Subscriber field kit + measured lab notes
Stop guessing which local AI run will fit, fail, or cost more.
Get the Local AI Field Kit now. Then get short field notes built from real runs on owned GPUs, including the misses.
Model fit worksheet
Estimate weights, KV cache, and headroom before a large download.
Repeatable benchmark card
Record the model, quant, prompt, hardware, speed, power, and stop reason.
Local agent failure checklist
Separate model, runtime, tool, memory, and GPU failures without guessing.
API or local decision sheet
Compare privacy, latency, utilization, and real per-run cost.
Get the Local AI Field Kit
Four copy-ready tools now, then one evidence-backed Local AI Lab Note on Friday when there is something worth sharing.
Get the requested artifact now, then at most one evidence-backed Local AI Lab Note on Friday when there is something worth sharing. One-click unsubscribe. No sponsored placements. Privacy.
Try one check. No signup needed.
Before you trust an agent's “done”
Give your next local or cloud agent task a clear finish line. Copy this template, fill in the brackets, and check the result. It is part of the field kit you receive by email.
# Before you trust an agent's "done" Source: https://bmdpat.com/newsletter#try-the-check Use this with a local or cloud agent. Fill in the brackets before the run. Task: [the result I need] Inputs: [the files or sources the agent may use] Allowed changes: [the exact scope] Stop condition: [time, cost, or retry limit] Before work, define a check for each requirement: | Requirement | How I will check it | What failure looks like | | --- | --- | --- | | [result] | [command, comparison, or human review] | [specific failure] | After work, report each requirement as PASS, FAIL, or UNVERIFIED. For each PASS, give the artifact and the actual check result. File existence alone does not prove the content is correct. Keep failed results. Do not invent a passing result or silently weaken a check. Name anything that still needs a person to review it. This is a review template, not an automatic verifier. Run the checks yourself or through a trusted test runner. An agent can still misreport a result. Do not paste secrets or private customer data into a public assistant.
Useful to someone on your team? Share this public sample. Your private kit link stays in your inbox.
What lands after the kit
One useful result. The setup. The failure. The decision.
I run open models on my own GPUs and publish the results. Each note names the model, quant, workload, hardware, and result when those details exist. No result means no padded email.
Sample note
A 26B model paid 140 seconds for one context change.
The useful part is the run conditions, why the reload happened, and the check you can apply before your next job.
Public proof before signup
Read the 5090 Reports first. They are the public artifact trail behind the private kit and email promise.
Read 5090 Reports