chore(main): release 4.1.0

[github-actions]

🤖 I have created a release beep boop

4.1.0 (2026-09-18)

Features

• memory: decision memory layer with commit-level invalidation (220a6fa)
• memory: eval harness + MCP tools for decision memory (e14ddfa)
• memory: tests/memory/ suite + register memory MCP tools (3664062)

Bug Fixes

• add ensure_test_chapters() to all tests using CHAPTERS_DIR (bbd2451)
• CI failures — black formatting + remove hardcoded paths (1198ce0)
• dedicated prose fixture for pipeline integration tests (b8e3737)
• memory: update() referenced nonexistent files attr — silent no-op (066a48b)
• parity gate faithfulness floor 0.0 -> -0.10 (6ab2082)
• remove erroneous .neuralmind.yaml from code fixture (2a31413)
• resolve remaining CI lint failures and Windows checkout error (edab170)
• reword absolute claim on measure-your-own page (f6f3d71)
• site: bump deps to clear Dependabot alerts + add cognition icon (4881eab)
• test_pipeline_integration — portable paths, prose mode, self-contained chapters (7e4bec8)
• update parity gate unit test for new -0.10 floor (a1174f3)
• widen self-benchmark faithfulness floor from +0.000 to -0.100 (23a5122)

---

This PR was generated with Release Please. See documentation.

[github-actions]

Backend parity gate — graphify vs built-in tree-sitter

✅ PASS — the built-in backend must stay within tolerance of graphify on the reference fixture.

Metric	graphify	built-in
code nodes	65	79
mean reduction	4.94×	6.30×
faithfulness delta	-0.061	-0.054
fact recall	0.527	0.451
grounding	0.889	0.843

Gate checks

• ✅ reduction within tolerance of graphify — built-in 6.30× ≥ 3.71× (graphify 4.94× − 25%)
• ✅ reduction ≥ absolute floor — built-in 6.30× ≥ floor 4.00×
• ✅ faithfulness delta within tolerance of graphify — built-in -0.054 ≥ -0.161 (graphify -0.061 − 0.10)
• ✅ faithfulness delta ≥ absolute floor — built-in -0.054 ≥ floor -0.100
• ✅ fact recall within tolerance of graphify — built-in 0.451 ≥ 0.427 (graphify 0.527 − 0.10)

Tolerances: reduction within 25% (floor 4.0×), faithfulness within 0.10 (floor -0.10). Override via NEURALMIND_PARITY_* env vars.

Automated by evals/parity/run.py — reproduce locally with python -m evals.parity.run.

Multi-language structural parity

Language	graphify symbols	built-in covers	dangling
typescript	54	54 (100%)	0
go	45	45 (100%)	0
rust	49	49 (100%)	0
java	52	52 (100%)	0
c	47	47 (100%)	0
cpp	51	51 (100%)	0
csharp	52	52 (100%)	0
ruby	46	46 (100%)	0
php	54	54 (100%)	0

• ✅ typescript: symbol coverage ≥ floor — 54/54 graphify symbols (100%) ≥ 90%
• ✅ typescript: no dangling edges — 0 dangling edge(s)
• ✅ go: symbol coverage ≥ floor — 45/45 graphify symbols (100%) ≥ 90%
• ✅ go: no dangling edges — 0 dangling edge(s)
• ✅ rust: symbol coverage ≥ floor — 49/49 graphify symbols (100%) ≥ 90%
• ✅ rust: no dangling edges — 0 dangling edge(s)
• ✅ java: symbol coverage ≥ floor — 52/52 graphify symbols (100%) ≥ 90%
• ✅ java: no dangling edges — 0 dangling edge(s)
• ✅ c: symbol coverage ≥ floor — 47/47 graphify symbols (100%) ≥ 90%
• ✅ c: no dangling edges — 0 dangling edge(s)
• ✅ cpp: symbol coverage ≥ floor — 51/51 graphify symbols (100%) ≥ 90%
• ✅ cpp: no dangling edges — 0 dangling edge(s)
• ✅ csharp: symbol coverage ≥ floor — 52/52 graphify symbols (100%) ≥ 90%
• ✅ csharp: no dangling edges — 0 dangling edge(s)
• ✅ ruby: symbol coverage ≥ floor — 46/46 graphify symbols (100%) ≥ 90%
• ✅ ruby: no dangling edges — 0 dangling edge(s)
• ✅ php: symbol coverage ≥ floor — 54/54 graphify symbols (100%) ≥ 90%
• ✅ php: no dangling edges — 0 dangling edge(s)

Coverage floor: 90% of graphify's per-language symbols (no gold-fact set exists for TS/Go, so parity is structural).

Optional SCIP precision pass

• ✅ precision: SCIP corrects the heuristic call edge — run() → A.handle under SCIP (heuristic wrongly linked B.handle)
• ✅ precision: strict no-op when disabled — graph unchanged when NEURALMIND_PRECISION is unset

Off by default (NEURALMIND_PRECISION); proven on tests/fixtures/scip_precision to replace a heuristic call edge with the compiler-accurate one a SCIP index resolves.

[github-actions]

NeuralMind self-benchmark

Status: PASS — floor 4×, measured 5.1×.

Phase 1 — Reduction on committed fixture

• Average reduction: 5.1×
• Top-k retrieval hit rate: 66.7%
• Naive baseline: 89,984 tokens (all fixture files concatenated)
• NeuralMind total: 17,812 tokens across 19 queries
• Estimated monthly savings @ 100 queries/day on Claude 3.5 Sonnet: ~$34.19

#	Query	Shape	Naive	NeuralMind	Ratio	Hit
1	auth-flow	cross-file	4,736	974	4.9×	33.3%
2	api-endpoints	focused	4,736	978	4.8×	100.0%
3	billing-flow	cross-file	4,736	1,019	4.6×	33.3%
4	user-storage	cross-file	4,736	914	5.2×	50.0%
5	jwt-verify	focused	4,736	821	5.8×	100.0%
6	stripe-webhook	focused	4,736	949	5.0×	100.0%
7	create-user	cross-file	4,736	880	5.4×	50.0%
8	refund	focused	4,736	1,028	4.6×	0.0%
9	db-choice	identity	4,736	1,012	4.7×	100.0%
10	invoice-send	cross-file	4,736	947	5.0×	50.0%
11	debug-login-silent	focused	4,736	852	5.6×	100.0%
12	debug-webhook-reject	focused	4,736	916	5.2×	100.0%
13	debug-invoice-missing	cross-file	4,736	870	5.4×	50.0%
14	debug-token-expire	focused	4,736	1,038	4.6×	100.0%
15	refactor-add-user-field	cross-file	4,736	919	5.2×	50.0%
16	refactor-swap-database	focused	4,736	928	5.1×	100.0%
17	refactor-add-mfa	cross-file	4,736	915	5.2×	50.0%
18	next-after-jwt-change	cross-file	4,736	833	5.7×	50.0%
19	next-after-charge-change	cross-file	4,736	1,019	4.6×	50.0%

Phase 2 — Synapse recall A/B (same warm graph, recall off vs on)

• Synapse edges after seeding co-editing sessions: 2954
• Top-k hit rate: 71.9% off → 80.7% on (Δ +8.8 points, mean of 3 runs)
• Per-run deltas: +8.8, +8.8, +8.8 points — published because this metric moves between runs, and a mean that hides its own spread is how a directional claim gets decided by luck.
• Reduction ratio: 5.2× off → 5.2× on (Δ +0.03× — budget-neutral by design)

The Hebbian synapse layer is now the single learning measurement (the old
learned_patterns reranker was removed). The hit-rate delta shows associative recall
surfacing co-edited modules a purely textual search ranks lower; the near-zero reduction
delta confirms it does so without spending extra tokens (recalled nodes displace the
weakest hits, not add to them).

Assumptions

• Baseline: every .py file in tests/fixtures/sample_project/ concatenated.
• Tokenizer: tiktoken GPT-4o encoding (per-model breakdown in multi_model.json if generated).
• Pricing: Claude 3.5 Sonnet input @ $3.0/MTok.
• Regression floor: 4× — well below NeuralMind's typical 12–50× on real repos.

Per-model token reduction

Model	Tokenizer	Naive	NeuralMind	Ratio	Source
GPT-4o / GPT-4o-mini	tiktoken o200k_base	4,739	927	5.1×	measured
GPT-4 / GPT-3.5-turbo	tiktoken cl100k_base	4,710	918	5.1×	measured
Claude 3.5 Sonnet	estimated: GPT-4o × 1.08 — install anthropic for an exact count	5,118	1,001	5.1×	estimated
Llama 3 (70B)	estimated: GPT-4o × 1.22 — Llama tokenizer requires model weights; estimate based on published vocab ratios	5,781	1,130	5.1×	estimated

Rows marked measured use the provider's real tokenizer. Rows marked
estimated apply a published vocab-size correction to the GPT-4o count —
honest approximations, not hardcoded claims.

NeuralMind retrieval-quality eval

Suite	Queries	MRR	Answerability	Recall@5	Precision@5	Gate
c	10	0.650	90%	0.900	0.205	PASS
cpp	10	0.725	100%	1.000	0.293	PASS
csharp	5	1.000	100%	1.000	0.463	PASS
go	19	0.939	100%	0.956	0.404	PASS
java	19	0.886	100%	0.904	0.350	PASS
php	4	0.875	100%	1.000	0.479	PASS
python	19	0.974	100%	0.912	0.413	PASS
ruby	4	0.875	100%	1.000	0.467	PASS
rust	19	0.974	100%	0.956	0.361	PASS
typescript	19	0.895	100%	0.912	0.372	PASS

Query category coverage (PRD 2):

• c: architecture: 10
• cpp: architecture: 10
• csharp: architecture: 5
• go: architecture: 10, bug-localization: 4, refactor: 3, next-edit: 2
• java: architecture: 10, bug-localization: 4, refactor: 3, next-edit: 2
• php: architecture: 4
• python: architecture: 10, bug-localization: 4, refactor: 3, next-edit: 2
• ruby: architecture: 4
• rust: architecture: 10, bug-localization: 4, refactor: 3, next-edit: 2
• typescript: architecture: 10, bug-localization: 4, refactor: 3, next-edit: 2

go vs baseline:

• mrr: 0.939 (▼ -0.011)
• answerability: 1.000 (= +0.000)
• recall@1: 0.614 (▼ -0.003)
• recall@3: 0.886 (▲ +0.053)
• recall@5: 0.956 (▲ +0.123)

python vs baseline:

• mrr: 0.974 (▲ +0.074)
• answerability: 1.000 (= +0.000)
• recall@1: 0.667 (▲ +0.150)
• recall@3: 0.860 (▲ +0.026)
• recall@5: 0.912 (▲ +0.079)

typescript vs baseline:

• mrr: 0.895 (▼ -0.005)
• answerability: 1.000 (= +0.000)
• recall@1: 0.570 (▼ -0.013)
• recall@3: 0.860 (▲ +0.060)
• recall@5: 0.912 (▲ +0.112)

Overall: PASS

---

Automated by .github/workflows/ci-benchmark.yml — regenerate locally with python -m tests.benchmark.run and neuralmind benchmark --quality.

[github-actions]

NeuralMind CI Compliance Check

Framework: all
Changed files: 7
Files with annotations: 0

Summary: No compliance annotations affected by this diff.

[github-actions]

🤖 Created releases:

• v4.1.0

🌻