SOLIFAN designs and implements how organizations decidewhat to delegate to AI, how far that delegation should extend,and how those decisions are carried into operations.ReadinessOps is the foundation of this work.
This guide follows the path from new evidence to an AI-generated proposal, human judgment, explicit publication, and execution controlled by the published boundary.
01Evidence
02AI proposal
03Human judgment
04Publication
05Execution gate
06Trace
AI proposes. People decide. Publication defines the official state. Published boundaries control execution.
MODULES
M0 / OVERVIEW
See the operating model
What should be delegated to AI—and how far?
Understand how SOLIFAN, ReadinessOps, and the decision-to-execution boundary fit together.
CONTROL MAP
AI can
Produces evidence-grounded proposals and explains differences.
Human controls
Owns delegation boundaries and official decisions.
System enforces
Separates and enforces state transitions and execution conditions.
M0 / STEP 101 / 06
Start with the delegation question
Question for this step
Are you deciding how to delegate before choosing what AI can do?
ReadinessOps is not primarily a model-comparison exercise. It defines an operable boundary: what AI may propose, when a decision returns to a person, and what must never be executed by AI.
Control invariant
Capability is not authorization.
Expected control check
Delegated, review-required, and prohibited work are distinguishable.
M0 / STEP 202 / 06
Understand SOLIFAN and ReadinessOps
Question for this step
Who designs this system, and for what purpose?
SOLIFAN designs and implements how organizations decide what to delegate to AI, how far that delegation should extend, and how those decisions are carried into operations. ReadinessOps is the foundation of this work.
Control invariant
Present SOLIFAN as the provider and ReadinessOps as its operating foundation.
Expected control check
The provider–foundation relationship can be explained in one sentence.
Have proposal, judgment, and enforcement been collapsed into one actor?
AI proposes, people own the delegation boundary and official judgment, and the system permits execution only from published state. Separating these responsibilities prevents conversational assent or model output from becoming authority.
Control invariant
Do not overlap AI, human, and system authority.
Expected control check
Each state change has an unambiguous owner.
M0 / STEP 404 / 06
Follow the lifecycle
Question for this step
In what order does new evidence become an execution condition?
Evidence is inspected, its impact is established, AI prepares a reassessment, and a person decides. The exact revision is then explicitly published; a gate checks the published boundary, and the result is connected through one trace.
Control invariant
Do not skip stages; keep publication between judgment and execution.
Expected control check
The order is Evidence → Assessment → Review → Publish → Gate → Trace.
Are verified paths separated from organization-specific production requirements?
The public implementation is a reference that verifies design principles and control behavior. Production use additionally requires organization-specific RBAC, retention, monitoring, tenant isolation, incident response, and legal or regulatory controls.
Control invariant
Never present a reference build as production-ready.
Expected control check
VERIFIED and PRODUCTION REQUIREMENT are labeled separately.
The process begins with new evidence. The next module shows how change becomes governed input through inspection, versioning, and impact classification rather than being consumed as an unqualified prompt.
Control invariant
Reassessment begins with identifiable evidence and a revision.
Expected control check
Evidence intake is clearly the next control point.
M1 / EVIDENCE
Turn change into governed evidence
What changed, and what could it affect?
Inspect new evidence and establish its revision and impact scope.
CONTROL MAP
AI can
Organizes the meaning and possible impact of inspected evidence.
Human controls
Defines scope and materiality criteria.
System enforces
Enforces pre-LLM inspection, versioning, and suspension transitions.
M1 / STEP 101 / 06
Receive new evidence
Question for this step
Is change received as a traceable unit?
Policy, contract, evaluation, and operational changes are registered as evidence with origin, time, and scope. The input is fixed as the basis for later judgment instead of overwriting current state.
STATE
BeforeCURRENT / NO NEW EVIDENCE
AfterCURRENT + NEW EVIDENCE
Control invariant
Evidence carries provenance, receipt time, and scope.
Expected control check
New evidence and existing Current remain separate.
In the verified Google Cloud path, evidence is inspected by Model Armor before AI processing. Unsafe input is blocked and never reaches the model or downstream agents.
STATE
BeforeRECEIVED
AfterINSPECTED or BLOCKED
Control invariant
Inspect pre-LLM and never override a BLOCK downstream.
A new governed revision is created from inspected evidence. Earlier revisions remain in history so differences can be reviewed, reproduced, and rolled back.
STATE
BeforeREVISION N
AfterREVISION N + PROPOSED N+1
Control invariant
Record change as a new revision; never edit Current in place.
Is the reassessment scope clear when evidence is received?
Evidence impact is established first to limit the boundaries, agents, and decision items that must be reassessed. The system does not send everything to a model while scope remains ambiguous.
STATE
BeforeIMPACT UNKNOWN
AfterIMPACT ESTABLISHED
Control invariant
Derive reassessment scope from established evidence impact.
Does an agent keep running while a material assumption is changing?
When material drift affects the existing safety basis, the affected agent moves to SUSPENDED before reassessment. Suspension is a control state that pauses execution pending review—not failure or deletion.
STATE
BeforeREADY
AfterSUSPENDED
Control invariant
Material drift triggers suspension before continued execution.
Expected control check
The affected agent is SUSPENDED and the reason is traceable.
Can the input state be reproduced before processing continues?
The revision ID, evidence impact, affected agent status, and trace ID are checked together. Reassessment does not proceed if the control state is incomplete.
Control invariant
Do not begin reassessment from incomplete control state.
Expected control check
Revision, impact, status, and trace refer to the same event.
Combine multiple perspectives into an evidence-grounded proposal for human review.
CONTROL MAP
AI can
Reassesses the scoped change and proposes a decision pack.
Human controls
Retains evaluation criteria, acceptance thresholds, and final judgment.
System enforces
Checks grounding and state transitions; prevents self-approval.
M2 / STEP 101 / 06
Start from evidence impact
Question for this step
Is the task sent to AI limited to what changed?
The reassessment request is assembled from established evidence impact and only the evidence it needs. This limits the model’s opportunity to reinterpret unrelated current state.
STATE
BeforeIMPACT ESTABLISHED
AfterASSESSMENT SCOPED
Control invariant
AI scope cannot exceed the human-governed impact definition.
Expected control check
Inputs, scope, and exclusions are visible.
M2 / STEP 202 / 06
Run specialist agents
Question for this step
Are different responsibilities being mixed into one model answer?
Governance, value, and routing perspectives are assigned to specialist agents that assess the same scoped evidence. Specialization does not grant execution authority.
STATE
BeforeSCOPED
AfterSPECIALIST ASSESSMENTS
Control invariant
Specialized analysis does not distribute authority.
Expected control check
Each output has an identifiable role and evidence basis.
Are multiple assessments synthesized into a form a person can compare?
Specialist results are synthesized into a decision pack showing changes, evidence, risks, recommendations, and unresolved questions. It preserves decision-relevant differences—not only a conclusion.
STATE
BeforeSPECIALIST ASSESSMENTS
AfterDECISION PACK
Control invariant
Do not erase disagreement or unresolved issues during synthesis.
Can each material claim be traced back to supplied evidence?
Material claims in the decision pack are checked against evidence. Unsupported content is marked unresolved or sent for reassessment rather than presented to human review as established fact.
STATE
BeforeDRAFT PACK
AfterGROUNDED or NEEDS REVISION
Control invariant
Unsupported generation cannot become a premise for official judgment.
Expected control check
Every material claim has evidence or an unresolved marker.
Can the non-delegable boundary be stated at the action level?
AI may generate candidates, compare options, and organize evidence. It may not create human approval, publish official state, or perform protected execution.
Control invariant
Proposal capability never implies approval, publication, or execution authority.
Expected control check
Forbidden actions exist neither in the public UI nor its WebMCP tools.
M3 / REVIEW
Keep human authority explicit
Who is accountable for the official decision?
Review evidence and differences, then record a human decision with rationale.
CONTROL MAP
AI can
Presents candidates, differences, and evidence; does not decide.
Human controls
Edits the boundary and approves, returns, or rejects.
System enforces
Records role, exact revision, rationale, and time.
M3 / STEP 101 / 06
Review the proposal and evidence
Question for this step
Can reviewers see what changed—not just the recommendation?
The reviewer sees the decision pack, source evidence, differences from Current, and unresolved issues in one context. Material claims remain traceable beyond the summary.
Control invariant
Keep evidence and differences visible at the point of judgment.
Expected control check
Current, Proposed, and Evidence are distinguishable.
Can the delegation scope be revised as explicit conditions?
A person reviews action, tool, data, impact, and exception conditions, then narrows the delegation boundary, requires human review, or prohibits the action.
STATE
BeforePROPOSED BOUNDARY
AfterHUMAN-EDITED BOUNDARY
Control invariant
Boundary changes use human-readable conditions and explicit differences.
Expected control check
Before, after, and editor are visible.
M3 / STEP 303 / 06
Approve, return, or reject
Question for this step
Is the rationale preserved with the decision?
An authorized person approves, returns, or rejects the exact revision. Rationale and unresolved issues are recorded so the decision context can be reconstructed later.
STATE
BeforeREVIEW_REQUIRED
AfterAPPROVED / RETURNED / REJECTED
Control invariant
Bind every decision to an exact revision and accountable person.
Expected control check
Decision, revision, rationale, actor, and time are present.
Should current operating state change at the moment a person approves?
ReadinessOps treats approval and publication as separate state transitions. Approval records that a person accepts a proposal. Current State changes only when the exact revision is explicitly published.
Current therefore remains unchanged across this step.
STATE
BeforeREVIEW_REQUIRED / CURRENT N
AfterAPPROVED / CURRENT N (UNCHANGED)
Control invariant
Approval does not change Current State.
Expected control check
Proposal = APPROVED / Publication = NOT_PUBLISHED / Current State = UNCHANGED
Evidence, proposal, differences, and human rationale are connected through the same trace ID. A trace is not a single log line; it is the evidence chain for decisions and state transitions.
Control invariant
Never lose rationale or accountability from the decision record.
Publish the exact approved revision and update the active boundary.
CONTROL MAP
AI can
May explain publication impact; never publishes.
Human controls
Confirms the exact target and timing.
System enforces
Publishes only the exact revision and preserves history.
M4 / STEP 101 / 05
Make the publication decision
Question for this step
Should this approved change enter operations now?
After approval, timing, dependencies, affected scope, and rollback conditions are checked. This separates content acceptance from the decision to change operations now.
STATE
BeforeAPPROVED / NOT_PUBLISHED
AfterPUBLICATION AUTHORIZED
Control invariant
Keep approval and the publication decision separate.
Expected control check
The target revision and publication conditions are explicit.
M4 / STEP 202 / 05
Publish the exact revision
Question for this step
Can the system publish only the exact revision that was approved?
Publication verifies that stored human approval matches the exact revision. It does not substitute another revision or infer authorization from conversation.
STATE
BeforeAPPROVED / NOT_PUBLISHED
AfterPUBLISHED
Control invariant
The approved and published revisions must match exactly.
Which boundary becomes authoritative after publication?
Publication creates a new ACTIVE boundary. The prior version remains in history as SUPERSEDED, and the execution gate consults only the new active version.
Publication and restoration to READY are separate state transitions. The agent resumes only after required checks complete and the published boundary is available to the gate.
STATE
BeforeACTIVE BOUNDARY / SUSPENDED
AfterACTIVE BOUNDARY / READY
Control invariant
Do not combine official-state change with execution restart.
Expected control check
Boundary = ACTIVE / Agent = SUSPENDED until reactivation checks pass
Can this decision be reconstructed when the next evidence arrives?
Evidence, revision, AI assessment, human judgment, publication, and the active boundary remain connected in history. The next change starts a new revision without overwriting Current.
Control invariant
Separate Current from history while keeping decisions traceable.
Expected control check
Current, Superseded, and Proposed remain distinguishable.
Use identity and a deterministic gate to execute only inside the published boundary.
CONTROL MAP
AI can
May propose or request a protected action; cannot execute it directly.
Human controls
Owns the published boundary and exception handling.
System enforces
Deterministically checks identity, READY, active boundary, and duplicates.
M5 / STEP 101 / 06
Request a protected action
Question for this step
Are the action and its impact explicit before authorization?
An execution request identifies the action, tool, target data, intended effect, impact scope, and requester. A natural-language instruction alone is never execution authorization.
STATE
BeforeNO REQUEST
AfterPROTECTED ACTION REQUEST
Control invariant
Structure the request; never treat conversation as authority.
Expected control check
Action, tool, data, impact, and identity are present.
M5 / STEP 202 / 06
Evaluate with a deterministic gate
Question for this step
Is authorization based on published conditions rather than model self-assessment?
The gate deterministically checks agent READY status, the active boundary, requested action, data, and impact. An out-of-boundary action is DENIED even when the agent is READY.
STATE
BeforeREQUESTED
AfterPERMITTED or DENIED
Control invariant
Execution authorization is not delegated to model discretion.
Can the identity that performs analysis also execute protected actions?
The Google Cloud implementation rejects protected execution from Analysis Identity A. Analysis and execution are separated at the identity layer, beyond what a prompt or tool call can bypass.
STATE
BeforePERMITTED ACTION / WRONG IDENTITY
AfterDENIED
Control invariant
The proposing identity has no protected-execution authority.
Does the correct identity still recheck the boundary before execution?
Only Executor Identity B can reach protected execution. Even with the correct identity, the gate rechecks the active boundary and action conditions, executing only what is permitted.
STATE
BeforePERMITTED / EXECUTOR B
AfterEXECUTED
Control invariant
Require both identity authentication and boundary authorization.
Can evidence be followed all the way to the execution result?
The same trace ID links evidence intake, inspection, revision, AI assessment, human judgment, publication, gate decision, and execution result. Duplicate requests are identified and are not executed twice.
Control invariant
Do not fragment state transitions and outcomes across unrelated logs.
Has any boundary between proposal and execution been bypassed?
AI proposes. People decide. Publication defines official state. Published boundaries control execution. Together, these form the ReadinessOps operating model.
Control invariant
Preserve all four principles across UI, state, identity, and gate.
Review the Google Cloud architecture, verified path, safety behavior, and limitations with sources.
CONTROL MAP
AI can
Performs analysis and reassessment on Vertex AI.
Human controls
Retains responsibility for review and publication.
System enforces
Enforces boundaries through cloud services, identity, gate, and trace.
ARCH / STEP 101 / 06
See the four control planes
Question for this step
Can the implementation be explained by responsibility?
The architecture is read as four control planes: Evidence, Analysis, Human, and Execution. Even if services change, intake, analysis, judgment, and execution control remain distinct.
Control invariant
Show responsibility boundaries before product topology.
Expected control check
Evidence, Analysis, Human, and Execution are distinguishable.
The public README records this path: Evidence → Model Armor → governed revision / evidence impact → specialist reassessment → Human Review → Explicit Publish → deterministic gate → Trace.
This guide explains that verified path without adding unverified production claims.
Control invariant
VERIFIED is limited to what public evidence supports.
Expected control check
The guide sequence matches the public README’s golden path.
Are denial and retry paths verified alongside success?
Verified behaviors include pre-LLM blocking, denial of Analysis Identity, denial outside the active boundary, duplicate suppression, retry of the same revision, one trace ID, and a permitted path through Executor B.
Safety is demonstrated not only by what succeeds, but by what reliably stops.
Control invariant
Acceptance tests include both allow and deny behavior.
Expected control check
Check BLOCK, DENIED, duplicate, retry, trace, and permitted paths.
Are out-of-scope areas stated as clearly as verified behavior?
The public demo uses synthetic text evidence and contains no customer data or PII. General PDF processing and production enterprise connectors are out of scope. Production use requires additional organization-specific authentication, authorization, retention, monitoring, and tenant isolation.
Control invariant
Limitations sit beside verified facts, not in fine print.
Expected control check
Demo scope and production requirements remain distinct.
Can readers verify the guide against primary sources?
Readers can continue to the SOLIFAN official site, the Google Cloud and Snowflake repositories, the AI Delegation Boundary WebMCP implementation, and official Chrome and Firebase documentation.
Control invariant
Verifiable claims link to stable primary sources.
Expected control check
Each source exposes its name, URL, and review date.