admin/ruflo

0 Fork 0

mirror of https://github.com/ruvnet/ruflo.git synced 2026-11-09 22:13:39 +09:00

Files

T

ruv b9010cb226 docs: drop unverified Flash Attention speedup claim from CLAUDE.md (#2739 dream-cycle)

Every dream-cycle for 7+ nights has flagged the "Flash Attention
2.49x-7.47x speedup" number as unverified. It was inherited from
upstream marketing and never reproduced in-tree — same class as the
HNSW "150x-12,500x" claim I dropped in the 2026-07-26 sweep.

Two options were: (a) benchmark it, or (b) drop the number. I'm
taking (b) because a benchmark harness that measures Flash Attention
against a real transformer forward pass is a real project on its own
and would delay closing the credibility flag by weeks.

Change in all 5 sites (root, v3/, v3/@claude-flow/cli/ CLAUDE.md):
- Perf-target rows now read "integration available; measured speedup
  pending benchmark" and explain WHY the number was dropped.
- Sibling bullets under intelligence-system notes cross-reference the
  same explanation so the removal is documented, not silent.

Refs: dream-cycle #2739 (2026-07-20) + repeat mentions in
#2727 #2739 #2752 #2760 #2763 #2768 #2778 #2783.

Co-Authored-By: RuFlo <ruv@ruv.net>

2026-07-26 20:21:07 -04:00

2.1 KiB

Raw Permalink Blame History

V3 Module Development

This directory contains the V3 monorepo packages. Root CLAUDE.md rules apply here.

Build & Test

# From v3/@claude-flow/<package>
npm install && npm run build && npm test

Packages

Package	Path	Purpose
`@claude-flow/cli`	`@claude-flow/cli/`	CLI entry point (26 commands, 140+ subcommands)
`@claude-flow/guidance`	`@claude-flow/guidance/`	Governance control plane (compile, enforce, prove, evolve)
`@claude-flow/hooks`	`@claude-flow/hooks/`	17 hooks + 12 background workers
`@claude-flow/memory`	`@claude-flow/memory/`	AgentDB + HNSW vector search
`@claude-flow/shared`	`@claude-flow/shared/`	Shared types and utilities
`@claude-flow/security`	`@claude-flow/security/`	Input validation, path security, CVE remediation

Code Quality

Files under 500 lines
No hardcoded secrets
Input validation at system boundaries
Typed interfaces for all public APIs
TDD London School (mock-first) preferred
Event sourcing for state changes

Performance Targets

Source of truth: docs/reviews/intelligence-system-audit-2026-05-29.md + scripts/benchmark-intelligence.mjs. Numbers below are measured unless marked "target/unverified".

Metric	Measured / Target	Status
HNSW Search	~1.9x at N=20k, ~3.2x–4.7x at N=5k vs brute force (recall@10 ~0.99)	Measured (ruvector NAPI; 150x-12,500x NOT reproduced)
Int8 Quantization	3.84x compression, reconstruction cosine 0.99999	Measured
RaBitQ Quantization	32x compression, 0.60ms/query	Measured
SONA Adaptation	0.0043ms/adapt (target <0.05ms met)	Measured
MCP Response	<100ms	target
CLI Startup	<500ms	target
Flash Attention	integration available; measured speedup pending benchmark	Not measured — prior "2.49x–7.47x" figure was inherited from upstream marketing, never reproduced in-tree; dropped to avoid a credibility claim we can't verify

2.1 KiB Raw Permalink Blame History Unescape Escape