site.py now lists a SKILL.md under skills/ only when /<name>.md is that file byte for byte and every file it points at by a relative path (references/, scripts/, templates/, tools/, a sibling page) is published. 120 entries pointed at files the site does not serve, reference docs were listed as skills, and 18 names served another file. 188 skills remain. tests/discovery.py holds the published index to the same rule. build.py deletes every page it did not write, except the hand-written pages in KEEP, walks skills/ in name order so a shared name resolves the same on every filesystem, and refuses to run without HIPS_DIR. Rebuilt against hanzoai/hips main: the 40 discover-* pages and 20 HIPs whose source is gone are removed, the stale code-review, engineering-code-review and ml pages match their source, and 66 pages that were never built (60 skills among them) and 4 new HIPs are published. .gitignore anchors the root symlinks, so public/gemini.md is not ignored on a case-insensitive checkout.
2.3 KiB
Hanzo ML Research: Core Stack Reference
This reference accompanies the hanzo-ml-research skill.
Canonical repositories
hanzoai/ml
Rust-native ML compute stack.
Important crates include:
hanzo-mlhanzo-nnhanzo-trainhanzo-transformershanzo-kernelhanzo-research
The hanzo-research crate is the canonical Rust client for the unified cloud research plane and automatically records source/runtime provenance.
hanzoai/engine
Native Rust inference/serving engine.
Use it when research touches:
- serving
- batching
- quantized inference
- model loading
- attention kernels
- KV cache
- speculative decoding
- backend-specific runtime behavior
- end-to-end latency
hanzoai/decision
Kai typed-decision research and production stack.
Use it when research touches:
- typed decisions
- decision programs
- calibration
- safety heads
- option retrieval/reranking
- joint decoding
- benchmarking
- training objectives
- StateHandle / cached decision execution
Hanzo DSL / kernel stack
Use the local Hanzo DSL/kernel implementation exposed by the active workspace. Prefer extending existing Rust DSL/kernel primitives to introducing a second implementation path.
Research API
The native Rust SDK in hanzo-research uses:
POST /v1/research/experimentsGET /v1/research/experimentsGET /v1/research/totalsPOST /v1/research/artifacts
Configuration:
HANZO_API_KEYRESEARCH_BASERESEARCH_PROJECT
The server keys canonical experiment identity by the research record's project/stable experiment identity, while provenance captures git/repository/runtime details.
Typical research object
hypothesis
prediction
control
treatment
primary metric
non-inferiority metrics
train/dev/final lineage
artifact hashes
raw observations
verdict
because
Keep the experiment falsifiable and durable.
Rust-first rule
Python/notebooks are acceptable for disposable visualization or inspection.
They are not canonical for:
- final model training
- benchmark generation
- feature extraction
- serving latency
- kernel timing
- final threshold selection
- production model behavior
unless the experiment explicitly studies a Python/PyTorch implementation.
Canonical results should be reproduced in Hanzo's Rust stack.