Skip to main content

W1 — Exposure Evaluation Logic

Implementation status: Spec complete. 8 finding types defined, evaluator rule patterns designed. Implementation planned — see W1 unified implementation plan.

W1 emits findings; the UI presents the path’s current exposure as the set of active findings.

This document defines the deterministic evaluation logic W1 applies to the normalized execution/authority path graph.

All outcomes must be evidence-backed. Unknown states are first-class results.

1. Authority path Qualification​

A workload qualifies as execution-backed when:

  1. It is an inventoried workload.
  2. It can execute without an active human session.
  3. A first-party execution_evidence record exists.
  4. The execution record can be deterministically linked to the specific workload construct (or its execution identity) without heuristic matching.

If condition (4) cannot be satisfied:

execution_status = unproven

Execution status is binary: proven or unproven.

2. Identity Binding Resolution​

It applies to the authority path.

For each workload:

Traverse:

workload → RUNS_AS → identity

Identity binding is valid when:

  • The relationship is deterministically recorded,
  • The identity object is uniquely identifiable.

If linkage fails:

identity_status = unknown

No heuristic correlation is permitted.

3. Data Reachability Computation​

Reachability is computed via bounded traversal.

Primary path:

workload → RUNS_AS → identity
identity → HAS_ROLE → GRANTS → APPLIES_TO → resource

Rules:

  • Only directly recorded edges are traversed.
  • No inference of implied permissions.
  • No RBAC inheritance modeling beyond explicit assignments.
  • No multi-hop AUTHENTICATES_TO traversal.

Each reachable resource is mapped to:

  • business_domain
  • sensitivity

Output:

  • Domain-level classification only.
  • If no deterministic resource mapping exists: data_origin = unknown.

4. Egress Determination​

Egress is determined via first outbound boundary traversal.

Primary path:

workload → INVOKES → connection → USES → credential → AUTHENTICATES_AS → identity

or first observable outbound endpoint configuration.

Classification is based on endpoint host/base URL:

  • LLM
  • External
  • Internal
  • Unknown

Constraints:

  • Stop at first outbound boundary.
  • Do not traverse further AUTHENTICATES_TO hops.
  • No payload/header inspection.
  • If no outbound configuration is observable: egress = none_observed.

5. Ownership Evaluation​

Ownership state is derived from:

entity → OWNED_BY → owner

Ownership is:

  • valid (at least one active identifiable owner)
  • invalid (no active owners or any one active owner is missing)
  • ambiguous (group-only ownership without individual accountability)
  • unknown (insufficient deterministic metadata)

Ownership evaluation is independent of execution status.

6. Deterministic Finding Triggers​

W1 emits deterministic findings derived from graph state.

W1 emits only a restricted set of finding types appropriate to autonomous execution (e.g., unproven execution, unknown identity binding, sensitive reachability, outbound boundary classification, ownership invalid/ambiguous/unknown).

Each finding must:

  • Be derived solely from deterministic graph evidence.
  • Include a clear, plain-language deterministic explanation.
  • Explicitly declare any evidence gaps.
  • Be persisted using the platform’s canonical Finding model.

W1 does not introduce new persistence semantics or lifecycle behavior beyond the platform Finding model.

6A. Evidence Declaration Requirement​

Every W1 finding must explicitly declare whether required evidence was:

  • Available and used,
  • Partially available,
  • Unavailable, or
  • Not applicable to the evaluated surface.

W1 must never imply evidence availability when it was not accessible.

Unknown or unavailable evidence is a valid and explicit outcome.

7. Triage Grouping​

Risk grouping is a deterministic roll-up of active findings.

Inputs:

  • Execution status
  • Data domain sensitivity
  • Egress classification
  • Ownership status

Grouping does not replace canonical findings and does not introduce new risk semantics.

8. Unknown Handling​

At any stage where:

  • Required edges are absent,
  • Deterministic identifiers cannot be joined,
  • First-party artifacts are unavailable,

The system must:

  • Stop traversal,
  • Emit explicit unknown state,
  • Avoid inference.

Unknown is not an error condition.

9. Architectural Constraint​

W1 may compute derived reachability or authority path relationships ephemerally during evaluation.

W1 does not require:

  • execution_chains persistence,
  • Chain versioning,
  • Composition fingerprints,
  • Temporal drift analysis.

Those capabilities belong to future wedges.