Expound/Resources

Every paper and article, in one place

The core and long technical papers, every published paper and article, and the recorded demonstration, with the LinkedIn edition and the PDF wherever both exist. Start with the core paper if you want the whole argument; start with an article if you want one piece of it; watch the demonstration if you want to see it run.

Technical papers

The core argument, and the full source record.

Core paper

Finality Assurance™: Core Technical Paper

The citable core argument: the problem, the formal result, the architecture, the benchmark evidence, the economics of Accepted Work, and the governed Work Management product.

PDF · 47 pages · August 2026
Long paper

Finality Assurance: Long Technical Paper

The full technical source record behind the core paper.

PDF · 241 pages · August 2026

On computing reliance

Who may rely on a result, for what, now.

Paper

AI Fails in Many Ways. Many of Those Failures Go Unseen.

AI fails in four places: the model, the system built around it, the people who approve its work, and the facts nobody has. Many of those failures look correct and come with no warning, and better models will not cure a large share of them. It builds on Satya Nadella’s post on treating AI models as insider risks.

PDF · 15 pages · October 12, 2026
Article

Your code didn’t change. The test results behind your AI approval did.

A supplier can correct the test results your AI approval relied on while your own system stays the same. A model name does not tell you which setup was tested. Recheck the decisions that relied on what changed.

LinkedIn article · September 24, 2026
Paper

AI Agents Need Governed Evidence and Computed Reliance

When is an agent done: when it says so, or when the evidence establishes it and a named party may rely on the answer now? Evidence before done, reliance after: the full argument behind the home page.

PDF · 26 pages · September 2026
Article

What is the reviewer actually deciding?

Three of the four challenges to human oversight are answered by equipping the reviewer. Compute the confirming; the person keeps the authorizing.

LinkedIn article · September 1, 2026PDF
Article

Finality Assurance Principles: An Architecture for Computing Reliance

The eight runtime decisions a single status value is usually asked to carry.

LinkedIn article · August 29, 2026PDF
Article

Can Your Enterprise Compute What It May Rely On?

The party, result, purpose and now question that a reliance license answers.

LinkedIn article · August 26, 2026PDF
Article

FAS Enables AI for Regulated Industries

Governance as a runtime checkpoint rather than a policy binder.

LinkedIn article · August 20, 2026

On what counts as work

Accepted Work, and what it cost.

Article

Is Your AI Safer, or Just Saying No?

OpenAI says some of one model’s higher safety scores may come from refusing more requests, including legitimate ones. Before you buy, ask what the AI stopped doing.

LinkedIn article · September 25, 2026
Article

Your AI Produces Outputs. Which Ones Count?

Accepted Work per Dollar applied to FinOps (cloud financial operations), and the gap between what a system generates and what it can count on.

LinkedIn article · August 26, 2026
Article

Measuring AI results: The cheapest disposition is refusing everything

Cost to disposition, and why Accepted Work per Dollar needs a numerator.

LinkedIn article · August 18, 2026
Paper

Accepted Work per Dollar (AWpD)

Accepted Work per Dollar as the measure that cannot be gamed by producing more.

PDF · 44 pages · August 2026

On routing and execution

Routes, not models.

Article

The Hugging Face exploit: the agent stole the answer key, and the benchmark gave it credit

Two failures treated as one: an isolation failure, and an acceptance failure. This work addresses what happens to a result once that boundary has already failed: whether it can still be accepted for reliance.

LinkedIn article · August 20, 2026
Article

Today’s routers pick a model. The work needs a route.

A route is not a model name. The comparison of what each layer carries in its object model, and why Governed Multi-Route Selection (GMRS) and Constrained Policy Reinforcement Learning (CP-RL) are released under AGPL v3, with a commercial license for embedding.

LinkedIn article · August 19, 2026

On friction, rights and economics

AI changes what friction costs, and who can afford it.

Paper

Music Has BMI. Software Needs Its Own.

A right you cannot afford to check protects little. AI is cutting the cost of finding out whether a possible violation of software rights matters. A shared organization, as music has in BMI, can spread the costs AI cuts least across many smaller publishers.

PDF · 19 pages · September 2026
Paper

OpenAI’s “Software Factory” Is Not a Factory. It Is a Heavily Automated Development Process.

A factory earns the name by what it can tell you about anything it shipped, not by how fast it shipped. Read from its own public description, OpenAI’s process automates a great deal and answers almost none of the questions a factory has to answer.

PDF · 21 pages · September 17, 2026
Paper

Medium-Friction Businesses Are Doomed in the AI Era

Some businesses earn because their customers find the work too expensive to do, or too hard to check. AI is making understanding, acting and proving cheap for the customer. For many of those businesses, that cost was the profit.

PDF · 17 pages · September 16, 2026
Paper

AI Companies Are Already Regulated. Who Can Afford to Hold Them to It?

A rule binds a large supplier and a small customer the same way. The work of making it count does not. The supplier holds the records a customer needs in order to complain, and that decides who can actually use the protection.

PDF · 23 pages · September 15, 2026
Paper

Fair Use Is Not Immunity

Four steps an owner can take before deciding whether to sue: know what you own, control what you disclose, keep the evidence, and get one legal review.

PDF · 18 pages · September 8, 2026
Article

The Right to Say Yes

A company can obey your software license and still take the category.

LinkedIn article · August 30, 2026PDF
Article

Satya Nadella Asked Who Keeps the Learning. The Harder Question Is How You Would Know.

Existing law may reach it; the hard part is making compliance independently checkable.

LinkedIn article · August 24, 2026PDF
Paper

Startup IP Is the Most Underpriced Asset on Your Cap Table

The right always existed; the cost of finding out protected the conduct around it.

PDF · 52 pages · August 2026
Paper

Who Can Afford Friction? How AI Reprices Economic Power

AI does not remove friction; it changes who can afford to bear it.

PDF · 23 pages · August 2026
Case note

A Trade Secret Claim Dismissed Because of How the Secret Was Made

Trinidad v. OpenAI: the claim failed on reasonable measures, because on the face of the complaint the act of creating the material was also the act of disclosing it.

PDF · 5 pages · September 7, 2026

Analyzing the analysts

What the analyst firms cover, and what it costs to be recognized.

Paper

Who Can Afford to Create or Change a Category?

A large company spreads the cost of analyst recognition across many products and customers. A small one pays it out of the same hours it needs for building and selling. The bill has three parts, and two of them cannot be bought to anyone’s deadline.

PDF · 41 pages · September 14, 2026
Paper

What Gartner’s AI Governance Research Covers, and the Eight Gaps It Leaves

Fourteen requirements decide whether an enterprise can rely on an agent’s work. Gartner’s public research covers six. The other eight are off its research agenda, and each one carries a cost.

PDF · 19 pages · September 11, 2026

Demonstrations

The invoice demonstration, in full and stage by stage.

One invoice, decided two ways: the ordinary workflow marks it payable; Finality Assurance Standards holds it, names the missing proof, passes it once the proof arrives, and withdraws permission when the basis changes. The full recording runs about twenty minutes; each stage is also posted on its own. All of it is on the Expound YouTube channel.

Full recording

FAS Invoice Demo: Full

All eleven stages, 0 through 10: the false pass, the hold, the independent check, the pass, the withdrawal, then routing, the kernel and the industry scenarios. Stages 8–10 appear in the full recording only.

Video · 20:28 · September 4, 2026
Stages 0–1

One invoice, two decision methods, three outcomes

Why FAS is different, and the whole demonstration in one screen.

Video · 1:46 · September 4, 2026
Stages 2–3

The workflow says PAYABLE with required proof missing

The ordinary workflow runs and marks the invoice payable. The omitted receiving proof never surfaces.

Video · 2:17 · September 4, 2026
Stage 4

How FAS knows which proof is required

Five of six obligations supported; the result is HOLD, naming the missing proof.

Video · 1:26 · September 4, 2026
Stage 5

The producer does not get the last word

An independent verifier returns HOLD by its own path. The Four-Corners Test™ passes.

Video · 1:04 · September 4, 2026
Stage 6

One authorized receiving record changes the result

All six obligations met; PASS, with Accepted Work 1.

Video · 0:54 · September 4, 2026
Stage 7

A past PASS is not permission forever

A revocation is admitted. Permission is withdrawn; the earlier pass stays in the record.

Video · 1:12 · September 4, 2026

PAYABLE, HOLD and PASS are the invoice workflow’s on-screen labels. The standard’s own verdict grades are BLOCKED, UNKNOWN, DEGRADED and VERIFIED.

Follow the work

New articles land on LinkedIn first.

Follow Expound on LinkedIn, subscribe to Expound on YouTube for demonstrations, or sign up below and we will send the next one when it is worth your attention.