Skip to content

Scope — v0.6.3.1 First-Principles campaign

What this campaign tests, what it deliberately does not test, and what each verdict in releases/v0.6.3.1/summary.json means.

This page is the docsite-rendered companion to repo-root SCOPE.md. The authoritative specification of how the campaign behaves is docs/governance.md — if anything below conflicts with governance.md, governance wins.

Subject under test

Field Value
Tag v0.6.3.1 (cut 2026-04-30)
Schema v19 (migration ladder v15 → v17 → v18 → v19)
Repo alphaonedev/ai-memory-mcp
CLI surfaces added since v0.6.3 ai-memory boot, install <agent>, wrap <agent>, logs, audit verify, doctor
Config surfaces added [boot], [logging], [audit], [audit.compliance.{soc2,hipaa,gdpr,fedramp}]
Tooling claim being verified 1,886 lib tests, 49+ integration tests, 93.84% line coverage, 17 documented integrations × 10 platforms, 7-section doctor health dashboard
Campaign repo alphaonedev/ai-memory-a2a-v0.6.3.1
Verdict surface releases/v0.6.3.1/summary.json (schema v2)
Umbrella issue ai-memory-mcp#511

The doctor surface is the v0.6.3.1 promotion of R7 from ROADMAP2 §7.2. The budget_tokens recall is the recovered commitment R1.

Agent scope

In scope (Principle 6 — scope discipline; see governance §2.6):

  • IronClaw — Rust framework, primary cert agent.
  • Hermes — counterpart agent for cross-framework substrate cells and Phase 3 NHI playbook.
  • OpenClaw — first-class third agent framework, in scope for v0.6.3.1 but dispatched separately from a higher-resource workstation via the local Docker mesh (16 GB+ per container — Basic-tier DO droplets used for IronClaw/Hermes don't have enough memory). Per Principle 6, OpenClaw runs produce their own scope-tagged artifacts; the Orchestrator joins all three frameworks via release=v0.6.3.1 linkage but never collapses them across framework boundaries (cross-scope contamination invalidates the artifact).

What carries forward from the v0.6.3 cert

Eight base scenarios S1 – S8 from the umbrella testbook run unchanged on the v0.6.3.1 binary:

Scenario What it asserts
S1 Per-agent write + read
S2 Shared-context handoff
S3 Targeted share
S4 Federation-aware agents
S5 Consolidation + curation
S6 Contradiction detection
S7 Scoping visibility
S8 Auto-tagging

These must remain green. Any regression on S1 – S8 is a release-blocker for the campaign verdict — they are not Patch 2 candidates.

The certification cell is ironclaw / mTLS at the carry-forward sweep. Target: 48 / 48 on that cell.

What is new for v0.6.3.1

Sixteen scenarios S9 – S24 exercise the surfaces added in this release:

ID Surface Assertion
S9 ai-memory boot multi-node Boot manifest agreement on version, schema_version, tier across 4 nodes.
S10 ai-memory doctor cross-node Storage / Index / Recall / Sync sections agree (in-flight delta tolerated).
S11 ai-memory install <agent> recipe handoff Recipes installed at different nodes share the same store cleanly.
S12 ai-memory wrap <agent> cross-vendor Boot context delivered for codex / gemini / aider / ollama.
S13 ai-memory audit verify tamper-evident Hash chain verifies independently per node; corruption detected.
S14 ai-memory logs operator surface Filters work uniformly; privacy default OFF holds.
S15 R1 budget_tokens recall Same query + budget → same ranked head on federated peers.
S16 Capabilities v2 honesty cross-mesh Asymmetric capabilities surface, never absorb.
S17 G9 webhook fanout link / promote / delete / consolidate fire from any originating node.
S18 G4 embedding-dim integrity Mixed-dim writes refused at boundary.
S19 G5 archive/restore preserves embeddings Archive at A → restore at B keeps embedding + tier + expiry.
S20 G6 on_conflict policy Concurrent same-key writes deterministic.
S21 G13 endianness magic byte x86_64 + arm64 nodes exchange f32 BLOBs without silent corruption.
S22 Schema v19 migration Heterogeneous mesh; warn manifest variant fires on mismatch.
S23 Issue #507 ~ expansion Tilde in db field expanded before SQLite open. Expected RED on v0.6.3.1.
S24 Issue #318 MCP stdio fanout MCP stdio writes trigger federation fanout. Expected RED on v0.6.3.1.

S23 and S24 are intentionally expected to fail. They correspond to known-open OSS bugs that will close in Patch 2 (v0.6.3.2). Their inclusion is an integrity check on the harness itself: if either ever turned green on v0.6.3.1, the harness would be lying. Both flip to expected-green in the successor ai-memory-a2a-v0.6.3.2 repo.

What is deferred (out of scope)

Per governance Appendix B, this campaign deliberately does not cover:

  • Infrastructure provisioning, mTLS cert management, Terraform. Confirmed-good before Phase 0 by a separate process.
  • Auto-tagging via Ollama. Opt-in feature requiring s-4vcpu-16gb droplet; deferred per the ai2ai-gate README.
  • memory_share / #311 targeted-share scenario. Depends on a v0.6.0.1 capability upstream of v0.6.3.1; revisit for v0.6.4.
  • Patch 2 (v0.6.3.2) regression run. Separate campaign in ai-memory-a2a-v0.6.3.2, using this repo as a template.
  • v0.7 attestation (Ed25519 signing of memory_links). Covered by the v0.7 cert.
  • v0.8 typed cognition / CRDTs / curator daemon. Covered by the v0.8 cert.
  • Performance / scale benchmarking. Owned by the ship-gate, not this correctness campaign.

Deferred work is not a gap in the v0.6.3.1 cert — it is owned by a different per-release cert.

Verdict criteria — two truth-claims, two evidence streams

Per Principle 1, the campaign emits two independent verdicts in releases/v0.6.3.1/summary.json (schema v2). They are stored in separate top-level keys (substrate_verdict, nhi_verdict) and never collapsed.

Substrate verdict (Phase 1 — binary, reproducible)

Read this for ship / no-ship gating. Computed from S1 – S24 outcomes on the ironclaw / mTLS cert cell.

Value Meaning
CERT All S1 – S22 GREEN; S23 and S24 RED as expected; cert cell is 48 / 48. The release passes A2A certification on this tag.
PARTIAL — pending Patch 2 Same as CERT except S23 and / or S24 are still RED. This is the expected v0.6.3.1 outcome — see governance §2.2. The release is conditionally certified pending Patch 2.
FAIL Any other red — including any regression on S1 – S8 or any unexpected failure on S9 – S22. Each red opens or updates a bug + v0.6.3.2-candidate issue parent-linked to #511.

If S23 or S24 flips GREEN, the harness halts the campaign and files a harness-integrity issue. Substrate is not allowed to lie.

NHI verdict (Phase 3 — behavioral, statistical, n=3)

Read this to assess whether ai-memory measurably changes NHI behavior under realistic agent workloads. Computed from the four-arm × four-scenario × n=3 = 48-run Phase 3 playbook.

The NHI verdict reports per-scenario treatment effects (Arm-T vs each control arm) and a cross-layer consistency assessment, never collapsed into a single ship/no-ship bit. See governance §6 for the full scenario design and matrix for the cell-by-cell view.

A campaign that ships a green substrate badge but produces no Phase 3 behavioral evidence does not satisfy the v0.6.3.1 cert.