Featured
Agent Memory Needs Typed Claims, Routing, and Maintenance
A three-month GBrain + Hermes run: why raw transcripts are not memory, and the typed-substrate, routing, and maintenance loop that made durable recall inspectable.
The Lab
Decision guides for the software and AI you're considering — what works, what it costs, and where it fails.
8 articles
Your AI Ops Bot Is Not Your Co-Founder: Guardrails First
Letting an agent touch production needs blast-radius limits, not trust. The guardrails to set first, the failure modes they stop, and where to keep a human.
LLM Fine-Tuning: Your Dataset Is the Hard Part, Not the Run
For an internal smart-contract audit SFT lane, hold the run until the dataset passes dedup, leakage, grounding, negative-coverage, and truncation gates.
Jentic One, Tested: Three Install Failures, Then a Sandbox Pilot
A shared-host Jentic One install hit release drift, a false health check, and an admin gate before any governed API call; here is what each one costs.
96GB Blackwell: SGLang NVFP4 + DFlash2 vs vLLM BF16, Measured
Measured Qwen3.8-27B serving on one 96GB Blackwell GPU: a BF16 vLLM baseline, an NVFP4 SGLang lane, and the throughput and latency that decide which holds.
Self-Hosted LTX 2.3: Failure Modes and Trade-Offs
Receipt-backed LTX 2.3 and ComfyUI workflows show the GPU, retry, and failure-mode trade-offs behind self-hosting versus an occasional hosted API call.
vLLM MTP: Faster Inference That Quietly Broke Tool Calls
Multi-token prediction cut our latency, then quietly corrupted tool calls. What MTP does, how the failure presented, and the settings that made it safe again.
The Workflow Decision Lab's Four-Part Method for Buying Decisions
The Lab tests AI and software buying decisions with a named scenario, a verdict, visible workflow and failure modes, a keep-or-drop rule, and receipts.