NeuralMind v3.10.0 — the v3.9.0 retrieval correction, and a public benchmark page

Release Date: 2026-09-10

Summary

Two threads: the faithfulness-gate regression that v3.9.0’s adversarial retrieval work introduced (found after release, root-caused, and fixed here — see the correction note in RELEASE_NOTES_v3.9.0.md), and a first-class public benchmark page on the marketing site plus SEO/indexability work.

Retrieval: v3.9.0’s adversarial-retrieval pull-in made opt-in

v3.9.0 shipped three retrieval “fixes” (synapse-seeded expansion, dependency-graph traversal, code-snippet extraction) that pulled extra candidate nodes into the L3 context slice. Measured after release: on the faithfulness fixture, appending those candidates moved the delta from +0.041 to -0.065 against a +0.000 gate floor — worse, not better. Making the pull-in budget-neutral (displacing a hit instead of appending) made it worse still (-0.107), because displacement evicts a real hit per candidate, and a candidate worse than what it replaces costs facts, not just tokens.

• Restored the faithfulness gate as blocking, and gated the v3.9.0 pull-in behind NEURALMIND_RETRIEVAL_EXPANSION (default off — see CLI Reference). Intent classification and the code-signal boost, the other two v3.9.0 changes, are unaffected and stay on by default — they re-rank existing hits rather than adding to them, and measured bit-for-bit neutral.
• Reset learned synapse state between faithfulness samples, and averaged the gate over 3 samples to absorb HNSW-index jitter — the earlier single-sample gate was measuring noise as often as signal.
• Ranked the merged L3 slice before flattening it, so an expansion candidate has to actually outrank what it would replace rather than being appended regardless of score.
• docs: corrected faithfulness figures that had been computed on a fixture that included its own answers — i.e., training on themselves.

Site

• Public benchmark page — the benchmark NeuralMind runs against itself is now a first-class page on the marketing site rather than only a results file in the repo.
• Fixed six pages that were canonicalizing to the homepage instead of themselves, and moved FAQ answers into the actual page HTML (they were previously JS-rendered only, invisible to a crawler that doesn’t execute JS).
• Added a 2026 SEO plan with a live indexability audit, and a comparison page against Graft and codebase-memory-mcp.
• Synced the Agent Zero / a0-plugins ecosystem listing status now that the a0-plugins registry submission (staged in v3.6.0) has actually merged, and documented both live registries.

Release tooling

• Repaired a release-please baseline drift and added a guard against orphaned release tags / manifest drift, after a version-manifest mismatch surfaced the gap. Added a ClawHub publish script.

What the agent sees post-install

NEURALMIND_RETRIEVAL_EXPANSION=1 opts back into the v3.9.0 pull-in behavior if you want it — it is a research setting, not a recommended default; see the correction note in RELEASE_NOTES_v3.9.0.md for the measured tradeoff. No other agent-visible behavior changed; the site/SEO work has no effect on the installed package.

Upgrading

pip install --upgrade neuralmind

If you had set NEURALMIND_RETRIEVAL_EXPANSION or relied on the v3.9.0 pull-in behavior being on by default, it is now off by default — set NEURALMIND_RETRIEVAL_EXPANSION=1 explicitly to keep it.