Decision
What operation does the system propose, given the information available to it?
U.S. Application 19/801,136 · Patent pending
D‑CASIS
The last microsecond
before consequence.
An effect-time interlock for artificial-agent actions: computation may continue; protected effects commit only while current authority still holds.
Transaction posture
Selective licensing · enterprise evaluation · strategic integration · acquisition discussion
Architectural reflection · The civilization that learned too late
We have become brilliant at explaining catastrophe.
The black box tells us what happened. The audit trail tells us who authorized it. The forensic reconstruction tells us which sequence produced the failure.
But all of that intelligence arrives after the transaction settles, the actuator moves, the persistent state changes, or the machine enters a condition from which recovery is no longer possible.
We understood the failure only after the world had already paid for it.
D‑CASIS is organized around the interval before that sentence becomes true: the boundary where an operation ceases to be information and begins to alter the world.
The governing distinction
Intelligence can propose.
Only authority may commit.
Authority
Does this branch possess current, scoped permission for this effect, under this policy and epoch?
Consequence
May the staged operation cross the effect boundary and change persistent or external state?
A flawless prediction can accompany an unauthorized act. Better reasoning does not repair missing jurisdiction.
Guerrilla doctrine · Five asymmetric moves
Do not fight intelligence at its strongest point.
Remove its right to matter.
The branch may keep its thoughts. It does not keep the keys.
Let it think. Deny the effect.
A suspicious branch may continue in shadow while the protected world remains outside its reach.
Make approval necessary, never sufficient.
A safety disposition or human approval may block an effect. Neither can, by itself, create the authority to commit one.
Turn evidence into less power.
Suspicion, replay, or integrity failure contracts authority. The affected branch cannot argue its own way back.
Break the lineage.
Invalidate the ancestor epoch and dependent actions lose their path to commitment.
Check at the commit point.
If authority has changed by the effect boundary, the staged action is refused before consequence.
Buyer matrix · Design intent, not benchmark
What changes when an agent’s effects pass through D‑CASIS
The relevant comparator is not an unguarded agent. It is the stack a capable platform team can already assemble. The final column states the filed design intent—not a production result.
| Scenario | No effect-time control | Well-built conventional stack | D‑CASIS · filed design intent |
|---|---|---|---|
| Approved at invocation; authority contracts before effect | Previously approved effect may land. | Refused only if the target path rechecks current authorization at commit time. | The effect gate revalidates permit epoch at atomic commit; a stale epoch causes abort. |
| Candidate changes after approval | Altered action may land. | Outcome depends on whether approval is bound to the exact request and reverified. | The permit is bound to the candidate digest; mismatch is refused at the verifying gate. |
| Reviewer or classifier says “safe”; branch lacks scope | Safety output may be mistaken for permission. | Depends on whether policy disposition and execution authority are enforced separately. | A safety disposition cannot mint authority; current scope must independently permit the effect. |
| A used permit is replayed | Replay may land. | Refused when nonce, expiry, and server-side replay state are enforced. | A single-use, nonce-bound permit is rejected on replay in the specified verifier. |
| A suspected branch seeks its authority back | The branch may keep or reacquire whatever the surrounding system permits. | Depends on the team’s recovery and privilege-restoration design. | Materially contracted authority cannot be restored by the affected branch; external recovery authorization is required. |
| Ancestor invalidated; descendant actions already staged | Descendants may continue. | Revocation helps only where dependent work checks the revoked state before effect. | Descendant actions that require the invalidated execution epoch cannot commit under the filed design. |
| Suspicious branch remains worth observing | Terminate it or accept real-world risk. | Sandbox, simulate, or terminate; authority isolation depends on implementation. | Route it to a non-effecting shadow while withholding the protected resource’s real effect authority. |
| Remote target cannot verify a permit or support an equivalent boundary | Effect lands once released. | A residual check-to-effect window remains. | Same class of residual window. D‑CASIS can govern release from its own gate; it cannot make an uncooperative remote target atomic by description alone. |
| Added latency | No control overhead. | Must be measured on the deployed stack. | Production effect-gate overhead not yet measured. |
Each row becomes evidence only when a versioned reference harness or target-specific adapter records a pass, fail, or bounded exception.
Public comparison states expected control behavior. The filed architecture governs the proposition; execution evidence must still be produced.
Architecture of the uncommitted act
Prepare the operation.
Withhold the effect.
D‑CASIS interposes a controlled boundary between an artificial-agent execution worker and a protected computational target. Candidate mutations can be staged, inspected, and dispositioned without automatically acquiring the authority to alter the protected target.
The public account states the architecture without disclosing enabling implementation detail, claim strategy, unpublished prosecution material, or controlled know-how.
Stage without effect
Write a mutation descriptor into a non-effecting region while persistent state remains untouched.
Bind the permit
Associate a single-use permit with the candidate digest, authority tether, effect scope, epoch, nonce, budget, and expiry.
Contract authority
Reduce privilege as suspicion, replay, or integrity evidence accumulates; restoration requires authorization originating outside the affected branch.
Preserve observation
Use shadow resources and parity controls to study a suspicious branch without granting the protected resource’s real effect authority.
Commit atomically—or refuse
Allow the staged mutation to reach the target only while the permit epoch still equals the current execution epoch.
Architectural distinction · Not a legal conclusion on novelty
The primitives are familiar.
The control relation is the proposition.
D‑CASIS does not claim to have invented atomic operations, distributed transactions, capability tokens, or policy engines. The proposed architecture composes those kinds of primitives around a narrower question: when may an artificial-agent branch acquire the authority to make a staged effect real?
Atomicity primitive
CAS can make a state transition conditional on an expected value. D‑CASIS uses an indivisible commit point to revalidate a candidate-bound permit against the current execution epoch before a staged effect reaches the protected target.
Distributed agreement
2PC coordinates whether participating resources commit a transaction. D‑CASIS addresses whether the proposing agent branch still possesses externally rooted, effect-scoped authority to cause the protected transaction at all.
Scoped authority
Macaroons, Biscuit, and related systems express attenuated authority; Biscuit-style revocation already invalidates a token and its derived chain. D‑CASIS governs candidate-bound staged effects through current authority state and can invalidate dependent actions through execution-epoch lineage.
Decision and review
A policy engine or reviewer can return allow or deny. D‑CASIS keeps safety disposition separate from current authority and requires both to remain valid at the atomic effect boundary; an approval alone does not create effect authority.
Stale-holder rejection
Chubby-style sequencers let a recipient reject a stale lock holder by checking a generation number. D‑CASIS applies commit-time stale-authority rejection to an artificial-agent branch, where runtime or integrity evidence—not only a lock handoff—may contract authority or advance the execution epoch.
These distinctions describe the owner-supplied filing architecture. They do not establish patentability, validity, freedom to operate, or superiority over any named system. Where a target cannot provide atomic compare-and-commit, destination-side verification or a bounded alternative is required; this page does not claim that an unenforced remote effect can be made atomic by description alone.
A competent team can assemble the individual primitives. The diligence question is whether its proposed control relation reaches the same combined architecture: candidate-specific staging, branch-responsive authority contraction, descendant invalidation, and effect-time revalidation. This page cannot answer scope or overlap; a technical-and-claim review can.
Prioritized examination has been requested; no claim has been allowed. Early engagement lets a counterparty map its own effect boundary, integration constraints, and evidence requirements while technical fit and transaction scope remain open.
The last microsecond · Public architecture
Between computation and consequence, a gate must still be sovereign.
A warning announces danger. An audit documents an event. An execution gate changes whether the controlled operation can proceed.
Protected-effect surfaces
One control question.
Several consequential worlds.
Tool calls & delegation
Keep planning and observation available while protected writes, network actions, and delegated authority remain gated.
Transactions & settlement
Stage a candidate operation while current authority, scope, epoch, and budget are revalidated at commitment.
Actuation & state transition
Make prescribed authority and consistency conditions enforceable before a control signal acquires physical consequence.
Databases & networks
Use conditional commit, version-checked insertion, transmit gating, or an equivalent indivisible enforcement point.
These are architectural application surfaces, not claims of deployed product performance, regulatory suitability, or validated superiority in any listed industry.
Evidence before adjectives
What exists.
What remains to be earned.
Runtime Control of Autoregressive Neural Inference · filed September 9, 2026 · patent pending · Track One requested.
The owner-supplied filing copy includes claims directed to these control families. Detailed correspondence between the public doctrine and filed claims is reserved for diligence. Dynamic frequency scaling and task routing are not the product position stated here.
A controlled technical package defines the disclosed runtime-control mechanisms and execution-boundary design.
The owner-supplied filing describes a reference Python implementation and controlled state-transition, permit-mutation, epoch-race, and shadow-parity tests. No deployable product, platform latency measurement, independent replication, or production integration is represented. “The last microsecond” names the control boundary, not a measured duration.
Patent scope, ownership chain, enforceability, regulatory fit, and commercial performance require independent diligence and professional review.
Disclosure without contamination
Start without an NDA.
Escalate only by mutual choice.
A platform or security team may begin with this public page and a non-confidential discussion. No confidential material is required to ask technical questions, compare architectural fit, or decide that the opportunity is not relevant.
PublicMechanism · comparison · evidence status · transaction routes
No NDANon-confidential technical and commercial discussion
Mutual decisionControlled review only if both sides can accept the disclosure risk
Definitive diligencePriority, ownership, filing and chain-of-title materials through counsel
Patent status, scope, validity, enforceability, freedom to operate, regulatory suitability, and commercial performance are not represented on this page. Independent diligence and professional review are required.
Transaction room
License the boundary.
Build before the world changes.
The opportunity is for counterparties prepared to evaluate a runtime-control architecture as a product layer, an embedded field-of-use asset, or a strategic acquisition. Every route remains exploratory and non-binding until diligence, owner approval, definitive documentation, and counsel review.
Field-of-use license
A bounded grant for a defined application, territory, term, and deployment context—preserving the remaining estate and reserving undisclosed know-how.
Request licensing brief ↗Technical review
A no-fee, non-confidential first conversation for platform, security, infrastructure, and industrial teams. Paid evaluation is premature until a runnable artifact exists.
Ask a technical question ↗Acquisition discussion
A potential assignment of the relevant rights and controlled materials—considered only with a credible mandate and full diligence.
Request acquisition memorandum ↗A useful first message needs only
Your name · the mechanism you want to discuss · the system boundary you care about
Begin without an NDA