Large-context research
Experimental long-context configurations on local hardware, validated with multi-needle recall tests. Boundaries — where a model or runtime caps what is possible — are documented rather than hidden.
Kraken Research Lab
Experiments, measured honestly — including the failures.
The Research Lab is where Kraken Systems publishes experimental evidence: model research, large-context work, heterogeneous compute, autonomous-agent research, benchmarking, and negative results. Claims here are traced to real measurements on self-hosted infrastructure — no marketing gloss.
Current focus areas
What the lab works on.
Heterogeneous compute
Engineering across mixed-generation accelerators managed as a single fleet: lifecycle management, thermal automation, throughput measurement, and honest comparisons between options.
Autonomous-agent research
Coordinated multi-agent architectures with human gates: inter-agent protocols, anti-loop controls, signed webhooks, and read-only validation — verified end-to-end before any production use.
Benchmarking
Measured model and tooling comparisons on real hardware — including the negative results: the community fork that was ~8.5× slower and was rolled back, and the build that produced blank tokens and was rejected.
Model research
Local model evaluation, quantization and configuration work, and capability mapping on constrained hardware — what runs well where, and what does not.
Trading research (paper-only)
Market and trading research conducted in controlled, paper-only environments. No real capital is traded; results are shared as research, never as investment advice.
Evidence, not marketing
Selected measured results.
Every item below is backed by real measurements in the build journal and internal evidence store. Figures are rounded and sanitized — no internal addresses, identifiers, or credentials.
Long-context validation on 16 GB cards
Multi-needle recall validated at 80K and 128K tokens on a single Tesla P4 (3/3 correct retrievals, 80–81 °C peak, 44–55 W under load). The 196K attempt was documented as a model/runtime training-context cap, not a hardware failure. On the 4× P100 node, a 35B-A3B model loads at 524K context and a 27B model is validated at 196K with real inference.
Multi-GPU engineering on 4× Tesla P100
A 64 GB P100 node (2× Xeon Gold 5122, 187 GiB RAM) runs measured model comparisons — 30B-A3B at ~50 tok/s vs ~31 tok/s for a larger alternative — and hosts a 284B-parameter deployment plan with 524K-context KV planning across all four cards. A naive scheduler claimed ~150 GiB for a workload that measured ~51 GB — a 3× overestimate, documented and corrected.
Thermal and lifecycle automation
Idle lanes unload after 15 quiet minutes (VRAM returns to ~7 MiB, GPU drops to its lowest power state) and auto-wake on demand with identity restoration and GPU-binding verification. A latent bug that had silently disabled unloads was found and fixed. Fans auto-quiet to ~5,500–6,700 RPM at idle.
Heterogeneous fleet, one registry
A single lane registry manages mixed-generation accelerators across multiple nodes. No node is considered production-ready until host-level and container-level GPU visibility both pass fleet-wide validation.
Autonomous agent architecture, gated
Specialized agents produce, challenge, validate, and repair one another's work through a defined PLAN → IMPLEMENT → CHECK → AUDIT → REPAIR → REVALIDATE → IMPROVE cycle. Inter-agent transport uses the A2A v1.0 protocol with per-peer tokens, anti-loop caps, and signed outbound webhooks; a read-only validator holds veto power. Verified end-to-end in dry-run with zero API calls.
Negative results are published too
A community inference fork promising speedups was A/B-tested on a live lane: ~4.8 tok/s vs ~41 tok/s stock — roughly 8.5× slower. The rollout was stopped and the lane restored. A Qwen3.6-35B build produced blank tokens on P100 and was rejected. What failed, what changed, and what was learned are all recorded.
Experimental by design. Research Lab results describe experiments on self-hosted infrastructure — they are not product certifications, performance guarantees, or endorsements of any model, tool, or vendor. Context sizes are reported as measured on specific hardware and workloads.
Trading research is paper-only. Trading research at Kraken Systems is conducted paper-only. We do not trade with real capital, and nothing we publish is investment advice or a solicitation to buy or sell any security. Our methods and results are shared as research — not as guidance. If you act on anything you read here, that is your decision and your responsibility.