HMIC / Studies
The science, in one place.
Benchmarks, definitions, whitepapers, reference software and anchored evidence. Each study states the question it asked, what was found, and where the claim stops. The report, the data and the downloads are linked from the card.
01
Model-chain benchmarks
Benchmark 01 / RTX 5090 / September 2026
Qwen3.8-27B versus Qwen3.6-35B-A3B
- Question
- Of two installed Qwen deployments on one RTX 5090, which should an operator start with, and which should they build around?
- Found
- 35B-A3B decoded 3.82× faster (205.3 versus 53.8 tokens per second over 48 timed calls). 27B held the stronger cold result on an unseen holdout, 20/24 to 12/24. One compact operating card moved 35B to 16/24 and restored the whole causal-transfer family.
- Boundary
- A deployment comparison between two source packages at different quant levels (27B Q6_K, 35B-A3B Q4_K_M). It does not isolate architecture from quantization or packaging. Every arm missed the integer-only chain; that failure calls for a deterministic executor, not more prompting.
Evidence pack SHA-256
c8e1a1ca602fa9f80063610f95ed395ad4820aeb74956d32918ec0aa67b24283
Model file hashes, per-call outputs and the calibration card are inside the pack and listed in the report.
Benchmark 02 / Now running
Q4_K_M versus Q6_K from one pinned source
- Question
- Where does Q6 buy measurable task capability over Q4_K_M, and where does that gain justify its memory and speed cost?
- Design
- One pinned model source. RTX 5090, one RTX 3090, and two RTX 3090s. Identical context, runtime settings, prompts and scoring. Measured: speed, time to first token, memory, semantic accuracy and exact format.
- So far
- In an earlier installed comparison of Qwen3.6-35B, Q4 ran nearly three times faster while retaining about 98% of the measured quality. The matched-source matrix above is what turns that observation into a result, and removes the ambiguity left by Benchmark 01.
02
Definitions and papers
Canonical definition / Version 2.0 / August 2026
Programmed Interface Compression
- Question
- When two parties share an engineered, evolving codebook, how small can their messages become before meaning is lost, and how do you recognize that it is happening?
- Found
- A canonical definition: messages shrink toward near-nothing because the shared state carries the meaning. A five-part recognition kit (shared reference, reduced repetition, correct reconstruction, evolving state, repair path), a boundary page on what PIC is not, and worked examples.
- Boundary
- This names a pattern and publishes explanatory material. It does not claim exclusive rights, measured performance, customer adoption or industry priority.
Whitepaper v1.0 and reference software v1.0.0 / August 2026
ALCOA+ by Public Consensus, with ALCOA RubberStamp
- Question
- Can audit evidence be verified by someone outside the system that produced it, without exposing the private record?
- Found
- A reference architecture for externally verifiable audit evidence: private payloads, deterministic commitments, and independently inspectable witness layers. ALCOA RubberStamp v1.0.0 implements the file-set commitment and verification primitive as free, MIT-licensed, dependency-light Python, with optional prior-state binding and OpenTimestamps submission.
- Boundary
- Reference software and a reference architecture. Not a validated GMP system, not a 21 CFR Part 11 or EU GMP Annex 11 compliance product, not an electronic-signature system, and not proof that a record is true. Output must be governed according to intended use.
03
Anchored evidence
Evidence lineage / 2026-04-18 to 2026-08-03
Six confirmed public anchors
- Question
- Can the anchoring method be demonstrated on a public network without making that network part of the architecture?
- Found
- Six confirmed public anchors, from genesis to the current lineage tip, each dated and linked to the full transaction on a public explorer.
- Boundary
- The protocol is what HMIC publishes; the network is selected per deployment, and HMIC does not operate, validate or maintain any specific network.
04
Follow, or send a failure
The Assembled Worker
This site carries the tables, methods, limits and durable corrections. The YouTube channel carries the narrated analysis and the finished visual evidence.
Open the channel
Bring a real failure.
If you run one of these models, send the exact configuration and the first real task where it breaks. Useful requests become candidate cells in the next benchmark.
What to include