Expound/Products
One standard, four products
Finality Assurance™ Standards (FAS) is a new kind of computation, not just another policy engine. It goes wherever a consequential decision is made, across the systems and AI agents you already run. It is targeted: start with one consequential decision and measure it against the way you handle that decision today.
Availability
What ships, and when.
FAS is available now. The three products that follow are in development, with no public release date; ask about design-partner access. The standard came out of building the Work Management product, which was used to govern its own development.
| Product | What it does | Status |
|---|---|---|
| Finality Assurance StandardsFAS | Computes whether consequential work is complete and what may happen next on that evidence. | Available |
| Model RouterGoverned Multi-Route Selection (GMRS) + Constrained Policy Reinforcement Learning (CP-RL) | Selects the whole way a job gets done, not which model answers. | In development · no public date |
| Mechanistic Conflation EngineMCE | Finds places where two different things were treated as one, and turns each confirmed finding into a control that stops it happening again. | In development · no public date |
| Work ManagementAccepted Work | The authority layer over organizational work. | In development · no public date |
FAS is the first product and the foundation. Each product that follows is built on it.
Questions buyers ask
Five questions buyers ask before they license FAS.
Will FAS make our AI less useful?
It is not built to. FAS leaves what a model writes or suggests to the model, and decides what each output may be used for: shown to a customer, used to pay an invoice, shipped, or passed to another system as fact. An answer nobody expected can be used once it meets the requirements for that use. A step that does something, such as spending money or opening confidential data, needs permission like any other action.
In our tests, FAS refused none of 270 correct answers that came with their evidence. It refused all 90 correct answers that came without it. That refusal is the cost of the check, and we publish it beside the result.
What do you mean by “rely on”?
Whether you may act on a result: approve the loan, sign the contract, close the account. An answer can be right without being enough to act on. The evidence the action needs may be missing. The person who approved it may not have had the authority. The facts may have changed since it was checked. FAS answers one question for each result: what may this be used for, by whom, and until when.
In our tests, accepting answers on the producer’s word let 7 wrong answers through out of 2,520. Checking the evidence let none through. Those were tasks whose answers could be checked by recomputing them.
Why isn’t this already part of my AI platform?
It can be. FAS can run at the application, the control plane, the model and orchestration layer, or the infrastructure, so a model or platform vendor can license it and build it into its own products. Producing an answer and deciding whether your organization may use it are different jobs, even when one system does both. Where the check runs changes what may be claimed about a result, not the rules it follows.
How model and platform vendors work with Expound →
Is FAS worth the setup outside high-stakes work?
Start where the consequence justifies the effort to measure the value. FAS is not limited to high-stakes work, but the case for ordinary work rests on setup and reuse: what a first governed decision takes to set up, and how much of that the second decision reuses. We have not published those numbers. Until we have, we make no claim about the economics of ordinary work, and the first engagement is one consequential decision on your own baseline.
How a first decision is measured →
Won’t better models and better alignment make this unnecessary?
Better models and better alignment can remove some failures. A more capable model does not, by itself, show that a decision has the evidence and approval it needs. A model released next year cannot add them to a decision your organization makes today.
Start with one workflow. Compare your current setup with the same setup plus FAS, using measures you fix before the test: wrong acceptances, wrong refusals, time, expert effort and full cost per accepted result. If your current setup meets your requirements for less, that is worth knowing. If a gap remains, you have measured it instead of waiting on a model release.
The products
Finality Assurance Standards
FASNot looks done, but proven done: on evidence a second party can re-check without calling you.
- Status
- Available now
- License
- Proprietary standard; licensed
- Conformance
- Tested and certifiable
- Implementations
- Multiple languages
- Integrates at
- Application, control-plane, model, or infrastructure boundary
What it does
FAS computes whether consequential work is complete and what may happen next on that evidence. For each call it identifies the required evidence, excludes routes that cannot meet the requirements, computes what may be relied on now, limits what may happen next, and decides whether the evidence-cleared result counts as Accepted Work.
The problem it addresses
Done is asserted rather than computed, and every department and tool is free to invent its own meaning for authorized, ready, accepted, closed and current. Nothing computes what a named party may rely on, for what, until when.
What changes
The settable status field stops being the answer. A verdict is computed from evidence attached to the decision, against requirements written down before the work rather than assembled after it, and it is recomputed when anything underneath changes.
Technical value
One vocabulary for verdicts, case files, evidence classes and claim limits, specified precisely enough for a licensed implementation to be tested against it for conformance. The evidence each call needs is looked up from its context and an approved profile, not hard-coded.
Business value
Permission to rely on a result becomes something you can grant, limit, transfer and withdraw on the record, instead of something people assume. An audit reads what is attached instead of piecing it together.
Capabilities
Finality Kernel
The core computation. It takes the requirements, the classified evidence, who has authority, whether it is all still current and the exact action being asked about, and returns a case file (the Account), an answer (the Verdict) and a permission to act (the reliance license).
Obligation rosters
The list of what must be proved for each kind of decision: versioned, written down in advance, and applied by one general evaluator rather than hard-coded case by case.
Graded verdicts
Four grades (BLOCKED, UNKNOWN, DEGRADED, VERIFIED), where the whole result gets the grade of its weakest requirement. More than one way to pass, each tied to a specific decision.
Claim and reliance licenses
What may be said about a result, and who may act on it, for what and until when: two separate records with a fixed shape. Your industry supplies the ranking of decisions.
Independent verifier
Read-only; can narrow, hold or refuse a result, and can never raise one.
Four-Corners Test™
Whether a second party can reconstruct the determination from what is attached and nothing else.
Continuous Finality™
Who relied on what is recorded at the time; when something changes, everything affected is found, permissions are narrowed or withdrawn with notice, and history is preserved.
Semantic Virtual Machine
If a step is not allowed by the rules, the SVM will not run it.
Confidence Ladder™
Limits what may be claimed about whether a result reproduces, from R0 to R3, based on kept evidence. It shapes what you may say; it never grants permission to act.
Accepted Work
Only work that cleared its evidence requirements counts. A refusal is a legitimate outcome, not a failure to deliver.
Model Router
GMRS + CP-RLToday’s routers pick a model. The work needs a route.
- Status
- In development · no public date
- License
- Dual: AGPL and commercial
- Built on
- FAS
- Selects
- Model or person · tools · evidence · checks · stopping rules · budgets · escalation
- Cannot
- Choose what evidence counts, or create Accepted Work
What it does
It selects the whole way a job gets done, not which model answers. A route is the model or the person, the tools, the evidence to be gathered, the checking to be performed, the stopping rules, the budgets, and when to escalate.
The problem it addresses
Routing today ranks models on cost, latency and a quality score. Nothing checks that the chosen path could even produce the evidence the question needs, and an optimizer that can loosen its own rules effectively has none.
What changes
Eligibility is decided before anything is ranked. Routes that cannot produce what the result has to prove are ruled out first; a control from the conflation engine enters as a hard rule, not a preference; and sending the job to a person is a route the rules can choose like any other. If no eligible route remains, the router stops and names what was missing rather than returning the best of a bad set.
Technical value
A route is a commitment, not a preference. Two routes can use the same model and still be different routes, because the evidence, checks and stopping rules attached to them differ, and so does the full cost of reaching an outcome. The route that was selected is tied to the route that actually ran.
Business value
Spend follows what the decision actually requires, and no score buys its way around a hard requirement. The learning half ranks eligible routes and can never widen the set.
Capabilities
Governed Multi-Route Selection
Keeps only the routes that can produce the evidence this call needs. Eligibility comes before ranking.
Constrained Policy RL
Ranks eligible routes only. It cannot decide eligibility, change evidence requirements, weaken obligations, widen what may be claimed, or promote itself.
Governed hold
When no route is eligible, the output is a clear refusal that names what was missing: something you can act on, not a silent best effort.
Escalation as a route
A person is a route like any other, with its own evidence requirements, chosen by the rules rather than bolted on when automation fails.
MCE controls as constraints
A route whose model has a known failure the work cannot tolerate is ruled out, not just marked down.
Promotion gate
A newly learned policy goes live only through a governed decision with its own evidence. The learner must be able to be told no.
Mechanistic Conflation Engine
MCEA defect found once becomes a defense everywhere it applies.
- Status
- In development · no public date
- License
- Separately licensable
- Releases
- No public date; follows the Model Router
- May
- Detect · cap · refuse · quarantine · propose
- May not
- Write the verdict, decide finality, or promote its own findings
What it does
A conflation is treating two different things as if they were the same: a command returning as the change happening, a valid signature as an authorized signer. MCE scans agentic work for them, and turns a validated finding into a control that stops it recurring.
The problem it addresses
Before a result reaches anyone, a run decides what question is being answered, which evidence is relevant, whether a failed step should be retried, whether an observed effect establishes the intended outcome, and whether the work is complete. Those are different questions, and success at one establishes nothing about another.
What changes
Known kinds of failure are enforced by fixed rules rather than rediscovered team by team, and a newly spotted one enters only as a candidate. Nothing the discovery side notices takes effect until it has been approved by someone outside it.
Technical value
It cuts across the whole system rather than sitting at one step in a sequence. It limits what the other parts may do, and it scans the controls themselves: a detector that cannot fail, a control present but never applied, a cure that reintroduces what it was written against.
Business value
A failure family under standing control stops being rediscovered team by team. What it hands the router is a hard rule, not a preference.
Capabilities
Execution conflation
A command that ran, read as the change having happened. A command returning is not the intended effect happening.
Method conflation
A test that failed, read as proof the safeguard works.
Authority conflation
A record existing, read as the record having been properly issued; a valid signature read as an authorized signer.
Trust-domain conflation
Two separate processes read as two independent sources.
Custody conflation
A secret deleted today, read as never having been exposed.
Governed promotion
A confirmed detector becomes a permanent, machine-enforced control only after approval from outside the discovery side.
Work Management
Accepted WorkThere is no field named state to set.
- Status
- In development · no public date
- Role
- The first full implementation of FAS; it governed its own development
- Organizes around
- Capabilities: build · stabilize · keep · retire
- Unit
- Accepted Work, and its cost
- Replaces
- Authority claimed by ticketing, CI, release, audit, observability
- Leaves in place
- The work each of those tools is actually good at
What it does
It is the authority layer over organizational work: the system that decides whether a piece of work can be relied upon, and does not let anyone overturn that by saying so.
The problem it addresses
Authority over what done means is claimed separately today by ticketing, requirements, test management, continuous integration, release and change, security closure, audit, observability and knowledge bases, none of which was built to agree with the others. A checkmark is three assertions in one gesture: that the work was planned, that it was done, and that acceptance was achieved.
What changes
Nobody drags a card to Done, types a percent complete, or signs a readiness checklist, because those are not operations here. Progress and readiness are computed from what actually landed. Typing a computed value into a status field is treated as a bug, not a feature.
Technical value
Every change of state passes through one gate, carrying its authorization and evidence, so no tool, dashboard, agent or person has a back door to the verdict. People set policy, handle exceptions and decide contested cases: they authorize, and they do not edit the verdict.
Business value
The unit of measure is Accepted Work and its cost, not output. A refusal is a legitimate outcome, not a failure to deliver. When permission to rely on something is withdrawn, the work that rested on it is found automatically and comes back with its reasons attached.
Capabilities
Capability-first, not project-first
The organizing unit is a lasting ability with an owner, evidence obligations, a map of what it depends on, and a computed status: build, stabilize, keep or retire. Every tool underneath plugs into it.
Computed lifecycle
Readiness and state derived from what landed, under obligations declared before the work starts.
One registered boundary
Every state change carries authorization and evidence through a single gate. No back door.
Producer never grades
The actor that produced a piece of work never grades it.
Governed exceptions
Held, classified, explained, with the missing requirement named: a record, not an override.
Accepted Work per Dollar
Governed value against the full cost of all attempts. Refused attempts stay in the denominator.
Dependency-driven reopening
When Continuous Finality withdraws reliance, dependent work returns to the queue with its reasons attached.
Governed feature flags
A flip is an approved change with declared evidence, and everything that depended on the old behavior is recomputed. A flag flip is a release.
Self-governed construction
Used to govern its own construction from early on; the standard is what remained after enough of those changes.
How Work Management is organized
Capabilities, not projects.
It begins from what an organization must be able to do: compute acceptance, control who can change state, produce evidence, verify independently, deploy accepted work, watch what is running, and preserve knowledge. Every harness, workflow engine, registry and deployment tool then plugs in underneath rather than being the architecture.
A capability is a durable ability, not a roadmap label. It carries an owner, who it serves, its business value, target maturity, what it depends on, evidence obligations, exit criteria, health metrics, governance debt, acceptance state, and its contribution to Accepted Work. From health, evidence and governance debt the system computes one of four statuses.
Because dependencies form a governed graph, the questions executives ask become queries: what can we delegate on the evidence, what blocks the release, what did accepted work cost.
| What becomes possible | What it replaces | Why it works |
|---|---|---|
| Governed feature flags | Mutable runtime booleans anyone with write access can flip | A flip is an approved change with declared evidence, recomputation of everything affected, and a dated record |
| Computed lifecycle and readiness | Dragged cards, percent-complete fields, status meetings | State is derived from what actually landed, with a receipt showing how |
| Computed admission | Approval clicks and sign-off chains | The decision is computed from evidence against a declared roster; a person contributes evidence and policy, not a checkmark |
| Delivery in a contracted order | Run-order conventions and tribal deploy knowledge | Prerequisites are an order a machine can check; violations are refused up front, not discovered as incidents |
| Recorded reliance edges | Institutional memory of who depends on what | Reliance is recorded when taken, so withdrawal, notification and unwind are computations |
| Blast-radius preview | Impact guesswork before a risky change | Everything a proposed change would affect, directly or indirectly, is computed before it is approved |
| Attempts that say what they were for | CI readouts judged by exit status | Every attempt says what it was for, so a test that was supposed to fail and passed is read as the failure it is |
Licensing at a glance
Proprietary, published, licensed, conformance-tested.
| Product | License model | Notes | Status |
|---|---|---|---|
| Finality Assurance Standards | Proprietary standard, published and licensed | Conformance testing; certification of conforming implementations | Available now |
| Model Router (GMRS + CP-RL) | Dual: AGPL and commercial | Open core, commercial terms for enterprise | In development; no public date |
| Mechanistic Conflation Engine | Separately licensable | Follows the Model Router | In development; no public date |
| Work Management | Commercial; reference implementation of the family | Governs its own construction | In development; no public date |
FAS is a proprietary, published standard, the model Dolby, UL and Qualcomm use. The specification is owned and published by its author, it is licensed, an implementation is tested against it for conformance, and a conforming implementation can be certified. Patent pending.
What makes it a standard is not who owns it. It is that the specification is precise enough for a licensed implementation to be tested against it for conformance. A reliance object only its author’s agents can produce, and only its author’s systems can read, is that author’s own status field under a new name.
Deployment
You choose where it runs.
The standard does not dictate one integration point, which is the practical reason to start with it. It plugs in at whichever layer suits each kind of work (the application, the control plane, the model and orchestration layer, or the infrastructure), with no single central runtime.
Where it can sit
Major cloud services; model gateways and agent runtimes; central or distributed enterprise control planes; libraries, sidecars and gateways; on-premises, cloud, hybrid, confidential, edge and disconnected environments; and selected firmware, secure-enclave and chip-level controls.
What placement changes
What may be claimed, not the rules. Evidence that comes from inside the system that did the work proves the process ran, not that the effect independently happened, and the Account records which.
What is not flexible
How results are read. Wherever the computation sits, anyone using a consequential result gets it by taking a license, because a read that goes around the license cannot be found when the result is later withdrawn.
How to start
Start with the standard, whether or not you adopt the systems.
Run the test
Take one determination your agents produced this week. Run the Four-Corners Test on it. No tooling needed.
Name one decision
One that you would struggle to defend afterward. Ask whether its obligations can be written down and its evidence classes named.
License FAS
The standard is the place to start. Implementations exist in multiple languages; you pick the integration point.
Conform, then build on it
Conformance testing and certification go through the lab. The router, MCE and the Work Management product arrive on the same standard.
Next steps
The standard is available now.
Licensing, conformance testing and adoption inquiries go to the lab directly.
Licensing and conformance
FAS is available and licensable now. The router is dual-licensed, AGPL and commercial; MCE is separately licensable.
Talk with us →EvidenceThe measured results
Every result, with the boundary on what it shows.
See the evidence →FoundationsThe technical record
The core technical paper and the long technical paper behind it.
Read the research →