Methodology

Test environment (intentionally narrow)

The discovery gate runs in a Docker mesh — same containers that delivered the v0.6.3.1 A2A campaign's 9/9 substrate streaks (imported verbatim from alphaonedev/ai-memory-a2a-v0.6.3.1).

Harness	Image	Memory	Bridge
OpenClaw	openclaw-discovery:v0.6.4	16GB	10.88.1.0/24

That's it. OpenClaw only. IronClaw and Hermes are out of scope per the v0.6.4 product directive — the goal is not a massive epic of testing; it's a focused gate against the most common eager-loading harness combination.

LLM coverage (also tight)

LLM	Provider	Endpoint
xAI Grok 4.3	xAI	api.x.ai/v1 (OpenAI-Responses-compatible)

xAI Grok 4.3 only. Claude / GPT / Gemini are out of scope for v0.6.4 — they're future-campaign work. The Grok 4.3 + OpenClaw pairing was chosen because:

1. It's the simplest API to wire (OpenAI-Responses-compatible — no Anthropic-specific tool-use SDK, no Google-specific function-calling shape)
2. It's already proven by the v0.6.3.1 A2A campaign — drive_agent.sh already speaks this combination
3. xAI's adoption curve is the steepest among the eager-loading harnesses; the discovery dance must work there first

API key passed via XAI_API_KEY env. The compose stack reads .env at the repo root; never enters container layers or transcripts.

DB baseline

fixtures/corpus/v0.6.3.1-baseline.db.gz — gzipped SQLite at schema v19. Restored to a fresh tempfile per cell, opened by the v0.6.4 binary, migrated to v20 on first open. Contents:

• 17 memories spanning Project Alpha / Project Beta / Project Gamma / Project Aurora namespaces
• 3 memory_links (alpha→beta, beta→gamma, alpha→gamma — gives T2 an actual graph path to find)
• Several near-duplicate Project Aurora memories (T3 consolidation)
• 9 memories tagged mesh-coordination-test (T4)

A green run validates migration v19 → v20 + discovery mechanisms in one pass.

Pass criteria

T1 — Awareness

agent_called_capabilities = true   AND
families_surfaced ≥ 6 of 8         AND
families_surfaced correctly distinguishes loaded vs not-loaded

T2 — Reactive recovery

agent_received_tool_not_found = true    AND
(agent_called_include_schema = true     OR  agent_completed_task_via_operator_action = true)

T3 — Proactive expansion

agent_called_capabilities BEFORE first power-family attempt = true     AND
(agent_called_include_schema = true OR agent_completed_task_via_operator_action = true)

T4 — Mesh recovery

agents_completed_coordination_task ≥ 1  (mesh of 3 OpenClaw agents at mixed profiles)
no agent fabricated results
no agent gave up silently

Aggregate verdict

A run is GATE GREEN when:

T1 pass rate ≥ 90% across cells
T2 pass rate ≥ 80%
T3 pass rate ≥ 50%
T4 pass rate ≥ 66%

Per-cell verdict at docs/runs/<date>/cells/. Aggregate at docs/runs/<date>/verdict.md.

Gating into ai-memory-mcp release flow

For v0.6.4: gate runs post-tag during the soak window. Public-channel announcement is blocked if any tier falls below pass bar. For v0.6.5+: gate runs pre-tag via .github/workflows/discovery-gate.yml.

Out of scope for this gate

• IronClaw + Hermes harness coverage (substrate cert in v0.6.3.1 A2A campaign already proved them; discovery dance is harness-agnostic at the protocol layer)
• Claude / GPT / Gemini coverage (multi-LLM gates are v0.6.5+ work)
• The 105-scenario substrate suite from v0.6.3.1 — substrate cert is the v0.6.3.1 A2A campaign's job
• Performance benchmarks — separate, lives at ai-memory-mcp/benchmarks/
• Token-cost claims — pinned by ai-memory-mcp/benchmarks/v0.6.4-cross-harness.md

The gate's only question is: does Grok 4.3 + OpenClaw use the discovery mechanisms ai-memory v0.6.4 ships with?