Downloads · MIT · Apache 2.0 · CC-BY-SA 4.0
Take what we built. Use it.
Save yourself a month of trial-and-error.
The native C++/CUDA runtime, the surgical pipeline, the PLANCK pack format, the Physarum surgery engine, the Black-Dog reinforcement loop, the NanoOS capsule substrate, the hologram cache, the doctrine, and ninety-five reports. Open under MIT, Apache 2.0, and CC-BY-SA 4.0. No registration, no telemetry — on the artefacts that ship now.
The 95-report archive on this page is the full corpus including doctrine docs, internal audits and pre-publication drafts. The 66-entry /r/ index is the public-rendered subset — each one its own page with markdown rendering, schema.org TechArticle markup, and raw .md source link.
Honest release tiers
- Available now — direct download link visible on the card. Click and go.
- Release pending — artefact is real and measured in our internal reports, but a public-grade build (cleaned, signed, dependency-pinned) is still being prepared. ETA where known.
- Partner access — given on request via email; we want to talk to who is running our weights so we can fix bugs faster while artefacts are pre-release.
- Restricted — items not yet public for reasons we name on the card (license, third-party weights, in-flight surgery).
We over-disclose. Each card states its tier explicitly. If you spot an artefact marked "pending" for more than 30 days with no update, email [email protected] and we will either ship it or move it to "Restricted" with a reason.
What you get here that you cannot get anywhere else.
Five things — every one of them measured, every one published with reverts and errata visible. If any one of them saves you a month of trial-and-error, that is a fair trade.
A real, measured surgical pipeline on open weights.
Not a tutorial that ends at "fine-tune your model". The full cycle: bench failure → poison dataset → QLoRA → merge → repack → strict 4-axis gate → flip pack → re-bench → keep or revert. Eight rejected passes recorded.
A native C++/CUDA LLM runtime that doesn't import Python at runtime.
7B Q4 at 11+ tok/s on a 3060 Ti, 83.58 tok/s production via clean-room kernel autopsy. ChatML wrapper. Identity-anchored against donor token leakage. Single binary, no daemon.
The surgical organism with conductance routing and a 860× cache.
Multi-organ orchestration, critic+wound repair before 7B fallback, hologram cache for identical-prompt repeats (860× speedup), DAG-recorded food/poison reinforcement signal.
A 284 B MoE flagship demo on a single 8 GB GPU.
DeepSeek V4-Flash driven through end-to-end inference on a 3060 Ti — full optimization trace from naive baseline to working decode, including the 380-vs-100-second negative result we kept on the record.
Honest documentation. Public losses too.
The TRUTH_LEDGER pattern, errata-first culture, reverts visible. Public reports of our LOSSES (Sovereign Gauntlet RED 59/60, BD8 0-rescue, BD6.x ceiling at 53 % anchor saturation), not just wins.
Licensing
Open. Attributed.
We require attribution and ask that derivative reports keep the "no GREEN without numbers" doctrine: cite measured artefacts, record reverts as negative results, never hide errata. Donor weights (Qwen 2.5 0.5B / 7B Instruct) are not redistributed — pull them from HuggingFace under their Apache 2.0; we ship the deltas.
Catalogue
01 · Surgical artefacts
Production .planck packs.
The actual deployed weights. Each pack carries its baseline number, post-surgery delta, anchor set, and frozen pack hash — no merge without those four artefacts. Direct download for small artefacts; large packs are released via GitHub releases due to bandwidth budget.
physarum05b_code_skeleton.planck
physarum05b_triz_contradiction_v2.planck
physarium7b_identity.q4planck
Physarum-05B-Organic / model.safetensors
D < mean·0.1. Killed 20.6 % of weights; PPL +15.3 %; MMLU-mini −22 %; GSM8K-mini −20 %; JSON-smoke 100→100 %; same disk size; throughput preserved. Demonstrates the doctrine: gate, measure, record, no hide.physarum_engine.cpp + .exe
block_size=256, n_iter=30, β=2.0, energy-conserving softmax, threshold D < mean·0.1. This is the seed code that grew into everything.02 · Native runtime
gigachad_native — single C++/CUDA binary.
The compiled inference loop. CUDA 12. Linux + WSL2. No torch, no onnx, no huggingface in the running process. Nine CUDA kernels plus dequant-fused Q4 GEMV, DP4A v3 inner loop, fused residual+rmsnorm, fused silu_mul + down GEMV, tile-K shared-mem staged swiglu+down, CUDA graphs and fusion (Phase 8E5).
gigachad_native (build/)
03 · PLANCK pack format
mmap-able weights. Byte-verifiable.
Specification + writer + reader + verifier. Byte-for-byte verify 50 / 50 PASS. .planck file is mmap'd by the runtime; the writer never touches Python at decode time. Q4 group=128 with delayed scale, FP8 e8m0 → FP32 bit-shift expansion.
PLANCK7B_PACK v1 — spec / writer / reader / verifier
docs/PLANCK_PACK_v1.md. Writer: build/planck7b_tool. Reader: src/runtime/planck_runner.cpp. Verifier: tools/planck/planck_pack_verify.cpp.04 · Surgery toolkit
Disposable Python training capsule.
Python lives in tools/surgery/ only. Every tool produces .planck / .jsonl / .md artefacts; the runtime never imports them. The full failure-driven QLoRA cycle with 4-axis gating (anchor / strict-schema / target-bench / no organ leak).
tools/surgery/* — 9 production tools
05 · Bench harnesses
The tests that gate every surgery.
Public benchmarks run under our doctrine: 3-mode A/B/C (7B-only / organ-only / organ+fallback), repeat-learning ×N rounds, parallel-retry k=1..3, sovereign cognition gauntlet. Every run produces a JSON snapshot; no result merges without one.
tools/bench/* — 7 harnesses
06 · Memory spine + hologram cache
Line-addressable spine. 860× cache.
The persistent memory layer + the exact-match output cache. Both are built and shipped today.
tools/memory/build_spine_index.py
data/memory_spine/manifest_v1.jsonl · 305 files / 58 996 lines, sha256[:16] per line.src/runtime/hologram_cache.cpp
sha256(input). 860 ms → 1 ms = 860× speedup on identical-prompt repeats. Clean and small — about 150 lines of C++.holo_log_pack — lossless spine compression skeleton
include/holo_log_pack.hpp + src/memory/holo_log_pack.cpp + tools/memory/holo_log_smoke.cpp. Smoke 5 entries (46–65 535 bytes), sha-anchored.07 · NanoOS capsule substrate
Proof-carrying execution. Replay on a stranger's machine.
Sandboxed shell environment plus a code-repair / parallel-retry loop in the native runtime. Every run produces dag/capsules/cap_*.json with replay-recipe — you can take any answer the model produced, hand it to someone else, and they will get the same stdout / stderr / exit code.
tools/capsule/shell_capsule.py
replay_recipe spec.src/runtime/code_repair.cpp + parallel_retry.cpp
cat > file <<'EOF' ... EOF, python3 - <<'PY' ... PY, trailing-\ continuations).08 · ARIZ / TRIZ reasoning kernel
Contradiction analysis as a separate organ.
Rule-based ARIZ stages 1 / 4 / 5 / 6: neutralizer, IFR (ideal final result), resources, TRIZ operators. Every DAG entry stamps an ariz_trace_id; full ariz_trace_json is saved to reports/ariz_traces/.
src/reasoning/ariz_trace.cpp + ARIZ_KERNEL.md
09 · Black-Dog learning loop
Conductance per route. Float, persistent, queryable.
Per-(pattern_hash, action_chain_hash) conductance store. EMA: c = (1-α)·c + α·(food − poison), clamped [-1, 1]. Persisted to physarium/route_conductance.json. The router queries the store before launching any organ.
src/runtime/black_dog.cpp + BLACK_DOG_LEARNING_LOOP.md
10 · Doctrine documents
The doctrine itself.
Twenty-four documents covering the chronology, the truth ledger, the architecture lock, the clean-room doctrine, the ARIZ kernel, the Black-Dog learning loop, the NanoOS execution substrate, and the organ-colony spec. Released together as a single zip.
doctrine-pack.zip · 24 documents
11 · Reports archive · 95 case-studies
The school.
Every surgery and bench writes its own report. The 95-file archive is a reference for how we work — the BD6 ledger of nine surgery passes (one keep, eight reverts), the BD7 TRIZ surgery ascent from 0 to 88/100 in seven stages, the autopsy reports of llama.cpp / ExLlamaV2 / AWQ-Marlin (kernel-level extraction, not framework dependency), and the public LOSSES (Sovereign Cognition Gauntlet RED 59/60).
reports-pack.tar.gz · all 95 reports
12 · Datasets we built
Surgical training data, shipped open.
Every dataset that drove a production surgery, plus the line-addressable spine manifest, plus the replay capsules from Terminal-NanoOS-30. CC-BY-SA 4.0.
data/organ_surgery/phys05_code_skeleton/poison_train.jsonl
data/organ_surgery/phys05_triz_contradiction/
ariz_tasks_v1.jsonl · 100 hand-curated ARIZ contradiction tasks. teacher_targets_v3_100.jsonl · 100 strict-validated 6-field JSON targets. triz_train_80.jsonl · 80 train / 20 eval / 10 anchor splits.data/organ_surgery/phys05_critic_lite/critic_dataset_v2.jsonl
data/memory_spine/manifest_v1.jsonl
tools/memory/build_spine_index.py.dag/capsules/cap_*.json — Terminal-NanoOS-30 replay capsules
replay_recipe.spec_inline — runs anywhere with the same stdout / stderr / exit code.Integrity
sha256 manifest.
Every direct-download artefact below has its sha256 listed in MANIFEST.sha256.txt. Verify after download.
MANIFEST.sha256.txt
Need something not yet listed?
Production weight packs (.planck), the gigachad_native binary, the surgery toolkit and the bench harnesses are gated through email while the GitHub release is being prepared. Send a one-line request — what you want and what you intend to do with it — and we will reply with a direct download link or a build.