Attributed intent exists.
Evidence ceiling E3Proof boundary
Finality Research · first-party category thesis
Outcome Finality
The control layer that reconciles a claimed completion against the source empowered to establish reality, then issues portable evidence or keeps the case open.
Formal definition
Completion is a claim. Finality is an evidence-qualified state.
Outcome Finality is the state reached when a consequential action has valid authority, the required outcome is observed in the authoritative source, material contradictions are absent, and a versioned Receipt records the evidence and reopen rule. An authorization, task completion, tool success, HTTP response, or agent assertion alone is never sufficient.
Finality Gateway
Acknowledged is not authoritative. Authoritative is not final until reconciled.
- 01ACKNOWLEDGED
A runtime or tool reports completion.
- 02AUTHORITATIVE?
Read the source empowered to establish the outcome.
- 03RECONCILE
Compare required state, observed state, authority, contradictions, and obligations.
- 04FINAL / CONFLICT
Issue a Receipt only when the supportable state is explicit.
- 05REOPEN
Preserve lineage when governing reality changes.
Control stack
Route each question to the layer that can answer it.
Identity, authority, policy, execution and observability remain necessary. Outcome Finality adds the post-action reconciliation boundary without claiming to replace them.
Agent control stack
The control question changes as an action moves.
Each layer answers a different question. The Finality Gateway begins where execution systems stop: was the claimed ending established by an authoritative source?
- 01Identity
Who or what is acting?
- 02Authority
May it act for this purpose?
- 03Policy
Is this action permitted now?
- 04Execution
Was the action attempted?
- 05Outcome Finality
Did authoritative reality reach the required ending?
- 06Evidence
Can another party verify and reopen it?
State machine
Ten explicit states. One supportable final state.
The protocol makes uncertainty, conflict, stops and supersession visible instead of compressing them into success.
Outcome Finality protocol v2
Explicit states prevent false closure.
Only VERIFIED_FINAL represents a supportable authoritative ending. Every other state remains open, stopped, conflicted, or superseded.
Authority and policy permit the attempt.
The bounded action has not acknowledged completion.
A tool or runtime claims completion; this is not FINAL.
The authoritative source is missing, stale, or invalid.
Evidence or authoritative reality contradicts the claim.
Required state is reconciled with valid authority and no material contradiction.
New governing evidence supersedes a prior VERIFIED_FINAL Receipt.
The protocol could not evaluate the case.
A policy, authority, or safety stop ended advancement.
Integration profile
Normalize upstream completion without promoting it.
A2A, MCP, REST, webhooks and events enter the Finality Gateway as claimed execution evidence. Domain authoritative sources remain external control owners.
| Upstream signal | Finality state | Why | Required next evidence |
|---|---|---|---|
| A2A TASK_STATE_COMPLETED | EXECUTION_ACKNOWLEDGED | A terminal agent-task lifecycle state establishes what the agent runtime reports. | Read and reconcile the workflow's authoritative source. |
| MCP tools/call isError:false | EXECUTION_ACKNOWLEDGED | A successful tool result establishes protocol-level execution semantics. | Correlate the result to the target system and required ending. |
| REST 2xx / accepted webhook | EXECUTION_ACKNOWLEDGED | Transport or application acceptance does not establish business finality. | Obtain an idempotent authoritative-state observation. |
| Authoritative observation matches | VERIFIED_FINAL | Valid authority, required state and observation agree with no material contradiction. | Issue Receipt v2 and monitor reopen conditions. |
| Later contradictory observation | REOPENED | New governing reality supersedes the prior conclusion without deleting history. | Link the superseded Receipt and re-run reconciliation. |
Public protocol assets
Machine-readable and independently challengeable.
These assets are first-party seller-controlled E3 material. Passing them does not establish buyer acceptance, target integration, legal clearance or independent assurance.
Receipt v2 schema
Machine-readable fields, classifications, lineage, digest and reopen requirements.
Explore →Conformance cases
Twenty deterministic positive, adverse, conflict and reopen fixtures.
Explore →Integration profile
A2A, MCP, REST, webhook and event mappings into one outcome-control boundary.
Explore →Proof Lab
Safe browser-local demonstrations of denial, reconciliation, hashing and REOPEN.
Explore →Adjacent landscape
A bounded category thesis, not a category-award claim.
The reviewed official public materials increasingly cover identity, authorization, governance, execution telemetry and explainability. Finality Group's first-party inference is that those surfaces do not clearly establish a universal cross-platform layer centered on reconciliation of claimed completion against an external authoritative ending with portable evidence and REOPEN lineage.
| Layer / example | What its public material establishes | Outcome Finality boundary | Primary source |
|---|---|---|---|
| Agent identity · Microsoft Entra Agent ID | Identity, lifecycle, ownership, access and governance for AI agents. | Identity establishes the actor; it does not by itself establish the authoritative ending. | Official source ↗ |
| Agent control · Microsoft Agent 365 | An enterprise control plane for observing, governing and protecting agents. | Governance and observability are adjacent to post-action reconciliation. | Official source ↗ |
| Agent runtime · AWS Bedrock AgentCore | Managed runtime, identity, gateway, memory, observability and policy services. | Runtime acknowledgement maps to EXECUTION_ACKNOWLEDGED until authoritative state is reconciled. | Official source ↗ |
| Enterprise AI governance · ServiceNow AI Control Tower | Central inventory, governance, lifecycle, risk and value management for AI assets. | Portfolio governance does not replace a domain-specific authoritative ending test. | Official source ↗ |
| Agent interoperability · A2A | Task and message interoperability, including terminal task lifecycle states. | TASK_STATE_COMPLETED maps to EXECUTION_ACKNOWLEDGED, not VERIFIED_FINAL. | Official source ↗ |
| Tool interoperability · MCP | Tool discovery and invocation with structured success and error results. | A successful tools/call result is evidence of tool execution, not authoritative outcome. | Official source ↗ |
| Tracing and controls · OpenAI Agents SDK | Tracing captures agent events; guardrails can validate or block inputs, outputs and tools. | These controls provide evidence and prevention surfaces; post-action authoritative reconciliation remains distinct. | Official source ↗ |
| Domain trust · protected evaluator Agent Trust | Identity and trust signals for agentic commerce and financial interactions. | Trust signals inform authority and risk; the required authoritative ending remains workflow-specific. | Official source ↗ |
Falsification conditions
What would disprove or narrow the thesis.
The category thesis must change if public evidence establishes an equivalent universal post-action reconciliation layer, if buyers find execution acknowledgement materially sufficient, if authoritative sources cannot be bound reliably, or if independent reproduction fails.
Equivalent control already exists
A public implementation proves the same authority-to-reopen semantics across platforms.
No material buyer problem
Qualified buyers do not distinguish task completion from authoritative outcome.
No reliable authority source
The target domain lacks an identifiable source empowered to establish the ending.
No reproducible control effect
Independent testing cannot reduce false FINAL without unacceptable cost or delay.