Frozen evidence
Every release names the dataset, split, licence, evaluator version, prompt or rule, and execution date.
Measured reports
Engineering evidence that can survive a release meeting
Reproducible tests for the gates between a promising model and a production system. Raw results and methods included.
Measured releases
Benchmarks with downloadable evidence
Frozen datasets, explicit selection rules, tail latency, limitations and raw results.
REPORT 01 · EXECUTED 2026-08-23
Which release gates catch unsupported answers?
2,675 held-out responses. Precision, recall, F1, false-negative rate, local latency, raw decisions and the configurations that failed.
Read the report
REPORT 02 · EXECUTED 2026-08-24
How much accuracy does lower-resolution object detection trade away?
5,000 COCO validation images. Twelve configurations across M4, Core ML, L4 CUDA and C4 ONNX, with 2,400 timed inferences and capacity scenarios.
Read the report
REPORT 03 · EXECUTED 2026-08-24
Can a citation-first RAG service clear one million requests at low latency?
One million successful requests through a public GCP edge, 13.45 ms p95, zero HTTP errors, forced database traffic and an explicit production cost boundary.
Read the report
REPORT 04 · EXECUTED 2026-08-28
When a tool fails, does the agent recover, abstain, or invent?
Six injected failure modes, three models, 219 runs. Nothing was fabricated. Two of the three models never retried a tool that explicitly asked to be retried.
Read the report
Publication standard
Measured means reproducible
Failures included
We publish false negatives, failed configurations and limitations—not only the metric that makes a gate look useful.
Reproduction included
HTML is ungated. Raw results, methodology and executable reproduction files are downloadable without an email.
Release queue
The next evidence assets
04AIoT anomaly and offline-recovery benchmark — planned with NASA C-MAPSS
05State of AI Project Readiness — only after 100 estimator submissions and 20 responses per segment