Beyond the Data Wall: How Orchestrated Synthetic Data Is Quietly Rewriting the Future of AI

The Problem Nobody Wants to Say Out Loud

The AI industry is running out of data. Not slowly, and not theoretically. The well that fed the first generation of large language models, that sprawling ocean of open web text, licensed corpora, and scraped documents, is drying up at the exact moment the field needs more specialized, higher quality training material than ever before. Specialized domains like medicine, finance, law, and public safety have never had enough publicly available data to begin with, and what little exists is wrapped in privacy constraints, legal restrictions, and cost structures that make human annotation economically painful.

Into that gap walks a new paper from Google and EPFL, published at the SynthData workshop at ICLR 2025, titled Orchestrating Synthetic Data with Reasoning. The authors introduce a framework called Simula, and while the academic community will rightly dissect the technical contributions, the strategic implication is much larger. Simula is one of the first public blueprints for how the next decade of AI training is actually going to work. At Detailed In Design, we have been building a parallel process for some time, and seeing Google publish openly on the same territory is a welcome validation of a direction we have already pushed further.

This newsletter walks through what Simula actually does, where our own methodology at detailedindesign.com diverges and extends it, and why the combination of these approaches matters for law enforcement, legal practice, manufacturing, fraud detection, and the post-transformer architectures that are already beginning to replace today's models.

What Google's Simula Actually Does

The elegance of Simula sits in a single observation: you do not need seed data from the target distribution if you can reason your way to global coverage instead. Most synthetic data pipelines today rely on either elaborate hand-crafted prompts, evolutionary algorithms, or large banks of existing examples to bootstrap generation. Each of those approaches leaks bias, limits scalability, or simply fails when no seed data exists, which is the usual case in high-consequence domains.

Simula replaces the seed-data assumption with a taxonomic approach. Given a description of the dataset you want, a multi-modal model (what the authors call an M3) proposes the prime factors of variation, expands those factors into full taxonomies in a breadth-first fashion, samples combinations of taxonomy nodes to create meta-prompts, and then generates and critiques data points against those prompts. A double-critic rejection sampling step at the end filters outputs for both correctness and complexity.

The reported results are strong. On multilingual MMLU subsets, fine-tuning Gemma 2B on Simula-generated data closed the gap with instruction-tuned baselines and in several cases exceeded them. Taxonomy completeness reached 0.74 against human expert taxonomies, versus 0.52 for naive zero-shot expansion. Critic refinement reliably lifted the accuracy of generated answers on MATH problems across all five difficulty levels. The underlying assumptions, that large models can generate sound taxonomies, act as effective critics of their own output, and distinguish complexity, all held up under controlled testing.

That is a real contribution, and it deserves to be read widely. The public PDF is available through the ICLR 2025 workshop proceedings for anyone who wants to dig into the experimental setup and the appendices.

Where Detailed In Design Pushes the Process Further

Our process at Detailed In Design shares the spine of Simula's idea, which is that synthetic data generation should be explainable, controllable, and built on reasoning rather than on pure stochastic expansion. The divergence starts at step one.

Google's framework builds taxonomies from the factors a single M3 proposes, then expands those factors along the tree structure the model considers most natural. That produces strong coverage for the paths the teacher model already knows how to walk. What it does not produce, and what the paper openly acknowledges as a risk in Figure 1.c, is coverage of the nodes the model silently drops along the way. Progressive coverage loss is the polite term for a blind spot.

We filter our data through a substantially wider set of paths than any single teacher model would volunteer. Our pipeline runs multiple divergent taxonomy generators against the same target description, cross-validates their outputs against domain-specific ontologies we have curated in forensics, legal, financial, and manufacturing contexts, and then deliberately injects edge-case branches that the base models consistently under-represent. The result is a dataset that carries two properties a single-pipeline approach struggles to reach at the same time: genuine diversity at the global level, and adversarial complexity at the local level.

Concretely, that means our synthetic corpora contain the everyday cases a law enforcement analyst or a compliance officer sees hundreds of times a week, and they contain the rare, ambiguous, legally hazardous cases that historically show up once every few years and that are precisely the ones a model trained on thin or biased data gets catastrophically wrong. A random simulated dataset cannot do that. Even a well-engineered Simula run will tend toward the center of the distribution unless you force the tails open from the outside, which is exactly what our filtering layer is built to do.

None of this is a critique of Google's paper. Simula is excellent work and the quality of the datasets we have previewed from the framework is very strong. What we are pointing at is that the synthetic data frontier is going to be defined by stacking techniques, not by picking one.

Why This Matters for Post-Transformer Training

Transformers train long, consume enormous amounts of data once, and carry the weight of everything they have ever seen whether it is relevant to the current task or not. That model is starting to show its ceiling, both in compute cost and in the data scarcity curve that researchers like Villalobos and colleagues have been charting for two years now.

The post-transformer architectures now moving from research into production operate very differently. They train in short, targeted bursts, prune aggressively based on the shortest useful routes through the data, and rely on high-signal examples rather than high-volume scraping. That shift changes what good training data actually looks like. Quantity matters less. Coverage, complexity calibration, and per-sample reasoning density matter far more.

This is precisely the regime where orchestrated synthetic data, built with the Simula-style taxonomic skeleton and extended with the multi-path filtering we use at Detailed In Design, stops being an incremental improvement and starts becoming the foundation of the training pipeline. A model that trains in short bursts and then prunes based on quickest paths is a model that will be shaped, for better or worse, by whichever examples survived the prune. If those surviving examples came from a narrow synthetic pipeline, the resulting system inherits the pipeline's blind spots. If they came from a deliberately wide, complexity-balanced, explainable synthetic process, the resulting system gets stronger faster and in less time than anyone building on 2023-era assumptions would predict.

Scenarios Where This Changes the Outcome

Law Enforcement and Investigative Work

This is the scenario our founder John McClain has lived inside for twenty years. AI tools currently deployed in law enforcement contexts are trained on datasets that are too broad, too biased toward conviction-weighted historical records, and too thin on the edge cases that separate a valid lead from a catastrophic false positive. Innocent people have been arrested because a facial recognition system or a risk scoring tool confidently produced the wrong answer from a dataset that never contained anyone who looked, spoke, or behaved like them in the relevant way.

Orchestrated synthetic data, filtered through multiple divergent paths and forced to cover adversarial edge cases, directly attacks that failure mode. A dataset that contains the one-in-ten-thousand scenario, rendered with the same fidelity as the common case, produces models that hesitate in the right places instead of confidently misfiring. Our OpenIntel product for OSINT work and our SolaceSentry flagship for high-consequence inference are built on exactly this premise: a real case solved is worth infinitely more than ten false responses that put pressure on the wrong person.

Legal Practice

Attorneys working on complex litigation do not need an AI that has read the whole internet. They need an AI that has reasoned carefully over the narrow slice of case law, statutory history, and procedural edge cases relevant to the matter on their desk. Attorney Minds, our legal practice management tool, benefits directly from taxonomic synthetic data that has been filtered for the kind of ambiguous fact patterns real litigators encounter, not the textbook examples that dominate publicly available legal corpora. The complexity expansion step in Simula maps almost one-to-one onto what a senior partner does instinctively when briefing a junior associate, and extending it with multi-path filtering means the training set reflects the full distribution of how cases actually unfold.

Fraud Detection and Risk Evaluation

Fraud Detect and Managed Risk live or die by their coverage of novel attack patterns. Fraudsters do not cooperate with your training data refresh cycle. A system built on historical fraud examples is a system that catches last year's fraud well and this year's fraud badly. Synthetic data orchestrated around adversarial taxonomies, with explicit complexity expansion toward the kind of schemes that have not yet hit the historical record, is the only plausible path to keeping pace. Game-theoretic reasoning, the backbone of Managed Risk, becomes dramatically more effective when the training distribution was deliberately built to include the moves an adversary has not made yet.

Manufacturing Quality Assurance

Solace Manufacturing addresses a domain where the rare defect is the entire point. Ninety-nine-point-nine percent of parts coming off a line are fine. The model needs to be exquisitely sensitive to the one in a thousand that is not, and it needs to be calibrated against defect modes that have not yet appeared in production. Orchestrated synthetic data lets us build that sensitivity without waiting for real failures to happen, which in a manufacturing context means catching problems before they become recalls, injuries, or liability events.

Dataset Integrity Itself

Perhaps the most recursive application is Dataset Reality Check, our pre-model drift detection tool. The techniques Simula introduces for assigning taxonomic nodes to existing data points, and the nearest-neighbor retrieval they enable, are directly usable for auditing the datasets your organization is about to train on. Before a single parameter updates, you can see exactly which corners of the intended distribution your data covers, which corners it misses, and which regions are over-represented in ways that will bend the resulting model. That is a capability most teams currently do not have, and it is the difference between shipping a reliable system and shipping a confident one.

Narrative and Expectation Analysis

Narrative Capture Radar is built to detect when public discourse, internal communication, or intelligence feeds contain expectation-violating signals that warrant attention. That requires training data spanning the full taxonomy of how narratives normally behave, so that abnormal behavior is actually detectable as such. Multi-path synthetic orchestration is close to the only way to assemble that kind of balanced corpus without running into privacy and licensing walls at every turn.

The Third-Way Alignment Thread

None of this work sits outside a values framework. Every dataset we orchestrate, every synthetic corpus we build, is tested against the principles documented at thirdwayalignment.com. Mutual respect means the data reflects the real diversity of the people and cases it will be used to reason about. Shared flourishing means we do not build systems that concentrate false confidence in the hands of whoever deploys them first. Ethical coexistence means the process is transparent enough that a regulator, an auditor, or a defense attorney can trace exactly how a given decision was supported by training.

The JULIA test framework, our five-domain self-assessment covering Justice, Understanding, Liberty, Integrity, and Accountability, is applied to every model we train on orchestrated synthetic data. That is what separates a pipeline that produces strong benchmarks from a pipeline that produces systems you can actually deploy in high-consequence settings without lying awake at night about the edge case you never thought to test.

The Honest Read on Where We Are

Google's Simula work is a serious contribution to a field that badly needed one. The quality of the datasets produced by the framework, from everything we have been able to preview and compare, is genuinely very high, and the paper earns the attention it is receiving. What we are doing at Detailed In Design is not a replacement for that work. It is an extension that addresses the specific blind spots any single-pipeline synthetic approach will produce, and it is built specifically for the domains where the cost of a blind spot is measured in wrongful arrests, missed diagnoses, undetected fraud, or failed parts.

The transformer era is closing. The post-transformer era will be built on targeted training runs over carefully orchestrated synthetic corpora, and the organizations that figure out how to build those corpora well, with real diversity, real complexity, and real explainability, will define the next decade of what AI is actually capable of doing. At Detailed In Design, we are building that capability now, product by product, under Third-Way Alignment principles, with the kind of forensic rigor the problem demands.

If this is a conversation your organization needs to be part of, the door at detailedindesign.com is open. The Google paper is worth reading in full. And the work ahead is worth doing right the first time.

Detailed In Design builds proprietary AI systems with asymmetric architecture for high-consequence domains. Services begin May 1, 2026. Learn more at detailedindesign.com and thirdwayalignment.com.