RuvNet Brain · Source-grounded intelligence

rUv ships 8–9 months ahead.
Claude was trained on the old world.

rUv is well ahead of the curve. But AI is trained looking back — so Claude drifts, doubts, second-guesses. The Brain grounds every answer in rUv's real source.

v0.5.0-dev 21 building blocks 90,842 grounded chunks every claim re-checkable

the story →

The problem · why it's hard if you're not rUv

Give Claude the best AI stack in the world.
Watch it not use it.

rUv's ecosystem: ~248 repos, Rust-first, barely in any training set. So Claude does what every model does with the unfamiliar:

Drift

Reaches for pgvector, Pinecone, Postgres — instead of RVF.

Hallucinate

Invents APIs that were never in the repo.

Doubt

Insists a tool can't do what it actually does.

Skim

Pattern-matches the README, not the real code.

Even with Ruflo + RuVector + Fable 5 wired up, it plateaus. The tools were there. The knowledge wasn't.

The insight

The bottleneck was never the model.
It was knowledge — and nerve.

Installing rUv's tools isn't enough — Claude can't wield them like rUv without the knowledge. So we built a brain: every line of his real source, vector-indexed. Before Claude answers, it grounds and cites the file.

BEFORE"RVF probably can't handle concurrent writes…" — so it reaches for Postgres.

AFTERCites ruvector/…/09-concurrency-versioning, uses the durable-writer crate, stays in-stack.

Capability-confidence gate

84%→100%

It never again wrongly doubts what a repo can do — proven on a held-out capability set, with controls proving it won't invent capabilities either.

How it works

Ground before you assert.
Enforced, not suggested.

You ask

"Build X on the RuvNet stack"

→

Brain grounds

search_ruvnet · RVF/HNSW over 90,842 chunks

→

Cites

rUv's real file + path

→

Builds

with the full stack, confidently

Source-grounded retrieval

An RVF single-file vector store + cross-encoder rerank over rUv's actual code, ADRs and specs — not a summary.

Behavioral hooks

ground-before-assert · anti-drift substitution · take-the-wheel. Structural, so Claude can't quietly skip the grounding step.

Drop-in

npx github:stuinfla/ruvnet-brain. Runs right beside Ruflo & RuVector — it never fights rUv's tools, it grounds them.

How it's built · grounded, then verified

It doesn't skim rUv's repos.
It reads every layer — then re-checks.

Reaches into 20+ core repos → pulls the real source — ADRs, manifests, Rust bodies — into one RuVector .rvf knowledge base. No server, no API.

Verify every layerADR ↓ Rust

✓ADR — design intent · labeled shipped vs proposed

✓DDD / domain model · bounded contexts

✓Crate manifest · Cargo, features

✓Source bodies · doc-comments

✓Rust functions · the real implementation

↻

grade → diagnose → fix → re-grade — looped until retrieval returns the right layer. Capability confidence climbed 84 → 100.

It even labels each ADR shipped vs. proposed — so design intent is never mistaken for what actually ships.

Always current · rebuilds itself nightly

rUv is prolific.
So the Brain never goes stale.

rUv is well ahead of the curve, committing across ~248 repos faster than any cheat-sheet could track. A static one would rot in weeks.

✓

Bundle status

CURRENT

last rebuild · last night · 3:15 AM

+12

+7

+21

today

Every night a LaunchAgent checks rUv's repos, rebuilds only what changed, re-verifies, and re-stamps — automatically.

Honest: nightly, changed-repos-only (safe scope); brand-new repos are added under supervision. Your instruction manual updates while you sleep.

Proof #1 · a live business

One live salon — made genuinely smarter. 61 → 93, live-verified.

61~~/100~~

Before · Jun 29

→

93~~/100~~

Live · Jul 2

+32 points → 96–97 target

parallel-swarm sprint · 61 → 82 → 88 → 93 live · verified against live production · chrisdavidsalon.com

Intelligence

70 89 +19

Test coverage

62 83 +21

Architecture

58 80 +22

Security

85 90 +5

Composite

61 93 +32

Three things a full Ruflo session had missed — the Brain caught them by grounding in rUv’s source

01

The “AI” was faking it

Left in place: the recommender was 7 hardcoded platitudes from a memory that reset every request. The Brain caught it wasn’t intelligent at all — rewired to real measured outcomes.Intelligence 70 → 89

02

It couldn’t tell cause from correlation

Left in place: the learning loop just guessed. The Brain grounded it in causal do-calculus — models uplift with confidence intervals, and 2 adversarial skeptics now veto confidently-wrong picks.reasons, then explains why

03

A 2,326-line monolith

Left in place: one giant file, thin tests. The Brain split it into a 515-line router + 8 modules and drove tests 90 → 245, 0 fail.Architecture +22 · Test coverage +21

Honest: 93 live-verified, not 96–97 — nothing hidden: the 2 biggest revenue engines are shipped but armed-off (credit lags until they’re switched on), and one leaked key is owner-gated. full evidence ledger →

chrisdavidsalon.com

Chris David Salon — the live salon website

Proof #2 · a SaaS re-architecture

A SaaS app, re-architected in a day.
54 → 95, live in prod.

Composite

54→95

+41 points live in prod · v6.11.2

the climb, each step proven live: 54 → 70 → 80 → 95
the last gate is closed — real PMID/LCD citations now surface in prod; letters cite real ERISA/CFR/USC with 0× fabrication.

Three things a full Ruflo session had missed — the Brain caught them by grounding in rUv’s source

01

A live medical fabrication

Shipped past Ruflo: the product cited a wrong CMS code (L35056) from a corrupted KB. The Brain’s adversarial vet caught it before it reached a user.0× fabrication · real ERISA/CFR/USC

02

The root cause was mis-framed

Would’ve been a prompt tweak. Grounding in the RVF single-writer spec + ADR-029 (data returned verbatim, no correctness layer), the Brain proved it was a DATA problem — fixed at the source (native-HNSW rebuild, 3,564 vectors).the fix that actually holds

03

A whole denial type was dead

Silently failing: the exact denial that hard-failed 3× now delivers a real letter + a 19-file packet; mobile hero un-clipped; CI green e2e.156 jest green · 8/8 live

Honest: 95, live in prod (v6.11.2) — the whole climb is grounded in the RVF single-writer spec + ADR-029, every number tied to a live-prod test, nothing inflated. full evidence ledger →

appealarmor.com

The unlock

Same stack.
Suddenly used to its full depth.

Both apps already had Ruflo and RuVector — one even had Fable 5. Both had plateaued. The unlock wasn't a bigger model or more tools. It was Opus 4.8 + the RuvNet Brain: the map that let it —

Audit→ Find the AI slop→ Re-architect→ Secure→ Prove it

— grounded in rUv's real source at every step, with every number tied to a re-run test.

"It builds with the RuvNet stack as fluently and confidently as rUv would."

— the bar we set for the project