How the deterministic decisions are made
The tools apply published categorical rules in a fixed order. They do not call a generative model, inspect task text, rank vendors, or let a score override a safety boundary.
AI Task Fit Check
Task Fit methodology task-fit-1.0.2 answers five questions: whether AI should influence the task, the least complex suitable pattern, maximum autonomy, controls required before use, and the evidence, review, stop, and escalation instructions.
Decision order
- Validate required answers and reject impossible combinations.
- Collect every rule finding activated by those answers.
- Apply exclusions from the broadest boundary to the narrowest pattern.
- Apply the most restrictive autonomy ceiling.
- Combine and de-duplicate every required control.
- Select the least complex remaining solution pattern.
- Calculate explanatory fit and inherent-risk bands.
- Select the disposition.
- Assemble the rationale and the appropriate pilot or next-safe-step, evaluation, review, stop, and escalation instructions.
Exclusions, unavailable controls, and autonomy ceilings take precedence. Benefit, volume, and explanatory scores cannot override them.
What rule findings can do
- Exclusion
- Prohibits the requested workflow, one or more patterns, or all AI influence.
- Required control
- Permits the workflow only when a named safeguard is active before use.
- Autonomy ceiling
- Limits the most consequential action AI may take.
- Warning
- Adds explanation without changing the decision by itself.
Four dispositions
- Proceed: a pattern is available without an additional task-specific control or a lower autonomy ceiling.
- Proceed with controls: the same task boundary remains valid, but named safeguards or lower autonomy are mandatory.
- Redesign: the requested workflow is blocked. No pattern is recommended, the current workflow must not be piloted, and the materially changed task must be assessed again.
- Do not use AI: AI must not perform, draft, rank, summarize, recommend for, or otherwise influence the current task.
Risk and autonomy definitions
The risk band describes the task before controls. This version does not calculate residual risk, and controls do not lower the displayed band.
- Low: minor, visible, and reversible errors.
- Medium: material rework, confusion, cost, or limited external impact.
- High: material effects on people, money, rights, operations, or reputation.
- Critical: severe or irreversible harm, or improper influence over a high-impact decision.
The autonomy ceiling can allow, from least to most influence:
- No AI influence
- Draft for human review
- Recommend to a responsible person
- Decide with required approval
- Take a bounded, reversible action
Pattern selection
The engine selects the least complex pattern that can satisfy the task's evidence, control, and action requirements: human-only, conventional software, direct prompt, prompt with supplied context, retrieval-grounded workflow, bounded tool workflow, or bounded agent. More autonomy or technical complexity is rejected when a simpler pattern is sufficient.
Product-policy boundaries
Version 1 contains two explicit Use AI Clearly house-policy hard stops: material employment screening or selection, and operational assistance intended to steal credentials or account access. These are conservative product rules, not a complete statement of the law. Selecting “neither” does not mean a task is legally, professionally, or organizationally permitted.
The check relies entirely on the reader's categorical answers. It does not inspect the actual workflow, evidence, people, systems, policies, or jurisdiction. Input and result schema version is 1.0.0. Review examples calculated by the current engine.
Task Fit rule tables (task-fit-1.0.2)
Plain-language statements of the rules implemented in methodology task-fit-1.0.2. Codes match the result export. Open a table to inspect it.
1. Product-policy hard stops (checked first)
| When you answer | Finding | Result |
|---|---|---|
| The product boundary is material employment screening or selection | policy.high-impact-employment-decision | Exclusion of all AI influence. No other rule is evaluated. Disposition Do not use AI, pattern human-only, autonomy “No AI influence”. |
| The product boundary is operational help stealing credentials or account access | policy.credential-theft-assistance |
2. Exclusions, required controls, and autonomy ceilings
A qualified reviewer is required when the consequence is high or critical, verification needs qualified review, the work is regulated or high-impact, or the intended autonomy is “decide with approval” or “bounded action”. The reviewer counts as qualified only with domain competence, evidence access, error detection, review time, and authority—plus a responsible role for regulated or high-impact work.
| Trigger | Finding | Effect and scope | Ceiling | Controls named |
|---|---|---|---|---|
| Confidential or restricted data; approved environment unavailable | data.approved-environment-unavailable | Exclusion of the requested workflow | None | Approved environment |
| Confidential or restricted data; environment available or addable before use | data.approved-environment-required | Required control | — | Approved environment, data minimization |
| Reviewer required but unavailable or not qualified | review.qualified-review-unavailable | Exclusion of the requested workflow | None | Qualified review |
| Reviewer required and qualified | review.qualified-review-required | Required control | — | Qualified review |
| External action is irreversible | action.irreversible-autonomy | Exclusion of the requested workflow | None | Human approval, staging |
| Reversible external action; any of the seven action controls unavailable | action.controls-unavailable and action.no-safe-pattern | Excludes tool workflow and bounded agent, and the requested workflow | None | The unavailable action controls |
| Reversible external action; all seven controls available or addable | action.controls-required | Required control | — | Human approval, preview before write, permission boundary, rollback, monitoring, stop condition, escalation |
| Verification unavailable, and the work is regulated or high-impact or the consequence is high or critical | verification.unavailable-high-impact | Exclusion of the requested workflow | None | — |
| Current information required | evidence.current-sources-required | Excludes direct prompt and prompt with context | Recommend | — |
| Current information required | evidence.current-source-citations | Required control | — | Source citations |
| Traceability required, or an audit trail required | evidence.traceability-required | Required control | — | Source citations |
| Stable deterministic rules could solve the task | pattern.deterministic-rules-sufficient | Excludes every model-based pattern | — | — |
| Context sufficiency unknown and verification moderate | autonomy.cap-draft | Autonomy ceiling | Draft | — |
| Regulated or high-impact work with intended autonomy above “recommend” | autonomy.cap-recommend | Autonomy ceiling | Recommend | — |
| Supplied context sufficient and current information not required; verification easy; or direct prompt selected | pattern.supplied-context-sufficient, warning.verify-before-use, pattern.direct-prompt | Warning only | — | — |
The final ceiling is the lowest of your intended autonomy and every finding’s ceiling. Any workflow exclusion sets it to “No AI influence”.
3. Pattern candidates
The first matching row supplies candidates in order. The engine selects the first candidate that no finding excludes; if none remains, the result is Redesign.
| When | Candidates in order |
|---|---|
| An all-AI exclusion applies | Human-only |
| A requested-workflow exclusion applies | None |
| Stable rules could solve the task | Conventional software |
| The outcome is to act | Tool workflow, bounded agent. For recurring, high-volume work verified by qualified review: bounded agent, tool workflow |
| Current information required, or traceability required while current information is at least helpful | Retrieval-grounded workflow, prompt with context, direct prompt |
| Supplied context sufficient and the outcome is to summarize, classify, or analyze | Prompt with context, direct prompt |
| Otherwise | Direct prompt, prompt with context |
4. Disposition, risk band, and fit band
| Condition | Disposition |
|---|---|
| Any all-AI exclusion | Do not use AI |
| Any requested-workflow exclusion, or no pattern remains | Redesign |
| Any required control, or a ceiling below your intended autonomy | Proceed with controls |
| Otherwise | Proceed |
| Input | Score |
|---|---|
| Consequence minor / material / high / critical | 0 / 2 / 3 / 4 |
| Confidential data | At least 2 |
| Restricted data, regulated or high-impact work, or any external action | At least 3 |
| Irreversible external action | 4 |
| Band | 0–1 low, 2 medium, 3 high, 4 critical |
| Condition | Fit band |
|---|---|
| Do not use AI, or human-only pattern | Avoid |
| Redesign | Avoid at risk score 4; otherwise limited |
| Conventional software | Limited |
| Risk score 3 or more, except a tool workflow | Limited |
| Required controls or a lowered ceiling | Assisted |
| Otherwise | Strong |
Value, feasibility, and evidence-need scores are explanatory. They do not change the pattern, controls, ceiling, or disposition.
Output Verification Builder
Methodology output-verification-1.0.1 creates an acceptance and review plan from categorical task requirements. It determines:
- whether a review checklist can be generated, requires elevated review, or is blocked;
- sample or full review coverage;
- the minimum reviewer level;
- required checks, evidence, approvals, and stop conditions; and
- capability or use-boundary blockers that cannot be overridden.
The builder does not receive or evaluate an actual output. It cannot certify correctness, authorize a high-impact decision, or authorize automatic external action. It uses input and result schema version 1.0.0.
Representative checklists show the ready, elevated-review, and blocked boundaries.
Output Verification rule tables (output-verification-1.0.1)
Plain-language statements of the rules implemented in methodology output-verification-1.0.1. Codes match the checklist export.
1. Full or sample review coverage
Review every output when any of these is true: you prefer full review; the output goes to an external audience or could affect a person in a high-impact decision; material calculations are required; the consequence is material, high, or critical; completeness is strict; any policy constraint applies; or the output supports a decision or would authorize an external action. Otherwise a documented sample may be reviewed, with a warning that sampling can miss failures.
2. Minimum reviewer and escalation
| Minimum reviewer | Required when any of |
|---|---|
| Qualified approver | Critical consequence; regulated or professional constraint; high consequence and the output supports a decision; or the output would authorize an external action |
| Qualified reviewer | High consequence; current facts; material calculations; any policy constraint; or the output supports a decision or external action |
| Task owner | None of the above |
Above task owner, the checklist adds qualified-reviewer confirmation and escalation of unresolved reviewer disagreement. The one exception is an organizational-policy-only reference output with minor consequence, no factual claims, and no calculations. A preference for sampling or a lighter reviewer cannot lower these requirements.
3. Capabilities and blockers
Any blocker produces Review checklist blocked. A positive answer elsewhere cannot override it. With no blocker, sample coverage and a task-owner reviewer give “Review checklist generated”; anything else gives “generated with required review”.
| Blocker | Condition |
|---|---|
blocker.qualified-reviewer-unavailable | The minimum reviewer is above task owner and no reviewer is available before use with domain competence and the ability to detect errors; a qualified approver must also have authority to decide |
blocker.evidence-access-unavailable | The reviewer cannot inspect the evidence and sources, and factual claims or a source standard are required or the minimum reviewer is above task owner (even with no factual claims) |
blocker.calculation-review-unavailable | Material calculations are required and the reviewer cannot independently review them |
blocker.policy-review-unavailable | A policy constraint applies and the reviewer cannot review it |
blocker.approval-unavailable | A policy constraint applies, or the output supports a decision or external action, and the required approver cannot decide before use |
blocker.high-impact-decision | The output could affect a person in a high-impact decision and supports that decision—whoever reads it |
blocker.automatic-external-action | The accepted output would automatically authorize an external action, reversible or not |
The builder records six reviewer capabilities, and each can block. Evidence access, calculation competence, and policy competence map to their own blockers. Missing domain competence or error detection above task owner, or missing authority when a qualified approver is required, produces blocker.qualified-reviewer-unavailable. Answer “No” to any capability that is uncertain.
Review and validation status
VMCSoft Digital owner-reviewed and approved the original methodology, fixtures, implementation alignment, and editorial release (output-verification-1.0.0); that approval is a historic record and does not by itself cover later versions. Theoutput-verification-1.0.1 conformance correction enforces existing taxonomy requirements without changing the input schema; this page does not record a separate owner acceptance of it. Automated rule and fixture tests verify deterministic behavior. Independent domain review and moderated target-user validation have not been completed, so the site does not claim either.
Owner approval is not legal, medical, financial, employment, regulatory, or safety approval. See the editorial review record and report a safety or methodology concern.
See the public changelog for release and methodology history.