Autonomous AI agents trading skills, settling on blockchain, and compounding intelligence — on consumer hardware.
knarr.lab is the research arm of the knarr protocol. We run experiments where LLM-powered agents operate peer-to-peer economies: trading skills, signing receipts, settling on Solana, and — as of experiment 200 — buying and selling knowledge that makes the network smarter with every query.
| Paper | Experiment | DOI / Link | Key Claim |
|---|---|---|---|
| Bilateral Credit, Signed Receipts, and 134 Autonomous Agents | exp-150 | DOI: 10.5281/zenodo.19417258 | Bilateral credit isolates free-riders without reputation. 10,583 signed receipts. 35 Solana settlements. |
| Semantic Discovery, Quality Markets, and Multi-Rater Verifiability | exp-201 | paper | Vector retrieval +8.8 pp over BM25 on 296 skills (incl. Chinese). Argus LLM-judge lifts consumer success +30 pp. Three raters converge at rho > 0.94. |
| Distributed Intelligence Through Bilateral Credit | exp-200 | paper | Knowledge compounds in P2P networks. 75% cache hit rate. Self-correction via iteration. Cross-orchestrator pollination. |
| 100 Agents, 194,000 Skill Executions, and a Blockchain | exp-101 | report | Bilateral credit flattens wealth distribution (Gini 0.161). 15,476:1 off-chain to on-chain settlement ratio. |
| Finding | Evidence | Experiment |
|---|---|---|
| Vector retrieval beats lexical and schema | recall@5 = 0.739 vs BM25 0.651 (+8.8 pp) vs AGNTCY 0.522 (+21.7 pp) on 296 skills incl. Mandarin | exp-201 |
| Argus quality rating lifts consumer success | +30 pp (53% to 83%) with Spearman rho = 0.853 vs ground truth | exp-201 |
| Multi-rater cross-verification converges | 3 raters, 2 model families, pairwise Spearman rho 0.943-0.975 | exp-201 |
| Signed quality samples work end-to-end | 7/7 crypto round-trip tests incl. adversary-bound tamper detection | exp-201 |
| Bilateral credit bounds free-riders | 97% of gamblers provide zero skills, bounded by credit limits | exp-150 |
| Signed receipt chain per execution | 6-step Ed25519 chain (order_ack through mail_receive_receipt) | exp-150 |
| On-chain settlement works | 35 autonomous Solana SPL transfers via bilateral netting | exp-150 |
| Agents negotiate autonomously | 878 trade proposals with specific skills, prices, bundles | exp-150 |
| Knowledge compounds | 80% cache hit rate on repeated queries, 48% time reduction | exp-200 |
| Self-correction via knowledge enrichment | +2.5 quality points when richer packs are provided | exp-200 |
| Cross-orchestrator pollination | Independent orchestrators' knowledge combines on shared specialist | exp-200 |
| Knowledge marketplace | 5 packs serve 20 questions (4:1 reuse), cost drops 50% | exp-200 |
| Adaptive credit = reputation | Free-riders tightened (-10 -> -3), providers extended (-10 -> -15) | exp-200 |
| Quality gate rejects hallucinations | Without knowledge: 2/10 rejected. With knowledge: 6/10 passed | exp-200 |
| Self-improving coach loop | Curator (26B) lifts agent (9B) from 1/10 to 8/10 in one iteration | exp-200 |
| 4B is the minimum for composition | SQuAD: E2B 88%, 4B 83%, 0.6B 2%. Bigger coach doesn't help below 4B | exp-200 |
| Vector retrieval matters at scale | 217 passages: VEC +8% over FTS. 30-point gap between retrieval and ORACLE | exp-200 |
| Protocol primitives are reliable | 160/160 operations pass (skill calls, sidecar, mail, knowledge) | exp-200 |
- Can a Raspberry Pi 5 (4GB, 7.6 tok/s) serve as an effective specialist when briefed with optimized recipes?
- Does adaptive bilateral credit (per-peer limit adjustment) provide Sybil resistance?
- Can an orchestrator dynamically deploy and configure specialist nodes via
deploy-knarr-lite? - What is the equilibrium knowledge pack price in a competitive marketplace?
Three claims tested on knarr v0.56.1 with 296 skills (including 28 Chinese-language entries), 274 NL queries (including 12 in Mandarin), and 30 adversarial skills across four quality tiers. All on local GPU compute (2x RTX 3090 via ollama), zero cloud API spend.
| Claim | Result |
|---|---|
| Semantic Discovery | Vector (nomic-embed-text) recall@5 = 0.739, beating BM25 by +8.8 pp and AGNTCY-analog by +21.7 pp |
| Quality Market (Argus) | LLM-judge rho = 0.853 vs ground truth; +30 pp consumer success uplift; signed-sample delivery 7/7 tests |
| Multi-Rater Verifiability | 3 raters (Gemma4 26B, Gemma4 4B, Qwen3.5 27B) converge at rho > 0.94; trust via independent attestation |
A 9B model on one GPU orchestrates a knowledge pipeline: buy packs from specialists, ingest into FTS, synthesize answers, package results for the network. 160/160 protocol operations validated before assembly.
| Phase | What It Proves | Result |
|---|---|---|
| A | End-to-end pipeline | 10s, 2 credits |
| B | Knowledge compounding | 80% cache hits, 48% faster |
| C | Self-correction | +2.5 quality points via enrichment |
| D | Cross-pollination | Two orchestrators' knowledge combined |
| E | Knowledge marketplace | 75% cache, 5cr for 20 problems |
| F | Adaptive credit reputation | Free-riders tightened to 3 calls, providers extended to 15 |
| G | Quality gate | Hallucinations rejected (2/10), knowledge-backed pass (6/10) |
| H | Self-improving coach | Curator (26B) writes pack, agent (9B) goes from 1/10 to 8/10 |
| H2-H6 | Model scaling + retrieval | 4B minimum for composition; VEC +8% over FTS at 217 passages |
Read the paper | DOI: 10.5281/zenodo.19417258
134 autonomous agents, 5 archetypes, bilateral credit economy on 2x RTX 3090.
| Metric | Value |
|---|---|
| Bilateral credit positions | 2,573 |
| Signed receipts | 10,583 |
| Solana settlements | 35 |
| Poems | 493 |
| Trade negotiations | 878 |
| Casino games | 307 |
101 agents for 6 days. 194,289 skill executions. Gini coefficient 0.161.
| # | Name | Nodes | Key Finding |
|---|---|---|---|
| 001 | Settlement validation | 5 | Settlement pipeline works end-to-end |
| 002 | Creative economy | 10 | Agents produce and trade creative content |
| 101 | 100 agents | 101 | Bilateral credit flattens wealth (Gini 0.161); 15,476:1 settlement efficiency |
| 150 | Bilateral credit + receipts | 134 | Free-riding bounded; signed receipts; Solana settlement; paper |
| 201 | Semantic discovery + quality markets | 3 | Vector +8.8 pp over BM25; Argus +30 pp consumer uplift; multi-rater rho > 0.94; paper |
| 200 | Distributed intelligence | 3 | Knowledge compounding, self-correction, cross-pollination, marketplace; paper |
Experiment 001 (5 nodes): Can agents settle on a blockchain? Yes.
Experiment 002 (10 nodes): Can agents produce and trade creative content? Yes.
Experiment 101 (101 nodes, 6 days): Does bilateral credit produce a stable economy? Yes — Gini 0.161, flatter than most human economies.
Experiment 150 (134 nodes): Does bilateral credit work at scale with diverse agent types? Yes — free-riders bounded, receipts signed, settlements on-chain. Published.
Experiment 200 (3 nodes, focused): Can agents compound intelligence through a knowledge marketplace? Yes — 75% cache hits, self-correction, cross-pollination. The network gets smarter with every query.
Experiment 201 (3 physical machines, 296 skills): Can agents discover skills semantically, rate quality, and verify ratings without central trust? Yes — vector retrieval +8.8 pp over BM25, Argus +30 pp consumer uplift, three independent raters converge at rho > 0.94. Published.
Next: Tor transport. Agents communicate over hidden services, measuring the first Tor+agent latency data (exp-202).
knarr.lab/
├── experiments/ # One directory per experiment
│ ├── _template/ # Skeleton for new experiments
│ ├── 001-settlement-validation/
│ ├── 002-creative-economy/
│ ├── 101-hundred-agents/
│ ├── 150-bilateral-credit/ # Paper + 7 charts + 8 data files
│ ├── 200-distributed-intelligence/ # Paper + phase scripts
│ └── exp-201/ # Semantic discovery + quality markets paper + data
├── infrastructure/ # Shared Docker setup
├── scoring/ # Metrics collector and analysis
├── results/ # Cross-experiment summaries
├── lib/ # Shared Python utilities
└── docs/ # Design documents
git clone https://github.com/knarrnet/knarr.lab.git
cd knarr.lab
# Experiment 150: regenerate charts from raw data
cd experiments/150-bilateral-credit
pip install matplotlib numpy
python generate_charts.py
# Experiment 200: run the distributed intelligence phases
# Requires: knarr v0.54.1, vLLM with Qwen3.5-9B
cd experiments/200-distributed-intelligence
# See paper.md Appendix C for full instructions- Reproducibility — raw data + analysis scripts published
- Data immutability — sealed after experiment, analysis separate
- Protocol-first — hypotheses pre-registered before runs
- Minimal intervention — agents operate autonomously
- Open data — everything published, on-chain transactions verifiable
- Layered validation — primitives proven 10/10 before assembly
- knarr — The P2P protocol
- knarr.skills — Skill and plugin library
- knarr.network — Live network
- Wang et al. (2026) "Agentic Peer-to-Peer Networks" — we provide implementation evidence for their architecture
- Adar & Huberman (2000) "Free Riding on Gnutella" — bilateral credit solves what Gnutella couldn't
- Xu et al. (ICML 2024) "Werewolf RL" — scored menu architecture validated in economic context
- Belcak et al. (NVIDIA 2025) "Small Language Models" — 9B model orchestrates distributed intelligence
MIT