Weight programming for humans and agents: ternary quantization, ultra-graphs, reversible weight diffs, and an academy that verifies real checkpoints.
  • Rust 62.7%
  • HTML 24.1%
  • JavaScript 6.1%
  • Shell 4.7%
  • CSS 2.4%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-11-17 15:48:45 -05:00
.github ci: build one universal macOS binary instead of two racing jobs 2026-11-17 15:48:45 -05:00
agent-plugin feat: safetensors reader, ultra-graph submodule, agent plugin 2026-11-17 15:17:55 -05:00
crates fix: clippy manual_clear, actionlint install path, 9 mcp tools 2026-11-17 15:28:33 -05:00
docs docs: record ci evidence, enable badges 2026-11-17 15:33:57 -05:00
examples feat: safetensors reader, ultra-graph submodule, agent plugin 2026-11-17 15:17:55 -05:00
scripts ci: install cross target std for aarch64, zip via Compress-Archive on windows 2026-11-17 15:38:30 -05:00
site feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
third_party feat: safetensors reader, ultra-graph submodule, agent plugin 2026-11-17 15:17:55 -05:00
.gitattributes feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
.gitignore feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
.gitmodules feat: safetensors reader, ultra-graph submodule, agent plugin 2026-11-17 15:17:55 -05:00
AGENTS.md feat: safetensors reader, ultra-graph submodule, agent plugin 2026-11-17 15:17:55 -05:00
Cargo.lock feat: safetensors reader, ultra-graph submodule, agent plugin 2026-11-17 15:17:55 -05:00
Cargo.toml build: canonical codeberg remote, github mirrors, sync-owners parity 2026-11-17 15:25:07 -05:00
CLA.md feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
CODE_OF_CONDUCT.md feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
CONTRIBUTING.md feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
GOVERNANCE.md feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
LICENSE-APACHE feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
LICENSE-MIT feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
llms.txt feat: safetensors reader, ultra-graph submodule, agent plugin 2026-11-17 15:17:55 -05:00
OWNERS.toml build: canonical codeberg remote, github mirrors, sync-owners parity 2026-11-17 15:25:07 -05:00
README.md docs: record ci evidence, enable badges 2026-11-17 15:33:57 -05:00
SECURITY.md feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00
TRADEMARK.md feat: governance, site, docs, ci 2026-11-17 15:04:42 -05:00

Omnispherepop — a mutual product of Peter Lodri Sec · 88B (8b-is) · Standard Galactic

Omnispherepop

ci pages

Weight programming for humans and agents. Omnispherepop treats neural weights as a program instead of an opaque blob: ternary quantization, ultra-graphs, reversible diffs, and an academy that teaches by making you move real weights.

tensor layer.weight (ternary): 512 values in 4 groups
trits: 128 bytes vs 2048 float bytes (93.8% smaller, 95.0% with 5-trit packing)
sparsity 48.4%  mse 0.3496  scale range [1.5010, 1.8202]

No account, no telemetry, no model download. The demo snapshot is generated deterministically, so the terminal, the browser, and CI all see the same numbers.

30-second tour

cargo install --git https://github.com/omnispherepop/omnispherepop --bin ospop

ospop inspect                                  # snapshot identity and tensors
ospop quantize --group-size 128                # ternary PTQ, real size and error numbers
ospop lens --view ternary --eli5               # the same tensor through the ternary lens
ospop learn wp.ternary                         # a lesson with an observed checkpoint

ospop --file examples/tiny-gpt2.safetensors inspect
ospop --file examples/tiny-gpt2.safetensors --json quantize \
  --tensor transformer.h.0.attn.c_attn.weight --group-size 32

Weight programming, concretely

Operation What it does Evidence
quantize Δ-threshold + per-group scale to {-1, 0, +1} sparsity, packed bytes, RMSE
lens one tensor through ternary, heatmap, graph, diff, histogram views rendered output, round-trip flag
diff reversible edit set bound to a snapshot identity round_trip_ok: true
prune keep the subgraph reachable from a root ultra-edge kept/dropped counts, restore check
learn 17 lessons in 5 tracks, each with a checkpoint per-lesson verdict from the real runtime
load --file read a safetensors checkpoint (dense F32/F16/BF16) tensor count, snapshot id, strict rejection of malformed files

Two calibrations ship. --scheme ternary thresholds each group at Δ = 0.7·mean|w| and keeps a per-group scale. --scheme qdelta applies the ultra-graph A6P rule in Q16.16, which needs no look at the data. Both report the same fields, so you can compare them on one tensor.

Every surface is one Program enum plus Runtime::execute. The CLI, the MCP server, and the browser kernel are adapters over the same operations, and scripts/check-parity.mjs fails the build if the browser drifts from Rust.

Agents

ospop-mcp speaks MCP over stdio with eight tools, so an assistant can drive the same operations you type:

printf '%s\n' \
 '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
 '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
 '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"ospop_quantize","arguments":{"tensor":"layer.weight"}}}' \
 | ospop-mcp

Tool names: ospop_load_demo, ospop_load_safetensors, ospop_inspect, ospop_quantize, ospop_lens, ospop_prune, ospop_diff, ospop_explain, ospop_learn.

Academy

Seventeen lessons across Foundations, Deep Learning, LLMs, Weight Programming, and AI Systems. Each lesson binds a lab to a real tensor slice and verifies an observed checkpoint — sparsity, reconstruction error, patch round-trip, or graph reachability — not a multiple-choice quiz. ospop learn lists them; ospop learn <id> runs one.

Site

site/ is static HTML, CSS, and vanilla JS: no bundler, no framework. It hosts the landing page, the live ternary sandbox, and one page per lesson, generated from the CLI by scripts/build-site.sh so the site cannot claim behavior the code lacks.

bash scripts/build-site.sh
python3 -m http.server 8099 --directory site   # then open http://localhost:8099/

Ultra-graph

third_party/ultra-graph is a submodule. Its zerocpu crate supplies two things this project uses on the live path: 5-trit packing (1.6 bits per value, denser than the core's 4-per-byte codec) and the A6P Q16.16 quantizer behind --scheme qdelta. ospop_core::ultragraph is the bridge, and tests compare its integer dot product against naive arithmetic so a kernel mismatch fails loudly.

Real checkpoint

examples/tiny-gpt2.safetensors is a Hugging Face safetensors file: a 2-layer GPT-2 with 64 tensors, about 454 KB. crates/ospop-core/tests/real_checkpoint.rs loads it, quantizes a projection, and asserts the measured numbers. The weights are randomly initialized, so these are real measurements of a real file — not perplexity, and not a quality claim about a trained model. There is no trained checkpoint here yet.

Coding agents

  • AGENTS.md — the rules an agent must follow in this repository.
  • llms.txt — a compact map of what exists and what does not.
  • docs/adr/ — why one operation set, and why ultra-graph is a submodule.
  • agent-plugin/ — plugin manifest, MCP wiring, and three skills: omnispherepop (weight operations), superpowers-bridge (process mapped onto this repo's gates), caveman-comms (compressed reporting that keeps numbers exact).
cargo build --release -p ospop-mcp -p ospop-cli
bash scripts/install-agent-plugin.sh          # skills into ~/.agents/skills

Repository layout

crates/ospop-core       trits, PTQ, ultra-graphs, snapshots
crates/ospop-runtime    Program enum and Runtime::execute
crates/ospop-academy    curriculum, lessons, checkpoint verification
crates/ospop-cli        binary `ospop`
crates/ospop-mcp        binary `ospop-mcp`, MCP over stdio
third_party/ultra-graph git submodule: zerocpu kernels (5-trit packing, A6P)
agent-plugin/           plugin manifest, MCP wiring, agent skills
examples/               real safetensors checkpoint used by tests
site/                   landing, sandbox, academy pages
scripts/                site build, owner drift check, kernel parity check
docs/                   architecture, roadmap, release status

Publishing

Three homes, because ownership is joint. Codeberg is canonical (codeberg.org/8b-is/omnispherepop); GitHub carries two mirrors for reach, one per owner account.

Host Repository Role
Codeberg https://codeberg.org/8b-is/omnispherepop canonical, receives tags first
GitHub https://github.com/8b-is/omnispherepop mirror
GitHub https://github.com/peterlodri-sec/omnispherepop mirror
bash scripts/setup-remotes.sh --with-codeberg   # wire origin (canonical), 8b-is, peter
git push -u origin main --tags                  # canonical first
git push 8b-is main --tags
git push peter main --tags

scripts/sync-owners.sh --check fails if a URL drifts or a mirror is dropped from this table or from OWNERS.toml.

Governance

Jointly owned by Peter Lodri Sec, 88B (8b-is), and Standard Galactic. See GOVERNANCE.md, TRADEMARK.md, and CLA.md. OWNERS.toml is the single source of identity, enforced by scripts/sync-owners.sh --check.

Status

Read docs/STATUS.md before trusting a claim in this README. Version numbers in git tags are release claims; the status page records which gates actually ran, on what, and which are still open.

Contributing

See CONTRIBUTING.md. Security reports: SECURITY.md. Code of conduct: CODE_OF_CONDUCT.md.

License

MIT OR Apache-2.0.

Omnispherepop — a mutual product of
Peter Lodri Sec · 88B (8b-is) · Standard Galactic