Sapien Weekly Digest

PoQ Updates:

Preparing for the new App:
This week, we are gearing up to run the first end-to-end PoQ consensus cycle on representative datasets. Contributors will submit work, validators independently review it, and the protocol produces a stake-weighted quality signal. The team is focused on proving that the mechanism works at scale before expanding the product around it.

Builder’s Corner:

Proof of Quality is built by developers for developers to find an answer to what we consider the major bottleneck in AI. On top of building for developers, we more importantly want to build with you, too. The most useful input to that end right now is concrete workflow detail.

What models are you building or fine tuning right now? What would the consequences of a wrong output be?

If you think what you’re building should inform Proof of Quality, respond to this email and we will be in touch and join the early adopters list for priority access when Proof of Quality goes live!

Under the Hood:
Most AI pipelines fail because the quality decisions shaping the model are hard to verify and even harder to audit.

The failure mode we’re addressing is familiar to anyone building with human review in the loop: data gets labeled, reviewed, approved, and pushed forward, but no one can say which judgments were reliable or what should have been removed before it touched the model.

The current build is designed to change that. Developers will submit data through an SDK or REST API, define the grading rubric, and get back a quality signal that shows where a dataset holds up and where it breaks. That matters whether you are training a model, evaluating outputs, or trying to understand failure points in the model. Right now, the interesting engineering question is whether the consensus mechanism produces useful signals on messy, real datasets.

Reference Workflow:

Starting today, we will be sharing with you what types of data and datasets we’re expecting to work well with Proof of Quality out of the gate  to give you an understanding of the process. Consider a  dataset used to train a model that flags potential dangerous medical conditions.

The source labels may be inconsistent, certainly some are low confidence. Some may simply be wrong. Because of how sensitive the data is, realistically, an expert in the field should have had a pass at labeling the data. But can we really expect a medical professional to sit down and label image after image?

A developer sends that dataset into Proof of Quality, defines a grading rubric for what counts as a valid classification, and lets the system run independent review against it.

What comes back is a quality report that highlights disagreement, flags suspect labels, and identifies the slices of data most likely to degrade model performance.

Proof of Quality creates a defensible process for deciding what data got allowed into the dataset and why in a situation where a hallucinated output or a factually wrong one could have serious consequences.

Our Voices in the World:

On Thursday, our Product Manager Ali Malik was invited by Fhenix to join a conversation on the AI Agent boom and what infrastructure needs to be built to make them viable to sensitive applications in high velocity applications like trading. Rewatch the livestream right here!

AI Market Breakdown:

The signals we see & build upon:

The market is moving faster on AI deployment than on AI verification. Here’s why:

In regulated environments, that is becoming a blocker. Buyers increasingly need proof of how data was reviewed, who reviewed it, and whether that process can survive audit or compliance scrutiny. At the same time, synthetic data is making provenance harder, not easier. More AI-generated inputs are entering training pipelines without reliable controls for lineage, duplication, or expert validation.

The GSA proposes a strict AI clause:
As of March 2026, the U.S. General Services Administration (GSA) has proposed a strict new AI procurement clause, formally known as GSAR 552.239-7001, “Basic Safeguarding of Artificial Intelligence Systems”. This clause is designed to apply to federal contracts through a mass modification expected in late March or early April 2026.

The draft says contracting officers must insert a new clause, “Basic Safeguarding of Artificial Intelligence Systems,” into solicitations and contracts for AI capabilities. It requires disclosure of all AI systems used in contract performance, use of only “American AI Systems,” human oversight and traceability, summarized intermediary processing steps, source attribution for retrieval methods, 72-hour incident reporting, documentation aligned with NIST AI RMF and LLM transparency requirements such as system cards, and government rights to run its own automated evaluations against production systems.

AI agents in Healthcare are advancing with breakneck speed:
At HIMSS on March 9, Mass General Brigham’s chief information and digital officer Jane Moran spoke about how healthcare organizations are treating adoption of AI as high stakes because of clinical safety, cybersecurity, privacy, and regulatory concerns. She also said Mass General Brigham built a secure internal “AI Zone” that gives employees access to approved large language models, assistants, and agents, rather than letting AI use sprawl uncontrolled across the enterprise. Moran said that the efforts to make use of the technology enabled by AI have to be stay ethical while also being quality driven thanks to human oversight. 

On March 11th, STAT reported that AI agents are spreading through health systems “faster than they can be counted,” with Epic promoting agents for documentation, billing, and patient communication, Oracle rolling out an agent for physicians in 30 specialties, and Amazon, Google, and Microsoft all adding health-focused AI personas at HIMSS. According to STAT, the issues facing adoption are that validation is lacking, and patients are seldom consulted on development and testing. Healthcare Dive reported the same week that regulators are struggling to keep up and that agentic systems will require new frameworks, continuous oversight, and more data sharing.

Model Governance Is the Bottleneck in Finance:
On March 13th, FinTech Global reported that new Hawk and Chartis research found nine in ten financial institutions now actively encourage AI in financial-crime and compliance operations.The study surveyed 125 compliance and risk leaders at banks globally and found broad support for AI alongside a governance bottleneck, with more than half of the technical challenges slowing broader deployment tied directly to model governance. Among the most-cited issues were poor or limited training data at 91%, integration problems at 86%, explainability at 83%, data and model governance itself at 73%, and degrading model performance over time at 70%.

Those issues intensify after deployment, with firms reporting rising concern about their inability to update live models, maintain governance across a growing model inventory, and sustain trust in model decisions.

The bottom line:
The constraint around continued AI adoption is moving away from raw model capability and toward validation, provenance, runtime controls, data residency, and audit evidence. The teams that solve those layers cleanly will not just look more responsible. They will move faster through procurement and into production.

More to show off next week!