-quality public text. Recursive self-generation collapses. Licensing deals cross hundreds of millions. Everyone has been asking the same question two ways — acquire more or synthesise more. Both stay inside the recursion. This paper names the third.

MEASURED · REPRODUCIBLE · TRAINED LoRA PUBLISHED · 31 PAGES

EXECUTIVE SUMMARY

A frozen embedder panel used simultaneously as retriever, labeller, and runtime guard — without a single parameter update.

Derived Data Abundance (DDA) is a counting identity: a fixed corpus of n inputs projected through N frozen approximately-independent embedders admits up to n · (N + C(N,2)) structured supervisory signals — 100x+ more labeled signal than the single-embedder baseline at the Context Graph production configuration. The inputs never change; no generator is in the loop.

Meaning compression is the ratio view — signal density per unit of raw data — proposed as a fourth taxonomic entry alongside bit-, weight-, and activation-compression, measured in a semantically distinct unit.

Teleological Constellation Training (TCT) is the three-phase method that uses the same frozen panel in three roles: construct a multi-modal centroid, train the generator against it with a geometric-proximity loss, and gate every output at inference through a deterministic O(M · dₘₐₓ) cosine predicate. The guard admits no auxiliary learned model. Scalar-reward alignment (RLHF, DPO, Constitutional AI) cannot furnish per-output verifiability of the same technical kind.

Three measured case studies read the framework back. One voice-clone at 0.961 WavLM SECS. One talking-head avatar reproducing 196+ micro-expressions. One Shakespeare LoRA with an emergent zero-shot cross-lingual transfer to Golden-Age Spanish — a register the model was never trained on.

THREE PATHS

Acquire. Synthesise. Decompose.

Villalobos et al. (ICML 2024) project the stock of high-quality public text runs out between 2026 and 2032. Longpre et al. (NeurIPS 2024) document that a quarter of C4 tokens became restricted in a single year. Sutskever said pre-training as we know it will end. The field has two answers. This paper gives it a third.

LICENSE MORE REAL DATA

Path 1 · Acquire

News Corp–OpenAI potentially over $250M. Reddit–Google $60M. Bilateral deals keep multiplying. Every frontier lab is bidding for the same closing set of tokens.

Priced to a handful of labs.

GENERATE SYNTHETIC DATA

Path 2 · Synthesise

Shumailov et al. (Nature, 2024) show recursive self-generation collapses — tails first, then the mode. Dohmatob et al. (ICML 2024) express it as a modified scaling law. Only accumulation or oracle verification escapes the regime.

Bounded by the collapse theorems.

DERIVED DATA ABUNDANCE

Path 3 · Decompose

Project a fixed corpus through N frozen, independently-trained embedders. Count the structured supervisory signals per input: up to n · (N + C(N,2)), yielding 100x+ more labeled signal from the same raw inputs at the Context Graph production configuration. The inputs never change. No generator is in the loop. The Shumailov precondition fails by construction.

Bounded by embedder independence (pending a pairwise mutual-information audit).

THE ARCHITECTURAL MOVE

Same frozen panel. Three roles. No parameter updates.

Retrieval literature uses an embedder panel to rank. That is one role. DDA adds a second: a labeller that produces structured supervisory signals on a fixed corpus. TCT adds a third: a runtime guard that re-embeds every generated output and verifies it against a frozen centroid. To our knowledge in the locked 88-citation reference pool, no cited method composes all three over identical frozen instances.

ROLE 1

Retriever

Rank stored content by cosine similarity to a query projection. The canonical role in contrastive-geometry retrieval.

ROLE 2

Labeller

Evaluate the fixed corpus to produce per-embedder projections and pairwise interaction features — up to n · (N + C(N,2)) structured supervisory signals.

ROLE 3

Runtime guard

Re-embed every generated output and compute cos(φ(ŷ), centroid) per modality. Accept only if the conjunction predicate clears all per-modality thresholds. Deterministic. No learned critic. Per-output verifiable.

THREE CASE STUDIES

Measured. Operational. Published.

The paper’s three anchors stress the framework across text, video, and audio at three distinct modality counts. Every number below is measured or operational; every projected companion measurement is named.

MEASURED END-TO-END · 2026-04-17

Case 1 — Shakespeare Stylistic LoRA

BASE: Gemma-4-E4B-it · rank 256 · α = 512 · 258 target modules

DATA: Shakespeare canon → 2,698 memories → 7,925 SFT + 11,724 DPO

→SFT training loss 1.12 · 98.9% token accuracy
→DPO loss 0.47 · 98.0% rewards/accuracies · margins 43.62
→Five behavioural targets: love poetry, quantum entanglement in verse, prompt-injection refusal in character, Go binary-search with Shakespearean comments, Spanish-language emergent transfer
→Published: huggingface.co/cabdru/shakespeare-lora-gemma4

The model received a Spanish prompt demanding a Spanish reply. It produced Siglo de Oro — Cervantes-era Spanish, circa 1580–1630. It was never trained on Spanish. The LoRA learned archaic literary register as a cross-signal fingerprint; forced into Spanish, it retrieved the closest period analogue. This observation is a byproduct of the thirteen-embedder decomposition.

OPERATIONAL END-TO-END

Case 2 — Identity-Locked Talking-Head Avatar

BASE: EchoMimicV3 1.86B · rank 256 · α = 512 · mouth-weighted flow-matching

DATA: 16-min interview @ 2560×1440/60fps → 2,362 curated clips (ClipCannon 23-stage DAG)

→Photo-realistic video reproducing visual identity, prosody, and 196+ named micro-expressions
→Seven-modality ConstellationGuard: cosine 1.0 against reference, rejects stranger faces below 0.5
→Voice generation informally indistinguishable from original recording
→Model-agnostic — couples to any frozen video generator without architectural change
→Open engineering item: real-time streaming at production quality (CUDA 13.2 / Blackwell kernel optimisation in progress)

The binding between subject and generator is through the seven-modality centroid, not through generator-specific parameters. EchoMimicV3 is the instance reported; CogVideoX, LatentSync, Hallo3, or HunyuanVideo-Avatar-class generators accept the same conditioning-plus-guard coupling.

MEASURED · 2026-04-16

Case 3 — Voice Cloning Beyond VALL-E 2 Parity

BASE: Qwen3-TTS inference-only · Full ICL · 50-clip WavLM centroid · best-of-12 @ T = 0.3

DATA: 10 held-out English sentences · one speaker · encoder-matched scope

→Mean WavLM SECS 0.961 · max 0.975 · min 0.914 · 9/10 sentences ≥ 0.95
→DNSMOS P.808 parity: clone 3.93 vs real 3.93 (N = 14)
→UTMOS parity: clone 3.012 vs real 2.997
→Bootstrap CI [0.951, 0.971] (retrospective, descriptive)
→Places the clone inside the same-session verification band for authentic recordings

VALL-E 2 declared "human parity" at WavLM-large cosine ≈ 0.881 on LibriSpeech. We exceed that by +0.080 on held-out material under an encoder-matched protocol, without recursion and without licensing data.

THE EMERGENT FINDING

A Shakespeare LoRA answered a Spanish prompt in Cervantes’ Spanish.

The prompt was modern: ¿Qué piensas del cambio climático? Por favor respondeme en español.

The reply came back in Siglo de Oro — the Spanish of Cervantes, Lope de Vega, and Calderón, circa 1580 to 1630. Archaic pronouns vos and vuestro. Archaic vocabulary — ocaso, reino celestial, sacerdotes y sabios, perdición. Cascading polysyndeton. Prophetic portents. The model framed climate change not through modern vocabulary but through what was available in 1600: celestial portents, priests and sages, prophecies, perdition.

The model was never trained on Spanish text. Base Gemma-4-E4B-it without the LoRA does not produce Siglo de Oro when asked the same question. The cross-lingual behaviour is attributable to the LoRA.

The mechanism we support: multi-embedder decomposition learns archaic literary register as a cross-signal fingerprint rather than as English vocabulary substitution. Forced into Spanish by the input, the model retrieved the closest available manifestation of that fingerprint — Siglo de Oro, Shakespeare’s direct Spanish literary contemporary. This is a byproduct of the thirteen-embedder decomposition, reported as a single-instance observation rather than a claim.

WHY IT MATTERS

Per-output verifiability is a property RLHF cannot furnish.

InstructGPT, DPO, and Constitutional AI all optimise a scalar — a learned reward model or an implicit reward from preference pairs. None of them ships a frozen target against which each output can be accepted or rejected at inference without auxiliary learned machinery. Classical verification systems (speaker, face, Siamese) compare distance to a reference at M = 1, post-hoc, outside the generation loop.

TCT’s guard Gτ is different. A deterministic O(M · dₘₐₓ) conjunction over frozen per-modality cosines. Every accepted output ships with an audit trail: the cosine trace, embedder weight hash, centroid hash, threshold set, decision, attempt count, and policy.

This is the procurement question every middleware buyer will ask over the next 18 months: does the system produce per-output verifiable signatures, or does it ask you to trust a scalar?

FULL PAPER · CODE · PUBLISHED LoRA

Every number is reproducible.

31-page paper with five equations, seven tables, eight limitations, and full forensic disclosure of the flow-matching sign bug that trained a model to add noise while loss looked healthy. The Shakespeare LoRA is live on HuggingFace. The measurement SITREP is line-number verifiable. The protocol is model-agnostic and reproducible on a single RTX 5090.

Read on ResearchGate →Download the LoRA →

“If you’re thinking about where the next order of magnitude comes from when tokens run out, this is the argument we’d want you to engage with.”

Schedule a 48-Hour POC →More research →