Skill: Determinism Engineering

Name: Determinism Engineering
Rating: 92
Author: Kristopherlb

Intent

•Enable an agent to identify and eliminate sources of nondeterminism so that identical inputs produce identical graph, intents, policy outcomes, and plan artifacts.
•Make determinism a mechanically enforced property (not a convention).

•In-scope: stable IDs, canonical ordering, canonical serialization expectations, deterministic evaluation phase rules, and determinism threat modeling.
•Out-of-scope: performance optimization beyond what’s needed for determinism; backend-specific diffing behavior.

•Stable identifiers: deterministic IDs for nodes/edges and derived artifacts.
•Canonical ordering: deterministic ordering rules for sets/maps/lists and traversal/phase execution.
•Canonical serialization: stable encoding of graph/artifacts for diffing and caching.
•Determinism hazards: time, randomness, iteration order, locale differences, external I/O, environment variability.

•Determinism specification: canonical ordering rules, stable ID rules, and canonical serialization expectations (versioned).
•Evaluation phase model: what phases exist conceptually (graph enrichment, binding, policy evaluation, artifact emission) and their ordering constraints.
•Golden cases: representative graphs + expected stable outputs.
•Environment invariants: what must be pinned/provided (e.g., pack selection, environment name) and what must not influence outputs.

•
Structured outputs (minimum):
- •Determinism risk inventory: identified nondeterminism hazards with scope and impact classification.
- •Canonicalization requirements: explicit ordering/ID rules relied on or strengthened.
- •Reproducibility statement: what inputs define the output (and what inputs must not).
•
Human-readable outputs:
- •Explanation of why determinism hazards matter and how the acceptance criteria are satisfied (tied to provenance).

•Replayable: the same inputs can reproduce outputs without relying on external state or time.
•Stable diffs: output changes reflect semantic changes, not incidental ordering or representation churn.
•No side effects: evaluation does not perform network calls, shell execution, or implicit discovery that changes results.

•Snapshot stability: byte-for-byte stable outputs for identical inputs (deny if violated).
•Property tests: repeated runs (and reordered inputs) produce identical results (deny if violated).
•Nondeterminism lint: detection of external I/O/time/randomness usage in determinism-critical paths (deny if violated).

•Forbidden: “fixing” nondeterminism by loosening tests or ignoring reorder-only diffs.
•Forbidden: introducing hidden input channels (environment variables, host metadata) that affect outcomes without being declared.
•
Escalation (HITL required):
- •Any change to stable ID rules, canonical ordering, or canonical serialization semantics.