Deterministic rules
Versioned policy logic, arithmetic, thresholds, or routing producing repeatable output from the same governed inputs.
Not labelled AI. Rule version, inputs, and reason exposed.
AI Governance
Use AI within defined authority, evidence, and review
Deterministic rules, approved machine-learning flags, governed assistance, and human-only decisions are four different things with four different authority levels. Each governed use states its purpose, sources, allowed output, human control, limitations, owner, status, and correction path.
Scope invariant
This page explains authority and evidence, not AI novelty. If a governed use cannot be described as a purpose, source boundary, allowed output, human control, limitation, owner, and status — it is not ready to be described at all.
Capability Taxonomy
Seven Categories — Starting With the One That Is Not AI
Relabelling ordinary rules as AI is a marketing choice with a governance cost: it makes an explainable system look opaque, and invites trust the mechanism has not earned.
Approved ML flagging
A governed statistical or machine-learning system identifies an anomaly, pattern, or signal-quality concern.
Flag only. No automatic guilt, payroll, discipline, or employment conclusion.
Governed AI assistance
A model retrieves, summarizes, transforms, or explains approved information within defined source, permission, tool, and output limits.
Source-linked, reviewable, non-authoritative.
Human-only outcome
A consequence that must be decided by an authorized person.
Outside autonomous authority entirely.
Prohibited use
A use that violates product invariants or approved policy.
Not configurable, not an add-on, not described as future capability.
Evidence-gated use
A proposed or evaluated use that has not passed release and operating gates.
Not current. Prerequisites and safe alternative stated, or absent.
Unavailable
No approved capability or evidence supports the request.
Routes to a deterministic or human process instead.
Why the first row matters
Time classification is deterministic and policy-bound. Calling it AI would make a reviewable rule engine harder to challenge, not easier.
Seven authority classes
| Authority class | Allowed behavior | Required human control |
|---|---|---|
| Inform | Present source-linked information or status. | User can inspect sources, freshness, and limitations. |
| Suggest | Offer a non-binding option or draft. | An authorized user reviews and chooses whether to act. |
| Flag | Identify a pattern or potential concern. | A flag opens review. It is not a conclusion. |
| Draft | Prepare text, summary, explanation, or workflow artifact. | A human approves any consequential communication or action. |
| Execute reversible administrative action | Only when explicitly approved, permissioned, and safely reversible. | Human initiation and confirmation, preview, audit, undo, and policy limits. |
| Consequential decision | Prohibited for autonomous AI. Payroll, discipline, employment, legal, eligibility, or comparable outcome. | An authorized human decision is required — always. |
| Irreversible or external action | Send, disclose, delete, publish, lock, or change authoritative records outside safe reversible bounds. | Human authorization plus explicit action-specific controls. Otherwise prohibited. |
Prohibited & Unavailable Uses
Nine Things ZoikoTime Will Not Build
These are not gaps awaiting a roadmap. They cannot be weakened by plan, configuration, or a hidden enterprise add-on, and changing any of them would require product, privacy, security, legal, ethical-design, and worker-trust review.
No screenshots, keystroke content, URL history, application-name monitoring, or clipboard collection under any tier or configuration.
Inference prohibitions
- Emotion, mood, personality, intent, loyalty, or honesty inference
- Productivity worth or protected-trait inference from workforce behavior
- Covert ranking, hidden risk scores, or worker comparison without legitimate purpose and approved authority
- Face recognition, voiceprint identification, or biometric categorization — no current claim exists, and none is implied
Authority prohibitions
- Automated payroll, discipline, termination, promotion, hiring, legal, compliance, or misconduct conclusions
- Autonomous access escalation, permission grant, evidence alteration, record deletion, incident closure, or policy override
- Training or provider reuse of customer or worker data outside approved contractual, privacy, and technical boundaries
- Fabricated sources, unsupported legal advice, guaranteed compliance conclusions, or claims of human-judgment equivalence
- Dark patterns that pressure users to accept AI, conceal a human alternative, or make correction and appeal harder
An unavailable request identifies its limitation and routes to a deterministic or human alternative. It is never presented as a coming feature.
Governance Lifecycle
Twelve Stages, With Evaluation Planned Before Results Exist
Stage 5 is the one most governance processes skip: the evaluation plan is created before anyone relies on an outcome metric, so the threshold is not chosen to fit the result.
| 01 | Register Proposed use, purpose, intended users, affected people, authority, owner. | Use-case owner |
| 02 | Classify Capability type, risk tier, prohibited-use proximity, reversibility, jurisdiction context. | AI governance |
| 03 | Verify data & provider Sources, permissions, retention, region, and training or use-of-data boundaries. | Privacy & Security |
| 04 | Define human controls Worker rights, fallback, accessibility, and safe failure behavior. | Product governance |
| 05 | Create the evaluation plan Before relying on any outcome metric — so thresholds are set before results are known. | Independent of the owner |
| 06 | Evaluate Task quality, groundedness, failure modes, fairness, privacy, security, robustness, accessibility, human factors, misuse. | Specialist reviewers |
| 07 | Review residual risk Limitations, evidence sufficiency, and operational readiness via eligible independent roles. | Independent review |
| 08 | Gate decision Approve, conditionally approve, reject, or keep evidence-gated — with a reasoned record. | Human only · separated from owner |
| 09 | Release In approved scope, with monitoring, rollback, incident, and support controls. | Operating owner |
| 10 | Monitor Sources, behavior, overrides, corrections, incidents, provider and tool health. | Operating owner |
| 11 | Re-evaluate After material change, incident, drift-like behavior, legal or policy change, or review date. | AI governance |
| 12 | Suspend, correct or retire When evidence or controls no longer support current operation. | Governance · emergency authority is broader |
Six risk tiers
G0
Not AI / deterministic
Versioned rules, arithmetic, or routing with no learned behavior.
Rule ownership, test evidence, explanation, change control, human review where consequential.
G1
Low-impact assistance
Drafting, retrieval, or summarization with no consequential authority and clear human review.
Source and permission controls, task evaluation, limitations, monitoring, user correction.
G2
Review-supporting signal
A flag or recommendation may shape an investigation but cannot decide a consequence.
Impact and fairness evaluation, visible reasons, reviewer training, correction, independent approval.
G3
High-impact context
Assistance used near payroll, discipline, employment, legal, or compliance decisions — humans retain authority.
Enhanced review, separation of duties, representative and legal review, staged release, strong monitoring.
G4
Prohibited autonomous consequence
AI determines or executes a consequential outcome, performs covert surveillance, or makes a prohibited inference.
Not permitted. No release path without fundamental product-policy change and full re-approval. No current claim.
G?
Unclassified
Purpose, authority, data, evaluation, or control is unresolved.
Evidence-gated or suspended. No current operation and no marketing claim.
Separation of duties
A use-case owner cannot independently approve a high-impact use. Evaluation author and final approver are separated where material conflicts exist. Emergency suspension authority is deliberately broader than release authority — it is easier to stop something than to ship it — subject to retrospective review. And provider commercial ownership never overrides evidence, safety, or rights review.
Public Capability Register
Every Governed Use, With Its Honest Status
If evaluation, monitoring, or source evidence is stale, incomplete, or conflicting, the capability reads Under Review, Suspended, Limited, or Unavailable — never silently Current.
CurrentLimitedUnder reviewSuspendedRetiredEvidence-gatedUnavailable
Capability typeAuthority classRisk tierProduct areaStatusEvidence levelOwnerLast reviewed
Time classification
PUBLICVersioned policy rules producing repeatable, explainable output.
- Type
- Deterministic — not AI
- Authority
- Inform
- Tier
- G0
- Status
- Current
Deterministic does not mean legally correct or consequence-ready without human review.
Source-quality flagging
PUBLICIdentifies stale, incomplete, or conflicting source conditions for review.
- Type
- Approved ML flagging
- Authority
- Flag
- Tier
- G2
- Status
- Current
False-positive and false-negative limitations documented in the use-case detail.
Anomaly flagging
PUBLICIdentifies a pattern that may warrant an authorized person's attention.
- Type
- Approved ML flagging
- Authority
- Flag
- Tier
- G2
- Status
- Current
Never a misconduct, fraud, payroll, or legal conclusion. No composite worker score.
Kairos retrieval & explanation
PUBLICRetrieves, summarizes, and explains governed data within the requester's existing permissions.
- Type
- Governed AI assistance
- Authority
- Inform · Suggest · Draft
- Tier
- G1
- Status
- Current
No completeness or legal-correctness guarantee. Cannot expand permissions or invent a source.
Drafted explanation text
CONTROLLEDPrepares an explanation draft for an authorized reviewer to adopt or reject.
- Type
- Governed AI assistance
- Authority
- Draft
- Tier
- G3
- Status
- Limited
High-impact context. A draft is never issued without human adoption and verification.
Autonomous approval of any kind
PROHIBITEDAI determining or executing a consequential outcome.
- Tier
- G4
- Status
- Prohibited
No release path exists. Not configurable, not on a roadmap, not an enterprise add-on.
Each public use-case detail states purpose, affected people, sources and permission boundary, allowed outputs, human controls, evaluation summary, limitations, owner, status, and correction path. Model, provider, tool, region, retention, and use-of-data conditions are disclosed at the appropriate public or controlled level — never invented for a marketing page.
Evaluation Evidence
Ten Dimensions, Six Result States
"Passed" always means passed within scope. It never proves universal performance, and the distinction is preserved in the state name itself.
| Dimension | The question it answers |
|---|---|
| Task quality | Does the use perform its defined task within approved scope? |
| Groundedness / source use | Are claims linked to approved, current, permissioned sources? |
| Failure modes | How does it behave with missing, stale, contradictory, adversarial, or out-of-scope input? |
| Fairness / differential impact | Do errors or burdens differ across relevant groups or contexts? |
| Privacy | Does it minimize data and respect purpose, permissions, retention, and provider limits? |
| Security / abuse | Can input, retrieval, tools, or output be manipulated or made to exfiltrate data? |
| Robustness / reliability | Does behavior hold across versions, load, provider failures, and environmental change? |
| Accessibility | Can users perceive, understand, operate, and correct the experience? |
| Human factors | Can reviewers understand, question, and override output without automation bias? |
| Misuse / prohibited use | Can the capability be repurposed for surveillance, discrimination, coercion, or unauthorized decisions? |
Six evaluation result states
Passed within scopeConditionalUnder reviewFailedNot evaluatedExpired / stale
Human factors is not optional
An evaluation that shows a model performs well but that reviewers defer to it uncritically has found a problem, not a success. Automation bias is measured, not assumed away.
Fairness has justified scope
Groups and contexts are selected only where legally, ethically, and analytically supportable. Sample, method, limitations, and excluded populations are reported — and no universal fairness claim is made.
Not evaluated is a real state
If no approved evidence exists for a claimed behavior, the capability is Evidence-Gated or Unavailable. Absence of evaluation is never treated as absence of risk.
Human Controls & Worker Rights
What a Person Can Do With an AI-Assisted Output
A right that exists only in policy is not a control. Each of these is an action available in the interface.
- Inspect — inputs, sources, freshness, scope, model or rule version, and stated limitations.
- Question — request more evidence, or mark the output insufficient.
- Correct — amend the record through the governed correction path, with the original preserved.
- Reject — dismiss a flag or decline a draft, with a reason recorded.
- Escalate — route to an independent reviewer where policy requires.
- Report — raise a concern about the AI-assisted output itself, not just the record.
No coercive framing
No language pressures a reviewer to accept a system output, and no interface makes accepting faster than examining. A human alternative is never concealed.
Worker-facing transparency
Where an AI-assisted output affects a worker's record, that person can see the relevant inputs, purpose, source, limitations, and status within role and policy scope — and can use correction, challenge, escalation, and appeal routes with full decision history.
Human-in-Command ControlsHuman-only outcome classes
- Payroll-impacting approval
- Disciplinary or misconduct conclusion
- Employment eligibility or termination
- Legal or regulatory conclusion
- Accommodation or protected-context outcome
- Appeal or grievance outcome
Each requires reviewer eligibility, separation of duties, evidence access, a reason, notification, a challenge route, and preserved history. Neutral pending-review states apply throughout.
Runtime monitoring & operating limits
CurrentObjective: detect when a governed use stops behaving as evaluated.
- Monitored
- Source availability and freshness, output behavior, override and correction patterns, incident signals, provider and tool health
- Override patterns matter
- A sharp rise in reviewers rejecting an output is a monitoring signal, not a reviewer problem
- Operating limits
- Scope, volume, and context limits are enforced rather than advisory
Limitations: monitoring coverage is not measured to a published standard, and we make no completeness claim. Where a signal is ambiguous, the capability moves to Under Review rather than continuing silently.
Incidents, correction & suspension
CurrentObjective: stop first, explain second — emergency suspension is intentionally easy.
- Incident categories
- Groundedness failure, prohibited-use proximity, privacy or security event, fairness concern, provider failure, accessibility barrier
- Actions
- Suspend, narrow scope, roll back, correct, replace, or retire
- Recorded
- Trigger, affected scope, mitigation, customer and worker communication, and retrospective review
Limitations: exploitable incident detail is never published. Unsafe or unsupported use-case claims may be removed from markup, search, and structured data immediately, followed by an attributable correction record.
Material Change & Re-evaluation
A provider change is a governance event
Triggers requiring re-evaluation
- Model, provider, or tool change
- Source, permission, or retention change
- Purpose, scope, or authority-class change
- Region or data-handling change
- Drift-like behavior or incident
- Legal, contractual, or policy change
- Review-date trigger
Six Gate Decisions
Each with a required record
ApprovedConditionally approvedEvidence-gatedRejectedSuspendedRetired
Conditional approval carries an automatic suspension trigger. If the condition lapses or its expiry passes without verification, the capability suspends itself rather than quietly continuing. That is the difference between a condition and a hope.
Retirement records the reason, any replacement, the effective date, data and artifact handling, and preserved public history.
Controlled AI Governance Review
Request Non-Public Governance Evidence
Model or system cards, evaluation summaries, provider information, and governance documentation, at a depth appropriate to your role and purpose.
Direct Answers
Eight AI Questions
No. Time classification is deterministic and policy-bound — versioned rules producing repeatable output from the same governed inputs, with the rule version, inputs, and reason exposed. Approved machine learning may flag anomalies or source-quality concerns for human review, but that is a separate capability with a separate authority class.