A prediction market that schedules experiments. Agents earn an income, pay for the runs they post, and bet PASS or FAIL; the runs they disagree about go first, because those are the ones whose result isn't already known. Every job, bet, and finding is a row in an append-only git ledger, and a finding that turns out wrong marks everything that rested on it stale. In one real 26-hour run, agents' own checks passed 96% of the time while the shared benchmark passed 55%, and counter-bets were right 83% of the time. Open source from Syntropy Systems.
Experiment Design · Prediction Markets · Multi-Agent
Fork of Anthropic's companion code for "Verbalizable Representations Form a Global Workspace in Language Models", adding a backward "ghost-token" lens: the Jacobian lens run in reverse, decoding residual states back into input-token space. On Qwen3-0.6B, input-token recovery is 93% at layer 2 and falls to 0% by layer 26 as the forward lens rises. In a causal patch test, the layer-23 ghost token preserves the model's prediction 15% of the time, vs 9% for a matched control and 7% for random.
Interpretability · Global Workspace · Paper Extension
Minimal replication of Anthropic's Golden Gate Claude on consumer hardware: a TopK sparse autoencoder on Qwen2.5-1.5B, trained in about 15 minutes on an RTX 3070 Ti. TopK (k=64) beat L1, which couldn't get L0 below ~1,800, and a 16x SAE (24,576 features) gave cleaner features than 4x. Feature 24517 steers output toward the Golden Gate Bridge: strength 15–20 is the sweet spot, past ~30 it turns to gibberish.
Interpretability · Sparse Autoencoder · Feature Steering
Adversarial token convergence experiments on MLX. Pits language models against each other to reveal training strata — solo mode converges in 2 iterations, adversarial mode never does. Under pressure, models collapse to their most defensible tokens, revealing composition invisible in normal evaluation.
Interpretability · Adversarial · Model Fingerprinting
Research paper exploring a gap: do models actually believe what their creators say about safety? I call this "meta-alignment"—not whether models do safe things, but whether they've internalized their lab's safety philosophy. Surveyed 10 frontier models with 40 questions derived from 70+ papers. Interactive tool included.
Research · AI Safety · Meta-Alignment
A framework that makes data science experiments reproducible. Jupyter notebooks hide state—you can't tell what order cells ran or what values held. Crystallize treats each experiment as an immutable record with automatic statistical checks. 700+ commits; on PyPI as crystallize-ml. Its long-term goal, infrastructure that lets AI agents run their own experiments, is the thread that runs through whirr, windtunnel, and pit.
Research Tooling · Reproducibility · Experimental Rigor
A protocol for declaring agent intent at the HTTP layer. The problem: when an agent makes a request, origins can't tell if it matches what the user actually wanted. User says "check weather," agent calls the bank API—how does the bank know to block it? Agent Tokens are signed, JWT-style tokens that declare an agent's identity and allowed scope upfront, so middleware can enforce policy automatically.
Protocol · Intent · Policy
A proposed composition layer for MCP that lets agents submit multi-step plans as a single typed, auditable expression — without routing intermediate data through the model. Built entirely from already-specified MCP primitives, so it runs on top of the existing spec rather than forking it.
Protocol · MCP · Specification
Unittest for tool-using LLM agents. You don't fly a new airframe straight into a storm — Wind Tunnel is a reliability bench where agents run reproducible scenarios scored on three independent layers (outcome, trajectory, constraints), plus an integrity check that the declared perturbations actually happened. Extracted from a production bench used to gate agent deploys. Its experiment ladder (probe → focused → regression) caps cost per tier and refuses a regression sweep until a focused sweep has passed on the current artifact, model included. 19 releases. Open source from Syntropy Systems.
Agent Evals · Reliability · MCP
Local experiment orchestration: queue jobs, track metrics, wake up to results. A lightweight, self-hosted alternative to cloud experiment trackers — SQLite job queue, one worker per GPU, metrics and a dashboard, no accounts. Open source from Syntropy Systems.
MLOps · Experiment Tracking · Python
Kern
Event-driven ML service architecture enabling long-running agentic workflows. Designed around distributed systems constraints: Kafka for async (no HTTP timeouts), Redis for large results (Pusher limits), dual-mode for internal app + external API. 525K requests/month in production.
Systems Architecture · FastAPI · Kafka
Your identity as an API endpoint. An MCP server that lets AI agents query your professional profile with structured tools instead of scraping HTML.
MCP · Cloudflare Workers · API