AGF

Decision Log

Terminology, structural, and positioning decisions with rationale — why AGF chose the path it did.

This is the public decision log. Each entry captures what was decided, why, and what alternatives were considered. The intent is that future contributors (and AGF itself, in retrospect) can reconstruct the reasoning behind every load-bearing choice.

The canonical DECISIONS.md on GitHub is the source of truth. This page summarizes by theme.

Vocabulary & Terminology

  • D1 (2026-03-19) — Replace "reliable" with "durable" in the five system properties. Durable adds a distinct architectural property (survives failures, preserves governance over time) that "reliable" — claimed by every framework — does not.
  • D2 (2026-03-19) — Replace "prior art" with "existing work". Removes patent/legal jargon for non-academic readers.
  • D7 (2026-04-21)Harness = governance envelope around an AI agent (source artifacts + runtime adapters, inseparable). Industry-converged definition; Agent = Model + Harness. Distinct from Wrapper (deployment mode), Hypervisor (AGT runtime), Test Harness (unrelated).

Positioning & Stack

  • D3 (2026-04-21)AGF Positioning Pillars (six). Category stake (Agentic Governance) · Four verbs (Synthesize / Unify / Prescribe / Operationalize) · OTAA invariant · Dual-form principle · Tempo taxonomy · NIST CSF parallel. Supersedes the earlier "integration layer" framing.
  • D4 (2026-04-21)Seven-layer stack. AGF at Layer 0 (architectural substrate), with OWASP at L1, MAESTRO/MS Failure Modes at L2, control catalogs at L3, ATF at L4, runtime references (AGT, CAF, Agent 365) at L5, FAIR/risk quantification orthogonal at L6. Resolves "AGF or X?" → "AGF and X, at different layers."

Gate Vocabulary & Audit

  • D5 (2026-04-21, refined 2026-04-22)Dual-form is load-bearing at gate boundaries. Every Gate Resolution and Domain Outcome MUST emit human-readable rationale AND a machine-readable GDR simultaneously.
  • D8 (2026-04-22)Gate Vocabulary Disambiguation: four-part vocabulary. Ring Control Signal (PASS / REVISE / HALT / GATE / DELEGATE / ERROR) · Gate Resolution (APPROVE / REJECT / MODIFY / DEFER / ESCALATE) · Domain Outcome (e.g., Authorized / Denied) · Audit Artifact (GDR). Each enum scoped to a different layer; nested per agentic-primitives.md.
  • D9 (2026-04-22)GDR is the canonical audit artifact. Reusable AGF primitive with polymorphic content. Required at every Gate Resolution and Domain Outcome. Five-state lifecycle. Schema at docs/schemas/gdr.yaml. Provides evidence in support of EU AI Act Art. 12, NIST AI RMF MANAGE 4.1, ISO 42001 documentation clauses.

Maturity Model

  • D6 (2026-04-21)L1 renamed "Awareness" → "Non-existent"; old L1 content (agent inventory, basic risk classification) absorbed into L2 Foundation. Honestly names the state most organizations are in. Complementary to (not substitute for) CSA ATF's 4-tier per-deployment scale.

Composition & Coherence

  • D10 (2026-04-24)Cohesion Pass v0.3 — three-part vocabulary reconciliation.
    • (a) Composition Patterns canonicalized at four (MVC / Validation Pipeline / GDF / Full Governed). "Phase 1–5" retired. Hardening stack (#11/#15/#18) demoted to a hardening posture modifier rather than a fifth pattern.
    • (b) Five-sense disambiguation for "governance"; qualifier conventions adopted across flagship pages.
    • (c) Three-layer observability disambiguation (Primitive #10 emits → Agentic Observability correlates → Observability Profile implements).

Licensing

  • D11 (2026-04-23)License: CC BY 4.0 (docs); Apache-2.0 / MIT for any reference implementations or tooling. Adoption > protection. Commercial value lives in tools, services, and consulting built around AGF, not in gating AGF itself.

Confidence Posture

Every load-bearing claim in AGF carries one of three confidence levels:

  • Established — evidence is clear, multiple sources would agree
  • Informed — AGF's synthesis; plausible but unverified at scale
  • Open — flagged but speculative; needs investigation

See Confidence Levels for the full posture rationale.


For the complete entry-by-entry log with full rationale and alternatives considered, see the canonical DECISIONS.md on GitHub.

On this page