HMIC / Articles
Long-form writing, papers and benchmark reports. Each one says what was found and where the claim stops, and links the data.
Article / November 2026 / Model evaluation
A model returned the right temperature. My grader read the unit ID instead. The audit traces that bug, misleading failure labels, and what the frozen results can actually support.
TJ Jung · November 10, 2026
Article / November 2026 / Tokenmaxxing
One person, one subscription, four months of ChatGPT Codex. How do you max out a flat-rate AI plan, how do you measure that you really did, and how do you use it to build things that outlast the subscription?
TJ Jung · November 3, 2026
Canonical definition / v2.0 / August 2026
When two parties share an engineered, evolving codebook, messages shrink toward near-nothing because the shared state carries the meaning. The definition, a five-part recognition kit, a boundary page on what PIC is not, and worked examples.
Whitepaper v1.0 and software v1.0.0 / August 2026
Audit evidence that someone outside the system can verify without seeing the private record. A reference architecture, plus ALCOA RubberStamp, free MIT-licensed Python that implements the commitment and verification step.
Benchmark 01 / RTX 5090 / September 2026
35B-A3B decoded 3.82× faster; 27B held the stronger cold result on an unseen holdout, 20/24 to 12/24. Speed and cold reliability picked different winners.
Benchmark 02 / Now running
Where does Q6 buy measurable task capability over Q4_K_M, and where does that gain justify its memory and speed cost? One pinned source across an RTX 5090 and RTX 3090s.
Every benchmark, with its question, finding, boundary and hashes, plus the anchored public evidence: all benchmarks · evidence lineage.
The benchmarks, narrated. The site keeps the tables, methods and limits; the channel carries the analysis.
Open the channelHMIC LLC · independent research · measured, bounded, correctable
