Track F — Heterogeneous AI-NHI Assessment v3 Pointer (2026-05-28)

Track F is the parallel-dispatched heterogeneous AI-NHI assessment v3, dispatched to a fresh Opus 4.7 (1M context) agent against the post-drift-fix release/v0.7.0 HEAD be3347d70 (with PR #1379 docs+Pages remediation, PR #1380 CoALA citation, and PR #1382 PostgresTestEnv fix pending operator merge).

Status

IN-FLIGHT at dossier write time. The agent was dispatched in parallel with this dossier-build task; the report lands when the agent returns.

Expected report path

/Users/fate/v07/v07-f5/docs/v0.7.0/heterogeneous-ai-nhi-assessment/report-claude-opus-4-7-v3.md

If the report is not yet present at the time you are reading this dossier, expect it to land at the path above. The directory docs/v0.7.0/heterogeneous-ai-nhi-assessment/ already exists and hosts the prior-version reports + the assessment-framework documentation.

Assessment scope

The v3 assessment dispatches a fresh Opus 4.7 agent to evaluate ai-memory v0.7.0 as a substrate for heterogeneous AI-NHI (Non-Human Identity) operations. The dispatch frame includes:

1. The post-drift-fix HEAD. release/v0.7.0 at be3347d70 with PR #1379 + PR #1380 + PR #1382 effects assumed merged.
2. The 73 MCP tools at --profile full surface. Including the v0.7.0-net-new tools (memory_reflect, memory_atomise, memory_skill_*, memory_persona_*, memory_check_agent_action, memory_rule_list, etc.).
3. The 26-field Memory shape + 6 MemoryLink variants for full provenance + confidence-calibration + recursive-learning coverage.
4. The lan-parity dual-agent topology as the realistic heterogeneous-NHI substrate (alice + bob at separate HTTP ports against a shared pg-age postgres).
5. The CoALA prior-art citation context (Track E / PR #1380) so the agent can position the substrate against the published cognitive-architecture literature without conflating the citation with a design constraint.

Scope discipline

Per pm-v3.3, the v3 agent operates under:

• Recompile-retest discipline (C5 step 7). Any load-bearing behavioral finding about a running MCP/HTTP/CLI daemon must reproduce against a freshly-spawned subprocess against the rebuilt binary, NOT the long-running PID 10338 (which may carry a binary snapshot that pre-dates code changes on disk).
• No banned phrases. “Non-blocking”, “DEFER-TO-V080”, “operator should…”, etc. are HARD-BLOCKed by the orchestrator safeguards.
• Issue-filed-at-discovery contract. Any defect surfaced is filed as a GH issue at the moment of discovery, not bundled at the end of the assessment.

Lineage

The v3 assessment is the third in a series:

The v1 + v2 reports live under docs/v0.7.0/heterogeneous-ai-nhi-assessment/. The v3 report (this campaign’s instance) will land at the path noted above.

The lineage matters because the #1315 stale-binary lesson (from v1 Phase-1) is the canonical example of pm-v3.3 step-7’s necessity: the v1 agent filed #1315 as a wire-layer regression that the QC subagent’s fresh-subprocess re-probe later proved was a stale-binary diagnosis (not a substrate defect). The v3 dispatch frame explicitly cites this lesson so the v3 agent applies step 7 mechanically from the start.

Cross-track relationship

• Track A’s 1hr dogfood + extended uptime evidence (16h+ on PID 10338) gives the v3 agent the load-bearing daemon-stability data point.
• Track B’s lan-parity A2A evidence gives the v3 agent the heterogeneous-NHI federation data point (alice + bob across shared pg-age substrate).
• Track C’s PR #1382 fix (test-isolation discipline for the lan-parity shared-container path) is one of the substrates the v3 agent will probe — the agent should NOT re-discover #1381 as a new defect.
• Track D’s docs+Pages remediation gives the v3 agent the canonical-count surface to reference (so the agent does not mis-cite v50 / 57 / 73-routes-instead-of-87, etc.).
• Track E’s CoALA citation gives the v3 agent the prior-art citation discipline to mirror in its own positioning sections.

Expected report shape

When the v3 report lands at the expected path, expect it to follow the v1 + v2 conventions:

1. Verdict at a glance. SHIP-recommended / SHIP-with-caveats / NEEDS-REWORK.
2. Phase-by-phase methodology. Bootstrap, capability discovery, write+recall roundtrip, federation roundtrip, governance refusal, KG traversal, signed-events chain, etc.
3. Per-phase findings. Each phase ends with PASS / FAIL / PARTIAL plus the per-finding GH issue numbers (filed at discovery per pm-v3.3).
4. Cross-cutting observations. Wire-shape consistency, error-envelope discipline, capability-overlay correctness, heterogeneous-NHI identity-attribution behavior.
5. Recommendations. v0.7.x follow-up items, v0.8 design considerations, NO banned phrases.

Audit trail (placeholder; will be populated when the report lands)

• Report path: docs/v0.7.0/heterogeneous-ai-nhi-assessment/report-claude-opus-4-7-v3.md (expected; may be IN-FLIGHT at dossier read time)
• Dispatch evidence: .local-runs/2026-05-28-ship-campaign/nhi-assessment-opus-4-7-v3/ (dispatch artifacts)
• Prior reports: docs/v0.7.0/heterogeneous-ai-nhi-assessment/ (v1, v2)
• Dispatch HEAD: be3347d70
• Dispatch frame: post-drift-fix (#1379 + #1380 + #1382 applied)

Cross-reference

For the full assessment-framework documentation, see docs/v0.7.0/heterogeneous-ai-nhi-assessment/ index. For the methodology that pm-v3.3 step 7 codified, see CLAUDE.md “Verify-before-claiming + no-operator-handoffs” §”Lineage of this step” (which cites #1315 as the canonical stale-binary example).

Verdict: POINTER (IN-FLIGHT)

This track-F document is a pointer to the parallel-dispatched v3 report. The final ship-campaign verdict (in README.md) does NOT depend on the v3 report’s verdict for SHIP-CLEARED on the 5 in-host tracks (A–E); the v3 report is an additional independent assessment that complements the in-host evidence with a fresh-eyes structural review.

Drafted by Claude (Opus 4.7, 1M context).