IDG Boundary Patterns
This page is optional. It is not a set of requirements, and not a reference implementation. It provides boundary patterns that reveal where determinability breaks and why an IDG blocks passage.
Not examples to follow.
Treat these as boundary lenses:
keep the structure, adapt the details locally.
Related repositories (pattern workbench):
interface-determinability-gate
Structural notice — non-direct execution boundary
In systems where non-deterministic or responsibility-bearing judgments are involved, this library assumes the existence of a structural boundary that prevents such judgments from being passed directly into execution, control, or promotion layers.
This boundary is structural, not algorithmic. It does not decide outcomes, enforce policies, or optimize results. Its sole role is to interrupt passage when determinability, ownership, or scope cannot be made explicit before downstream effects occur.
Whether the judgment source is human, rule-based, or AI-driven is intentionally left unspecified. Only the placement of the boundary is fixed:
- No judgment is promoted to an execution-bearing domain without passing a gate
- No ambiguity is resolved implicitly or retroactively
- No component beyond the gate assumes responsibility by accident
The patterns below illustrate where such boundaries commonly appear—not how they are implemented.
Minimum gate conditions (ambiguity signals)
IDG does not decide outcomes. It blocks when any of the following is ambiguous.
- Auditability: Are the input and its grounds traceable later (in a stable form)?
- Explainability: Is the reject reason transferable to a third party (as a structure, not a story)?
- Ownership: Is responsibility non-ambiguous (no implicit owner, no shared fog)?
- Scope: Does the interaction stay within the declared boundary (no overreach)?
Pattern 1 — API Boundary Gate
Boundary: External API client → Internal service
Input
- payload fields, headers, auth context, rate-limits
- optional AI-generated fields (if any) must be marked as candidates
Why this boundary exists
- Prevent silent coupling (unknown fields become “features”)
- Prevent responsibility leakage (who owns the request meaning?)
What becomes ambiguous
- Unspecified parameters that change meaning downstream
- Implicit ownership (“the client assumed you would interpret this”)
Gate conditions (ambiguity signals)
- Scope: fields outside declared contract → block
- Ownership: no accountable caller identity / role → block
- Auditability: missing correlation / request fingerprint → block
- Explainability: reject reason cannot be stated as contract clause → block
Fail-closed behavior
- Return
4xxwith explicit reason code and boundary reference - Do not “fix” or “guess” missing fields
Non-goals
- Not validating business correctness
- Not deciding acceptance policies beyond the declared boundary
Pattern 2 — AI Input Gate (Prompt / Context Boundary)
Boundary: Human/system context → AI model input
Input
- prompt, retrieved context, tool outputs, attachments
- policy constraints and declared scope of assistance
Why this boundary exists
- Prevent “probabilistic drift” from entering a responsibility-bearing flow
- Keep the model inside the declared assistive role
What becomes ambiguous
- Facts vs. guesses mixed without markers
- Missing provenance of retrieved snippets
- Requests that implicitly ask for decisions / authority
Gate conditions (ambiguity signals)
- Auditability: no source/provenance for key claims → block
- Scope: request exceeds declared assistive scope → block
- Ownership: user seeks delegation of responsibility → block
- Explainability: cannot express refusal in boundary terms → block
Fail-closed behavior
- Return “cannot proceed” with a boundary reason (no speculative completion)
- Request missing provenance or explicit ownership statement
Non-goals
- Not maximizing helpfulness
- Not producing a “best guess” when determinability is broken
Pattern 3 — OT → IT Data Handoff Gate
Boundary: Physical observation (OT) → Business artifact (IT)
Input
- measurements, timestamps, collection conditions
- metadata mapping (tag, unit, equipment context)
Why this boundary exists
- Prevent meaning injection from becoming “fact” by accident
- Preserve where responsibility begins (OT facts vs IT interpretations)
What becomes ambiguous
- Unit / scale drift, sensor replacement, wiring changes
- Interpretation promoted as “official KPI” without owner
- Scope creep (using OT data to justify decisions without resolution)
Gate conditions (ambiguity signals)
- Auditability: missing collection conditions / fingerprint → block
- Scope: attempts to attach decision meaning at handoff → block
- Ownership: KPI promotion lacks explicit owner → block
- Explainability: cannot explain the mapping change as a tracked artifact → block
Fail-closed behavior
- Quarantine the record / mark as non-promotable
- Allow downstream to see “blocked” reason without auto-correction
Non-goals
- Not deciding correctness of the physical world
- Not replacing field verification or RP for promotion
Pattern 4 — Contractual Boundary Gate (Clause Pattern)
Boundary: Vendor deliverable → Client responsibility domain
Input
- deliverables (reports, AI outputs, configs), acceptance criteria, scope
- ownership statements and escalation path
Why this boundary exists
- Prevent “implicit acceptance” from becoming responsibility
- Make rejection explainable without arguing about intent
What becomes ambiguous
- Who owns the final decision when AI is involved
- What is inside/outside contract scope
- Whether evidence is auditable after delivery
Gate conditions (ambiguity signals)
- Ownership: no named accountable role for promotion → block
- Scope: deliverable exceeds declared boundary → block
- Auditability: missing evidence/provenance package → block
- Explainability: reject reason cannot be stated as clause reference → block
Fail-closed behavior
- Reject with explicit clause/boundary reference and reason code
- Require explicit re-submission with ownership + provenance
Non-goals
- Not evaluating quality of the vendor’s “best effort”
- Not turning IDG into governance—only a rejectable boundary