Skip to content

Release 0.7.0

Choose a tag to compare

View all tags
@cdcavell cdcavell released this 22 Aug 10:51
· 128 commits to main since this release
v0.7.0
5142f30
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Version 0.7.0 expands the repository's AI Integration curriculum, improves problem-oriented navigation, consolidates repeated long-form architecture material, and strengthens maintenance controls around the customized DocFX site template.

This release completes the current Milestone 9 — Expanded AI Integration scope while preserving the repository's central architectural boundary:

The model may propose. The host retains execution authority.

Expanded AI Integration

Governed Multi-Tool Workflows and Recovery Boundaries

New advanced material extends the single-tool governance model into multi-step AI-assisted workflows.

Topics include:

  • Whole-plan validation versus per-step authorization.
  • Step-scoped policy evaluation and execution authority.
  • Policy and resource drift between steps.
  • Partial success and partial failure.
  • Idempotency and replay considerations.
  • Compensation and cancellation.
  • Bounded replanning and recovery.
  • Human escalation.
  • Prevention of privilege accumulation across workflow steps.

A key invariant remains:

Step N is allowed
      ≠
Authority for Step N+1

AI Proposal Rejection, Uncertainty, and Recovery Patterns

A new focused architecture explanation addresses what happens when an AI-generated proposal cannot safely proceed.

The material distinguishes:

Invalid model output
        ≠
Valid proposal rejected by policy

and:

Low confidence
        ≠
No authority

It covers:

  • Parse and schema failures.
  • Unknown tools and invalid arguments.
  • Host-authoritative context conflicts.
  • Low-confidence and unavailable signals.
  • Policy denial and acknowledgment outcomes.
  • Infrastructure failures.
  • Stable rejection reason codes.
  • Safe model-visible feedback.
  • Retry budgets and loop detection.
  • Replanning, escalation, and terminal states.

The guidance explicitly rejects "retry until something passes" as a governance strategy.

Agent Memory and Governance Boundaries

This release adds a dedicated treatment of persistent and reusable AI memory.

The central lesson is:

Memory may inform a future proposal, but remembered information does not become authority merely because the system retained it.

The material covers:

  • Session, workflow, and persistent memory.
  • User-, host-, tool-, model-, and externally derived memory.
  • Provenance and source identity.
  • Freshness, expiration, retention, and deletion.
  • User, tenant, workflow, and agent scope.
  • Cross-agent memory sharing.
  • Memory write and read policy.
  • Host validation before persistence.
  • Bounded retrieval.
  • Sensitive-data minimization.
  • Memory poisoning and persistent prompt injection.
  • Stale and conflicting remembered facts.
  • Audit evidence for consequential memory use.

Important distinctions include:

Remembered information
        ≠
Authoritative current fact

Prior decision
        ≠
Current decision

Prior approval
        ≠
Standing permission

Memory
        ≠
Capability

Memory
        ≠
Audit record

Memory
        ≠
Credential

The article also keeps stateless and session-only designs explicitly valid where persistent memory is unnecessary.

Milestone 9 Complete

With multi-tool workflows, bounded proposal recovery, uncertainty handling, and agent-memory governance now covered, the current Expanded AI Integration milestone is complete.

The established AI Integration path now includes:

  • Typed AI-proposed intent and schema-validation boundaries.
  • Host-authoritative context reconstruction.
  • Deterministic and probabilistic policy inputs.
  • Governed multi-tool workflows.
  • Proposal rejection and bounded recovery.
  • Agent memory and governance boundaries.
  • Experimental multi-agent execution boundaries.
  • Scoped host-owned execution.

Future work can now focus more heavily on executable companions, labs, threat-model exercises, architectural comparisons, and refinement of the established curriculum rather than filling foundational AI Integration gaps.

Problem-Oriented Learning Paths

A new Find Your Path page provides an alternative to reading the repository sequentially.

Readers can now begin from common architecture problems such as:

  • Seeing the core decision-before-execution boundary quickly.
  • Determining whether ASP.NET Core authorization is already sufficient.
  • Governing a consequential administrative operation.
  • Governing AI-proposed tool execution.
  • Reasoning about trust boundaries and operational security.
  • Preserving architectural decisions with ADRs.

Each route points to existing canonical tutorials, samples, labs, and simpler alternatives rather than creating a duplicate curriculum.

Editorial Consolidation

Selected Governance and Security material received a focused editorial pass to reduce repeated explanations while preserving architectural depth.

The consolidation improves information density around:

  • Human-review boundaries.
  • Escalation concepts.
  • Risk-based governance.
  • Deterministic and probabilistic policy inputs.
  • Secret handling and secure logging.
  • Threat-modeling comparisons.

Canonical cross-links are preferred where a concept already has a dedicated treatment.

DocFX Template Maintenance

The documentation platform now has explicit protection against drift between the pinned DocFX version and the repository's customized modern _master.tmpl.

The release adds:

  • Machine-readable DocFX template baseline metadata.
  • A template-baseline validation tool.
  • CI and publishing validation for the baseline.
  • A documented DocFX upgrade and synchronization procedure.
  • Guidance to reconsider whether the full template override remains necessary during future DocFX upgrades.

This turns a previously documented maintenance obligation into a visible, reviewable validation step without introducing remote-template downloads into CI.

Documentation and Navigation

Navigation and cross-linking have been updated throughout the AI Integration and foundational material so the new articles are discoverable from their related architectural boundaries.

The repository continues to favor:

Depth before breadth.

New material is intended to strengthen established learning paths rather than create disconnected documentation.

Release Metadata

The release updates:

  • CITATION.cff
  • .zenodo.json

for version 0.7.0.

Scope

ASI Backbone Learning remains an educational architecture resource.

It provides tutorials, labs, samples, architectural comparisons, and working-reference links. It does not represent:

  • A compliance certification.
  • A security guarantee.
  • A legal or regulatory standard.
  • An AI model.
  • An autonomous-agent platform.
  • An AGI or ASI implementation.

Its purpose remains:

Teach architectural reasoning through patterns, examples, executable boundaries, tradeoffs, and working references.

Assets 2
Loading