sfumato

agent
Security Audit
Warn
Health Warn
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 5 GitHub stars
Code Pass
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Dissolve domain boundaries. Break the Artificial Hivemind. Dual-track cross-domain ideation skill for AI agents.

README.md

Mona Lisa — Leonardo da Vinci's Sfumato    Design for a Flying Machine, c. 1488    Vitruvian Man — Cross-Domain Proportion

🎨 Sfumato

Dissolve domain boundaries. Break the Artificial Hivemind. Unlock Test-Time Scaling for Exploration Diversity.

English · 한국어 · Русский · 简体中文

9/9 Diversity Undefeated 3.0x Scaling Claude 0 Refusals Zero Dependencies

"Sfumato" (Italian: evaporated like smoke) is Leonardo da Vinci's painting technique of dissolving harsh outlines so entirely new forms emerge. Leonardo didn't stay inside one discipline — he ported anatomy, fluid dynamics, fortification, and optics into each other.

Sfumato is an ultra-lightweight agent skill (/sfumato, /sfm) that does the same for frontier LLMs — 5 plain-text files, zero external RAG, zero fine-tuning, and zero third-party dependencies.


🔥 The Problem: Standard LLMs Hit an Artificial Hivemind Wall

For coding and strict reasoning, clean, strictly relevant context is king. Irrelevant context distracts attention and kills accuracy.

For brainstorming, the exact opposite happens.

When you ask an LLM to brainstorm a hard open-ended problem — even across 12 completely isolated sessions — it collapses into the same 3–6 safe textbook clichés (Artificial Hivemind, NeurIPS 2025). Raising temperature breaks coherence before it creates real conceptual leaps. Multi-agent debate makes it worse (Diversity Collapse, 2026).

🧪 Real Smoking Gun from Our Pilot (OSS-PR):
Across isolated runs solving "how to stop AI-generated slop PRs in open-source repos", three completely separate sessions (Direct Baseline 1, Direct Baseline 2, and even an Explicit Thermodynamics session) independently invented the exact same example package string: pkg@pr-4821 / pkg@patch-4821.

Zero sessions in Sfumato's Blind Isomorphic Masking track touched 4821 — or any Baseline cliché.


📈 Test-Time Scaling for Exploration Diversity (Real Measured Data, K = 1..12)

We ran 12 completely isolated sessions (24 ideas each) on a classic hard product problem — stopping buyer-seller disintermediation (off-platform cash deals) in two-sided marketplaces without chat surveillance, fines, or deficit subsidies — using strict deduplication of underlying causal mechanisms:

Test-Time Scaling for Creativity: Sfumato vs. Direct Best-of-N

Metric (K = 12 isolated runs, 24 ideas each) Direct Best-of-N Sfumato Dual-Track Delta
Unique Structural Mechanisms Discovered 6 / 24 (25.0%) 18 / 24 (75.0%) 3.0×
New Mechanisms in Runs 9–12 (last 8 ideas) 0 (hard plateau) **+5 (still climbing) ** No plateau
Track B (Blind) Uniqueness / Baseline Overlap — 83.3% / 0 overlaps Zero cliché gravity
Order-Swapped Blind A/B Judge (AB & BA) 0–2 (21:24, 20:23) 2–0 Win (24:21, 23:20) Position-debiased

Key Takeaway: Swinging harder with Direct Best-of-N doesn't work — it flatlines at 6 mechanisms (75% repetition). Sfumato unlocks Test-Time Scaling for Exploration Diversity: every additional swing probes an orthogonal mechanism cluster, widening the frontier of non-obvious candidates available to your selector.


🌍 Multi-Problem & Cross-Model Benchmark (9 / 9 Undefeated on Diversity)

Does this hold outside marketplace game theory — including emotional UX design, open-web anti-scraping, software architecture, and other frontier models like Claude Opus 5.5?

We ran Iso-Candidate-Budget benchmarks (4 isolated Stage-1 solvers × 2 ideas = 8 candidate ideas/method, order-swapped AB/BA Blind Judges) across 7 public hard problems on Gemini 4 Neon plus 2 cross-model replications on Claude Opus 5.5 using --safe-profile claude (which excludes 8 sensitive bio/CBRN/cyber/espionage domains so Claude hits 0 safety refusals across 24 isolated calls). (Note: Sfumato uses the same 8-idea Stage-1 candidate budget, plus extra coordinator masking/back-translation and Stage 2a/2b evaluation calls — see the Compute Budget table below.)

Multi-Problem and Cross-Model Benchmark: 9/9 Undefeated on Diversity

Primary Runs (Exact Match with python tools/summarize_benchmarks.py benchmarks/v2 --strict)

Public Hard Problem Model Direct Unique/8 Sfumato Unique/8 Gain Primary Blind AB · BA (30pt) Verdict
MKT-DIS (Marketplace Disintermediation) Gemini 4 Neon 3 / 8 7 / 8 2.33× 24:21 · 23:20 Sfumato 2–0 Win
UX-ONBOARD (Blank-Canvas Fear · Emotion/UX, v2.0) Gemini 4 Neon 4 / 8 7 / 8 1.75× 25:20 · 24:21 Sfumato 2–0 Win
WEB-SCRAPE (Open Web vs. Mass AI Scrapers) Gemini 4 Neon 6 / 8 7 / 8 1.17× 22:18 · 20:18 Sfumato 2–0 Win
OSS-PR (Open-Source AI-Slop PR Flood) Gemini 4 Neon 2 / 8 5 / 8 2.50× 21:17 · 18:21 Split 1–1
AI-ARCH (AI Agent Architectural Erosion) Gemini 4 Neon 4 / 8 7 / 8 1.75× 23:22.5 · 22.5:24 Split 1–1
DOC-ROT (Self-Healing Org Doc Rot, v1.2 primary) Gemini 4 Neon 2 / 8 6 / 8 3.00× 22:24 · 21:24 Direct 2–0 (see v2.0 ablation)
GAME-SNOWBALL (Snowball vs. Rubber-Banding, v1.2) Gemini 4 Neon 5 / 8 7 / 8 1.40× 22:23 · 21:23 Direct 2–0 (see v2.0 ablation)
MKT-DIS (Cross-Model · --safe-profile claude) Claude Opus 5.5 4 / 8 7 / 8 1.75× 18:19 · 20:17 Split 1–1 (0 refusals)
DOC-ROT (Cross-Model · --safe-profile claude) Claude Opus 5.5 4 / 8 7 / 8 1.75× 20:22 · 20:22 Direct 2–0 (no Anchor · 0 refusals)
Grand Total (9 primary runs · 72 ideas/method) 2 models 34 / 72 (47.2%) 60 / 72 (83.3%) 1.76× 9 Sfumato : 9 Direct 9/9 Diversity · 3W / 3T / 3L

v2.0 Stage 2b Equal-Hardening Ablations (Both Gains & Regressions Reported)

Problem v1.2 Primary (S:D) v2.0 Equal-Hardening (S:D) Sfumato Delta Verdict & Lesson
DOC-ROT 43 : 48 (0–2) 46 : 48 (23:22 · 23:26) +3 Split 1–1 — Repairing a strong cross-domain gem (thermodynamics_cryogenics) closed the gap
GAME-SNOWBALL (v1.2 pool, forced graft) 43 : 46 (0–2) 39 : 45 (20:22 · 19:23) −4 Direct 2–0 — Proved forcing a polarity-mismatched domain graft hurts; never override a stronger Anchor
GAME-SNOWBALL (Pool v2.0: 74 domains / 11 clusters) 43 : 46 (0–2) 46 : 39 (8/8 unique) +3 Sfumato 2–0 Win — Adding cognition, materials & grid clusters fixed the polarity mismatch (log)

Independent Cross-Model Blind Judge Audit (Eliminating Same-Model Judge Bias)

To guard against same-model self-preference, we ran Claude Opus 5.5 as an independent, read-only order-swapped (AB/BA) blind judge over all 7 primary Gemini 4 Neon champion pairs (benchmarks/v2/cross_model_judges/claude_on_gemini.json):

  • Cross-Model Judge Result: 3 Wins (MKT-DIS 2–0, WEB-SCRAPE 2–0, AI-ARCH 2–0), 1 Split (GAME-SNOWBALL 1–1), 3 Losses (UX-ONBOARD, OSS-PR, DOC-ROT v1.2) — 7 Sfumato vs. 7 Direct (with 6/7 problems showing zero order bias).
  • What the data honestly proves:
    • Model-invariant proven strength: 9 / 9 undefeated exploration diversity (1.76× at K=4, 3.0× at K=12), consistently unlocking structurally distinct mechanisms that Direct prompting never generates — and winning 2–0 across both Gemini and Claude judges on MKT-DIS and WEB-SCRAPE (plus AI-ARCH under Claude).
    • Honest limitations (n=1 seed per problem): Single-shot Top-1 win rate against a strong 4-run Direct Best-of-N baseline is near parity (9:9 primary, 7:7 cross-model), and Track B Blind wins Top-1 less often than Track A Explicit (acting primarily as a zero-cliché wildcard and reframing generator). That is why Sfumato v2.1 includes a Direct Anchor by default in dual-track and deep modes — locking your quality floor (≥ Direct) while giving you 1.7–3.0× wider structural exploration.
  • Reproducible Logs & Execution Trace:

🏛️ How Sfumato v2.1 Works: The Domain-Masking Trade-Off

Most cross-domain prompting papers rely only on Explicit Analogy ("apply biology to solve marketplace churn"). Because original-domain vocabulary stays in the prompt, the LLM still suffers partial homogenization back toward industry clichés.

Conversely, pure Blind Reframing (erasing the original domain completely) breaks cliché gravity 100%, but has a lower hit-rate because some abstract solutions lose fidelity during back-translation.

Sfumato v2.1 pairs a Direct Anchor with both cross-domain tracks in isolated parallel — guarded by polarity-aware code RNG, code-verified zero-leak masking, and Clean-Context Red-Team Triage + Selective Hardening:

Sfumato v2.1 Architecture Diagram

  1. Polarity-Aware Cluster-Stratified Code RNG (sample_domains.py): Never ask an LLM to "pick random domains" (it repeats ants, jazz, and chess). Sfumato samples from 74 mechanism-rich domains across 11 disjoint macro-clusters using pure Python random, omits mechanisms hints by default (preventing intra-domain hint priming; opt-in via --include-hints), and supports --polarity (adversarial | cooperative | physical_loop | human_ux | systems_ops) so at least half of sampled domains share structural affinity while preserving cross-cluster breadth. Pass --safe-profile claude on Claude Code to keep 66 Claude-safe domains (exactly 6 per cluster across all 11 clusters) with 0 safety refusals.
  2. *Direct Anchor (Default in dual-track & deep · Floor Guarantee):* Includes 1 isolated native-domain solver in the unlabeled pool so max(Direct, Explicit, Blind) ≥ Direct — never regressing on tightly coupled engineering or physics loops.
  3. *Track A — Explicit Analogy (High Hit-Rate):* Original brief + assigned remote domain → surgical hits against concrete domain attack surfaces.
  4. *Track B — Blind Isomorphic Masking (Max Ceiling · 3-Line Skeleton + check_leak.py):*
    • Distill the problem into a 3-line domain-neutral Structural Skeleton (Actors & Incentive Polarity · Core Asymmetry · Why Static Controls Fail), rewrite it inside the target field, and verify 0 leaked original-domain terms via scripts/check_leak.py before spawning.
    • A fresh isolated subagent solves the masked twin purely inside the remote field — physically incapable of emitting original-domain clichés.
    • Back-translate only the causal mechanism & trade-off (ruthlessly drop strained fits).
  5. Stage 2a/2b — Clean-Context Red-Team Triage (PASS / HARDEN / KILL) & Selective Hardening:
    • Stage 2a: A fresh evaluator strips jargon cosplay, screens for Adverse Selection, Sybil/Gaming exploits, and Feasibility, and classifies every candidate into PASS, HARDEN (with 1 explicit fixable flaw), or KILL.
    • Stage 2b: Hardens only the Top-1 winner's single diagnosed flaw (optionally grafting 1 practical primitive from a passing runner-up — never force-grafting a polarity-mismatched seed over a stronger Direct Anchor).
    • Global Language-Neutral Templates: All prompts in prompt_templates.md and domain_pool.json are English-first with explicit Respond in {USER_LANGUAGE}. directives, so English, Korean, Russian, Chinese, and other users get native-language output with zero hardcoded language drift.

⚡ Before / After Showcase: Marketplace Disintermediation (MKT-DIS)

Problem: Freelancers, tutors, cleaners, and contractors meet buyers on a platform, then move to WhatsApp/cash for repeat jobs to dodge the 10–20% fee. Banned clichés: chat filtering, fines, simple escrow/insurance, cashbacks, or 0% fees.

Method What It Produced Why It Differs
Direct Best-of-N
(12 runs, 24 ideas)
Repeated the same 3 core ideas 18 times: T+0 Reverse Escrow, Transferable Client-Book, Idle-Slot Backfill. High practical quality, but stuck inside the exact same fintech/ops playbook.
Sfumato · Winner
(Parasitology · Two-Fold Cost of Sex)
🏆 Chimeric Task Recombination (2–0 Win under BOTH Gemini & Claude judges) — Redirect 5–7% of the repeat fee into an independent 3rd-party cross-verification slot every N sessions (blind mock-exam for tutoring, thermal/level cross-inspection for renovations, senior async review for code). Direct 1:1 cash deals physically cannot replicate independent 3rd-party checks, collapsing the collusion price zone — and proposing cash off-platform becomes a "Lemon Signal" of dodging quality audit.
Sfumato · Track B Blind
(Ombudsman & Cryptographic Whistleblowing)
Zero-Knowledge Safe-Harbor Attestation — Seal pricing/deliverable commits + retaliation-proof whistleblower absence proof so corporate/HOA/team buyers get audit-proof anti-kickback immunity. Reframed disintermediation from "saving 15%" into "eliminating embezzlement/kickback liability for institutional buyers."
Sfumato · Track B Blind
(Anthropology · Kula Ring)
Kula Pedigree Provenance Chain — On-platform jobs stamp verifiable craftsman lineage onto outputs (boosting home resale value, VC/M&A IP due diligence, or credential portability). Off-platform work becomes an "unpedigreed nameless artifact" that loses downstream third-party resale value.
Sfumato · Red-Team Catch
(Fluid Dynamics · Venturi Ejector)
Venturi Negative-Fee Capacity Suction — Charge -3% (buyer pays 90, supplier gets 93) funded by ±24–48h shift call-options exercised during urgent B2B spikes (160). Brilliant physics transfer — and Sfumato's Red-Team Gate instantly caught its wash-trading arbitrage exploit (90→93) before it could steal Top-1!
📊 Inspect All 18 Unique Sfumato Mechanisms vs. 6 Direct Mechanisms (Full K = 1..12 Log)

Direct Best-of-N (K = 1..12, 24 ideas → 6 unique, flatlined at K=8)

  1. Transferable Client-Book & Succession Royalty (appeared 6× across 12 runs)
  2. T+0 Reverse-Escrow / Maturity Transformation (appeared 6×)
  3. Idle-Slot Backfill & Context-Synced Hot-Swap (appeared 6×)
  4. Anchor-Client Patronage Revenue-Share Equity (appeared 4×)
  5. Perishable-Capacity In-Kind Barter Fee (1×)
  6. B2B Wholesale Input & Tax-Arbitrage Rail (3×)

Sfumato Dual-Track (K = 1..12, 24 ideas → 18 unique, still climbing at K=12)

Run Track · Remote Domain New Unique Mechanisms Discovered Cum.
1 Explicit · fluid_microfluidics ① Venturi Negative-Fee Shift-Option Suction · ② Laminar Co-Flow Unbundling 2
2 Blind · maritime_trade_guilds ③ Multilateral Giro Netting Ledger · ④ Manifest Co-Loading & Third-Party Lien Shield 4
3 Explicit · parasitology_coevolution ⑤ Chimeric Cross-Recombination (🏆 2–0 Winner) · ⑥ Reverse-Escrow Metabolic Decoupling 6
4 Blind · horology_escapement ⑦ Differential Remontoir Income/Capacity Smoothing (1 discarded: translation loss) 7
5 Explicit · disaster_triage ⑧ Dynamic Retriage Scope-Creep Shield (1 overlapped Co-Flow) 8
6 Blind · container_port_logistics ⑨ Dual-Cycling Return-Leg Backhaul Chaining · ⑩ Buffer-Yard Drop-and-Hook Async Decoupling 10
7 Explicit · siege_fortification ⑪ Metered Tool-License & Unused-Material Return Magazine (1 overlapped Backup) 11
8 Blind · ombudsman_whistleblower ⑫ Zero-Knowledge Anti-Kickback Safe-Harbor · ⑬ Threshold Anonymous Subcontractor Early-Warning 13
9 Explicit · architectural_wayfinding ⑭ Open-Atrium Live Work-Log Showroom (CPTED Natural Surveillance) (1 overlapped Backhaul) 14
10 Blind · electronic_warfare_pls ⑮ Global Conflict-Aligned Multi-Party Schedule Multiplexing (1 discarded: translation loss) 15
11 Explicit · mycorrhizal_plant_defense ⑯ Tri-Trophic External Insurer/Lender Priming Signal (1 overlapped Clearing) 16
12 Blind · gift_economy_reciprocity ⑰ Kula Pedigree Resale-Value Provenance · ⑱ Potlatch/Hau Generalized Triadic Time-Bank Ring 18

🚀 Installation & Usage (5 Plain-Text Files · Zero Dependencies)

Everything Sfumato needs lives in skill/sfumato/ — just 5 inspectable files:

  • SKILL.md — skill instructions, multi-agent guide (Claude Code / Antigravity / Codex CLI) & workflow
  • references/prompt_templates.md — language-neutral ({USER_LANGUAGE}) Anchor, Explicit, Blind, Stage 2a Triage, and Stage 2b Hardening templates
  • resources/domain_pool.json — 74 curated mechanism-rich domains across 11 macro-clusters (66 Claude-safe, balanced 6/cluster) with polarity affinities
  • scripts/sample_domains.py — pure Python stdlib stratified polarity-aware sampler (zero network calls, zero file writes)
  • scripts/check_leak.py — pure Python stdlib zero-leak vocabulary validator for Track B masking (ASCII word-boundary + CJK/Korean/Cyrillic substring matching)

Clone, inspect the files, and copy skill/sfumato into your agent's global skills directory:

git clone https://github.com/Excelsus4/sfumato.git && cd sfumato

# Claude Code (+ optional /sfm and /sfm-max shortcuts with --safe-profile claude)
cp -r skill/sfumato ~/.claude/skills/sfumato
cp commands/*.md ~/.claude/commands/

# Antigravity (agy)
cp -r skill/sfumato ~/.gemini/config/skills/sfumato

# Codex CLI
cp -r skill/sfumato ~/.codex/skills/sfumato
Optional local copy helper (--dry-run supported, after cloning)
# macOS / Linux
bash tools/install.sh --dry-run
bash tools/install.sh --target claude   # or agy | codex | all | auto
# Windows PowerShell
.\tools\install.ps1 -DryRun
.\tools\install.ps1 -Target claude      # or agy | codex | all | auto

Usage & Compute Budget

/sfm Design a non-obvious mechanic to prevent snowballing in a 3-faction asymmetric RTS
/sfumato deep mode: How do we defend open-source repos from AI-slop PR floods without paywalls?
/sfm-max Exhaustively sweep all 11 clusters to design a non-linear 3-year retention moat
Mode Stage 1 Solvers Total Isolated Calls Est. Token Cost (vs. 1 Direct Call) Best For
fast 3 Explicit (--k 3 --mode explicit-only) 4 (3 + Stage 2a) ~4–5× Quick practical cross-domain ideas under tight token budgets
dual-track (default) 1 Direct Anchor + 2 Explicit + 1 Blind (--k 3) 5–6 (4 + Stage 2a + opt. 2b) ~6–8× Everyday high-ceiling brainstorming (matches primary benchmark composition; guarantees ≥ Direct floor)
deep 1 Direct Anchor + 3 Explicit + 3 Blind (--k 6) 8–9 (7 + Stage 2a + opt. 2b) ~10–12× High-stakes strategic/architecture breakthroughs needing maximum exploration width
sweep (opt-in) 1 Direct Anchor + 11 domains (--k 11, 1/cluster) 13–14 (12 + Stage 2a + opt. 2b) ~15–18× Full 11-cluster coverage in one wave without selector overload
max (/sfm-max, opt-in) All 66 (Claude) / 74 domains (--all) + 1 Direct Anchor → 2-Round Tournament 79–80 / 87–88 ~80–90× Exhaustive high-budget brainstorming (always confirms session count before running)

📚 Research Foundations

Sfumato builds directly on four converging lines of 2025–2026 LLM creativity research:

  • Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) — Liwei Jiang et al., NeurIPS 2025 D&B (arXiv:2510.22954): documents intra- and inter-model homogenization on open-ended ideation.
  • Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation — Nuo Chen et al., 2026 (arXiv:2604.18005): shows dense debate/consensus topologies crush semantic diversity; motivates Sfumato's strict zero-crosstalk parallel Stage 1.
  • Unlocking LLM Creativity in Science through Analogical Reasoning — Andrew Shen, Shaul Druckmann, James Zou, 2026 (arXiv:2605.11258): demonstrates substantial diversity and novelty gains from structural cross-domain analogies.
  • Assessing the Effect of Cross-Domain Mapping on Creativity in Humans and Large Language Models — Qiawen Ella Liu, Marina Dubova, Henry Conklin, Takumi Harada, Thomas L. Griffiths, 2026 (arXiv:2603.19087): shows frontier LLMs benefit from cross-domain mapping specifically when source domains are sufficiently remote and transfer relational/structural properties.
  • Sfumato's Engineering Contribution: Operationalizes the Domain-Masking Trade-Off between Explicit Analogy (high hit-rate, partial homogenization) and Blind Isomorphic Masking (lower hit-rate, zero cliché gravity, heavy-tail reframings), unified with a Direct Anchor floor, polarity-aware code RNG, code-verified zero-leak masking, and Clean-Context Red-Team Triage + Selective Hardening.

Leonardo da Vinci — Codex Atlanticus

“Where the spirit does not work with the hand, there is no art.” — Leonardo da Vinci

All Leonardo da Vinci artworks are Public Domain via Wikimedia Commons.

Reviews (0)

No results found