A
ADAM · Adaptive Dual-torus Autonomous Mind
Aliases: ADAM engine · Program 05
A C++17 sovereign cognitive engine — a single binary (~45 000+ lines of C++17 in total; semantic.cpp ~13 k lines holds scoring + HTTP, the rest sits across legion.h, merkaba_heart.hpp, vortex_cuda.cu and others) that runs a 1.2-million-concept Legion graph, Clifford algebra Cl(3,0) + Cl(4,1), dual-torus MerKaBa dynamics, and biological Physarum routing for inference. Not a wrapper around a language model. ADAM accumulates a persistent binary graph of concepts, learns continuously from interaction, and produces answers grounded in its own internal geometry rather than in token statistics.
ADAM is the long-horizon program at the centre of the CyberdyneLabs stack. As of 2026-05-18, ADAM holds #1 on the public ARC-AGI-3 leaderboard: 25/183 (13.66 %) autonomous and 183/183 (100 %) hybrid harness — see /arc-agi-3. The runtime answers in ~500 ms end-to-end (raw graph traversal: ~80 ms). CPU mode runs the full substrate; CUDA acceleration is optional and unlocks 1024-clone parallel hypothesis evaluation. Live chat at /adam-chat (currently undergoing scheduled upgrade — v2 launches 2026-05-21); full architecture page at /adam.
Source: /adam · serialised state in adam_memory.bin · official scorecard at arcprize.org/scorecards/6a5888ac-21e1-40b9-abac-5fecbe62cb42.
Anchor gate
One of the four axes of the surgery 4-axis gate. A pre-defined set of probe questions (default 19) that a new .planck pack must answer correctly before it is flipped into production. If even one anchor fails, the surgery pass is reverted regardless of its bench numbers.
BD6.x surgery showed the anchor gate has a hard ceiling: anchor replication saturates at ~53 % on the Physarum-05B donor, which is why eight surgery passes were reverted before BD6 pass-1 was kept.
See BD6_8D_RANK_FINAL_FREEZE.md in the downloads pack.
ARC-AGI-3 · ARC Prize 2026 benchmark
Third generation of the Abstraction and Reasoning Corpus benchmark from ARC Prize. 25 interactive game-like environments, 183 levels total. Successor to the static ARC-AGI-2; solvers receive a stream of frames, take discrete actions (click · keyboard · keyboard_click), and the environment responds. Offline, no internet at eval time, no game source code.
As of 2026-05-18, ADAM holds #1 on the public ARC-AGI-3 leaderboard on both tracks: 25/183 (13.66 %) autonomous and 183/183 (100 %) hybrid harness. The 22 → 24 → 25 climb across 9 days is closed-loop substrate learning (see ARC closed loop, warmed world_model).
Live leaderboard: /arc-agi-3 · official scorecard: arcprize.org/scorecards/6a5888ac-21e1-40b9-abac-5fecbe62cb42.
ARC closed loop · substrate-learning loop
The four-endpoint cycle through which ADAM plays ARC-AGI-3 and improves across runs: FRAME → ADAM endpoints (/game_search_init · /game_search_expand · /game_search_next · /game_procedure_learn) → substrate scoring (Cl(4,1) convolution + MerKaBa + HDC) → ACTION → world_model update → procedure memory → next run starts warmed.
The signal is not the absolute percentage on any given run — it is the loop closing: experience changes memory, memory changes procedure selection, procedure selection improves future runs. This is what "cognition lives in the substrate" means in practice.
Visual diagram on /adam ARC section.
ARIZ kernel
Aliases: ARIZ trace · TRIZ-40 kernel
A rule-based contradiction-resolution kernel built on the ARIZ algorithm (Алгоритм Решения Изобретательских Задач). Stages 1, 4, 5, 6 are wired in src/reasoning/ariz_trace.cpp: a neutralizer, an Ideal Final Result (IFR) pass, a resources sweep, and the 40 TRIZ inventive-problem operators.
Every DAG entry that goes through the kernel stamps an ariz_trace_id; the full ariz_trace_json is saved to reports/ariz_traces/ for audit. ARIZ surgery is a separate organ — see BD-series BD7 (TRIZ contradiction organ, 88/100 strict 6-field JSON).
Spec: ARIZ_KERNEL.md.
Asabiyyah cycle
Aliases: Ibn Khaldun cycle · five-phase civilizational cycle
A 14th-century model of civilizational rise and decay described by Ibn Khaldun: solidarity (asabiyyah) accumulates in early-state populations, drives expansion, then attenuates as luxury and prestige accumulate, until the polity collapses and a new lineage rises. CyberdyneLabs Hypercolony measured this cycle from local agent behaviour on a 4D substrate without any scripted phase transitions.
The cycle's five phases — Rise · Zenith · Luxury · Decline · Collapse — emerge from per-clan accumulation of cohesion, luxury, and prestige; nothing in code tells a clan to fall.
Live in the /hypercolony-app simulator.
B
BD-series surgery · BD = Black-Dog
A numbered sequence of QLoRA surgery passes on the 0.5B specialist organs of Frankenstellm. Each BD-pass is gated by the 4-axis gate; passes that fail are reverted, passes that pass are merged into the production .planck pack.
Trajectory through 2026-05-05: BD6 code-skeleton organ kept after eight reverted passes (MBPP B 13/100, HE B 6/164). BD7 TRIZ contradiction organ at 88/100 strict JSON. BD8 critic-lite + wound rescue surgery — ARIZ rescue rate 0/n; wound v2 retained for in-chat path. BD9 swept four organs in one session — json_repair (10/10 GREEN), claim_extractor (GREEN), test_writer and cache_matcher (YELLOW), renderer (RED, queued BD9.1). Production state: 5 of 8 organs surgered.
All BD reports in reports-pack.tar.gz · history page /history#era3.
Black-Dog learning loop
Aliases: BD loop · conductance store · food/poison reinforcement
The reinforcement memory of the Frankenstellm runtime. Per (pattern_hash, action_chain_hash) tuple it stores a conductance value updated by exponential moving average: c = (1-α)·c + α·(food − poison), clamped to [-1, 1]. Persisted to physarium/route_conductance.json.
Verifier-pass writes food=1; verifier-fail writes poison=1. The router queries the store before launching any organ; selects the chain with maximum conductance for the (route, organ_chain) pair.
Spec: BLACK_DOG_LEARNING_LOOP.md.
C
CGA address · Cl(4,1) Conformal Geometric Algebra
The address scheme of PhysarumChain. Standard derivation: first 20 bytes of SHA-256(Ed25519 public key). In addition, each address maps to a null vector in Cl(4,1) — a 32-component multivector encoding position and orientation in 5-dimensional projective conformal space.
Geometric proximity in Cl(4,1) corresponds to routing distance: nodes close in CGA space relay between each other at lower latency. The address is no longer just an identifier — it is a coordinate the routing layer can read.
See /physarum Feature 02.
Clean-room doctrine
A CyberdyneLabs principle: external systems (llama.cpp, ExLlamaV2, AWQ-Marlin, Claude Code) are autopsy specimens, never runtime spine. We extract kernel-level techniques (e.g. DP4A inner loop, fused dequant GEMV) and reimplement them inside gigachad_native; we do not import the external library as a dependency.
The doctrine prevents the "framework drift" failure mode where a research project becomes a thin wrapper around someone else's binary.
Doc: CLEAN_ROOM_DOCTRINE.md · the doctrine pack.
Clifford algebra · Cl(p,q) geometric algebra
An algebra over a vector space equipped with a quadratic form, where vectors square to scalars and the geometric product unifies dot product and wedge product. The CyberdyneLabs stack uses Cl(3,0) for ADAM concept geometry (8-dimensional multivectors representing scalars · vectors · bivectors · trivectors — equivalent to 3D rotations and reflections without the singularities of Euler angles), and Cl(4,1) Conformal Geometric Algebra for PhysarumChain addresses.
Concepts in ADAM are not floating-point vectors. They are geometric objects with spin, phase, and orientation. Semantic operations are geometric products — rotation, reflection, projection — not dot products.
Used in ADAM L2 and PhysarumChain Feature 02.
D
DP4A · dot-product-of-4-byte-accumulator
A CUDA hardware instruction (__dp4a) on Pascal-and-later GPUs that computes the dot product of four packed int8 values into a single int32 in one cycle. Used by the gigachad_native backend to accelerate Q4 GEMV: with the Q4_GEMV_DP4A=1 opt-in flag, Physarium-7B Q4 decode rises from 18.27 to 28.99 tok/s (+59 %) on RTX 3060 Ti, and to 41.69 tok/s with tg128.
Phase 8E.8a · PHASE_8E8A_DP4A_NATIVE_BACKEND.md.
E
Expert streaming
The choreography that makes Mixture-of-Experts inference possible on a small GPU. For DeepSeek-V4-Flash, only ~1.6 GB of resident weights live in VRAM (the Singularity Monolith); the routed experts live on disk and are streamed in on demand per layer.
Expert streaming was the breakthrough that turned 284 B-total / 13 B-active inference into a problem of disk I/O scheduling rather than VRAM capacity. Roofline measurement: 89 % wall-time = expert I/O; 60 % of that = pure disk wait at 48 ms/expert. The PLANCK pack format and the hot-expert cache both descend from this insight.
V4_FLASH_TECH_BRIEF.md.
F
Four-axis gate
The merge criterion every CyberdyneLabs surgery pass must satisfy simultaneously:
1. Anchor 19/19 — pre-defined probe questions still answered correctly.
2. Strict-schema — output matches the production verifier's regex / JSON schema / compile gate.
3. Target-bench — no regression vs the previously kept pack on the relevant bench.
4. No organ leak — organs_used set equals expected; no unexpected fallbacks.
Failure on any axis triggers a revert to the previously frozen pack. This is the gate that produced eight reverted BD6 passes before BD6 pass-1 was kept. The doctrine: "no GREEN without numbers; reverts recorded in full".
See /surgery#cycle.
G
gigachad_native
The single C++/CUDA binary that runs the Frankenstellm organism. CUDA 12. Linux + WSL2. No torch, no onnx, no huggingface in the running process. Nine CUDA kernels (gemv · rmsnorm · rope · kv_cache · gqa · silu_mul · residual · embed · argmax) 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).
Production speed on RTX 3060 Ti: 83.58 tok/s via the llama.cpp clean-room backend; 18.27 tok/s native default; 28.99 tok/s with DP4A flag.
Acceptance suite 18/18 · identity probe 14/14 · architecture audit 10/10 · gigachad_acceptance_run_v14_llamacpp.json.
Grain · smallest unit of MKB
The atomic monetary unit of PhysarumChain. 1 MKB = 100 000 000 grains (8 decimal places). The minimum transaction fee is 0.0000001 MKB = 10 grains, which is also the anti-spam floor. A typical contract call uses ~10× the floor (~0.000001 MKB); a heavy multi-action call ~100× (~0.00001 MKB). All on-chain math runs in 128-bit integers — no floating-point rounding.
Economics block at /physarum#fees.
H
HDC vector · hyperdimensional computing
A 1024-bit dense binary vector representing a concept in the Legion graph. HDC vectors compose via element-wise XOR (binding) and majority vote (bundling), giving an algebra in which "the capital of France" can be derived from "Paris" + "France" + "capital_of" without any neural network call.
ADAM uses 1024-bit HDC for every concept; bonds between concepts carry their own 1024-bit binding vector, which is how a graph traversal can synthesise an answer without a generation step.
See /adam L3 Legion graph.
Hologram cache
An exact-match output cache for the Frankenstellm runtime, keyed on sha256(input). On a cache miss the runtime answers normally; on a hit it returns the prior answer in ~1 ms instead of ~860 ms = 860× speedup, with the same identity, the same provenance, the same DAG entry.
The hologram cache is one of the four sovereign axes a frontier API cannot reproduce by design: an API charges full price every call, every time. Implementation is ~150 lines of C++ in src/runtime/hologram_cache.cpp.
EXACT_REPLAY_CACHE_V1.md.
Hot-expert cache
A pinned-RAM cache of the top-N most frequently activated MoE experts. Built during the V4-Flash phase to attack the 89 % expert-I/O wall. Empirical measurement: top-500 hot experts → 56 % RAM hits → −21 % wall time. The honest negative result preserved on the same site: top-1000 caused page-cache thrash (380 sec/q vs 100 sec/q baseline = 3.8× worse), recorded in the trajectory as a local-optimum trap.
Optimization passes table at /history#era1.
I
IndexNow protocol
An open submission protocol for telling search engines about new or changed URLs without waiting for the next crawl. Bing, Yandex, Naver and several other engines accept IndexNow pings; pinging api.indexnow.org propagates to the shared subscriber list.
CyberdyneLabs publishes its IndexNow key at /b9dc68d35f64b3f2476d11f3de37eebb.txt. Every site update fires a fresh ping; new-page indexation latency is typically < 24 hours on Bing/Yandex.
L
Legion graph
The hypergraph at the centre of ADAM: 1.2 million concepts and 6 million bonds in honeycomb topology. Each node carries a 1024-bit HDC vector, a position in Cl(3,0) Clifford algebra, an I-Ching hexagram state, and a Physarum conductivity score per outbound bond.
The graph reorganises around use: bond conductance evolves under the Physarum equation, active reasoning pathways reinforce themselves, stale paths decay. Self-supervised learning runs continuously in the background via JEPA + InfoNCE objectives.
/adam L3.
Line-addressable memory spine
The persistent memory layer of the CyberdyneLabs runtime. Every line of every report, doc, and surgery transcript carries a sha256[:16] address; the manifest at data/memory_spine/manifest_v1.jsonl indexes 305 files / 58 996 lines across data/organ_surgery/ · docs/ · reports/.
The spine is shipped indexed; exact-lookup CLI and TF-IDF semantic ranker are in build, not done — explicitly. (We say so on the program page.)
MEMORY_SPINE_INVENTORY_V1.md.
llms.txt · proposed AI-crawler standard
A short markdown file at the root of a website that tells LLM-based crawlers (ChatGPT, Claude, Perplexity, Bing AI, Google AI) what the site is about, what URLs to ingest, and under what license. Proposed by Anthropic; analogous to robots.txt for search engines.
CyberdyneLabs publishes both /llms.txt (short summary, links, key numbers) and /llms-full.txt (every program's full content as one markdown document for one-shot ingestion).
M
MACHINA · Program 06
An autonomous world simulator. Cognitive engines for ground robots, drones, and hybrid swarms that build worlds rather than navigate them. Two directions: Cognitive Mechatronics (sensorimotor cognition in a single embodied machine) and Dynamic Cognitive Engineering (colonies of machines engineering their own world in real time).
The substrate is N-dimensional cognitive space, not XYZ. Machines reason inside configuration manifolds where dimensions are degrees of freedom, energy budgets, and goal axes. A path in that space is a plan; a colony in that space is a factory.
/machina · live simulator at /machina#sim.
MerKaBa oscillator
The dual-torus oscillator at the top of ADAM's stack (Layer 0). Two interlocked tori at slightly different frequencies pulse semantic attention across the Legion graph; 1024 quantum clones process in parallel at every tick. The MerKaBa is not an attention mechanism — it is a rhythm that modulates which concepts are active in any given reasoning window.
/adam L0.
Mode A · Mode B · Mode C
The three test modes of every CyberdyneLabs benchmark.
Mode A — 7B-only baseline. The top-brain handles every task without any organ assistance.
Mode B — organ-only, 7B fallback forbidden. Measures what the surgery did to the 0.5B specialist alone. This is not the system's production performance — it is the surgical delta.
Mode C — production. Organ-first, with 7B fallback if verifier fails. The full system as users see it.
BD9 production state by mode: MBPP Mode-C 60/100 vs Mode-B 13/100; HumanEval Mode-C 81/164 vs Mode-B 6/164.
MBPP_HE_3MODE_V1.md.
N
NanoOS capsule
A sandboxed shell environment that captures stdout, stderr, exit code, and file-artefact hashes for every command a model emits. Each run produces dag/capsules/cap_*.json carrying a replay_recipe.spec_inline — the entire execution can be replayed on a stranger's machine and produce the identical stdout/stderr/exit code.
This is what we mean by proof-carrying execution: every model answer comes with a capsule that can be replayed for verification. Measured on Terminal-NanoOS-30: MONSTER 22/30 vs PARROT 20/30.
Spec: PHASE_12_NANO_OS_EXECUTION_SUBSTRATE.md.
O
Organ · specialist 0.5B-class model
One of the 8 narrow-task models in the Frankenstellm stack. Each organ has its own .planck pack, its own decoder spec (rep_penalty · ngram · max_tokens), its own prompt template, and its own verifier. As of 2026-05-05 (post BD9), 5 of 8 organs are surgered: code_skeleton · triz_contradiction · wound v2 · json_repair · claim_extractor. 2 are YELLOW (test_writer, cache_matcher); 1 is RED queued BD9.1 (renderer).
/frankenstellm Stack 02.
P
Physarum routing
A self-organising routing algorithm derived from Physarum polycephalum, a slime-mould species that solves the shortest-path problem between food sources without a central nervous system. Each edge maintains a conductivity score D updated by
dD/dt = |Q|^α − μD
where Q is the flow through the edge, α=0.6 keeps strong and moderate routes both viable, and μ=0.008 lets a dead route fade in ~125 steps. The CyberdyneLabs stack uses Physarum routing in three places: the PhysarumChain P2P layer, the bond-conductance update inside ADAM's Legion graph, and the inter-trail-pheromone diffusion in Hypercolony.
/physarum Feature 01.
PLANCK pack · .planck file format
The mmap-able weight format consumed by gigachad_native. PLANCK7B v1 supports Q4 group-128 quantisation with delayed scale, FP8 e8m0 → FP32 bit-shift expansion, and byte-for-byte verifier coverage (50/50 PASS). Writer is build/planck7b_tool; reader is src/runtime/planck_runner.cpp; verifier is tools/planck/planck_pack_verify.cpp.
Q4 group-128 packs Physarium-7B from 15.23 GB BF16 down to 5.55 GB resident in VRAM on RTX 3060 Ti, all 28 layers in one shot.
Phase 8A · 8E2 NUCLEAR.
Procedure memory
The persistent store of action sequences that ADAM has either discovered autonomously or ingested from outside sources. Held on disk inside adam_memory.bin as crystal_forms entries with a TSV sidecar (adam_arc3_procedures.tsv). Each procedure is keyed by a grid signature (FNV-1a over the 64×64 frame) so it can be retrieved when a similar scene reappears.
In ARC-AGI-3 the procedure memory is the difference between the autonomous 25/183 score (procedure memory empty at start) and the hybrid 183/183 score (procedure memory pre-loaded with Crystalline MIT-0 trajectories, ARC-SAGE Apache-2.0 trajectories, ADAM's own discoveries, and two human boss-level demonstrations).
See ARC closed loop and /adam ARC section.
Q
QLoRA
Quantised Low-Rank Adaptation. A training method that freezes a 4-bit quantised base model and trains only a small low-rank adapter on top. CyberdyneLabs uses QLoRA throughout the BD-series surgery: typical configuration is r=16, α=32, lr=2e-4, 6 epochs on 25–280 row training sets, then PEFT merge_and_unload + .planck repack. The strict-schema 4-axis gate decides whether the new pack ships.
tools/surgery/train_*_lora_*.py in the downloads page.
R
Repeat-learning
A test of whether the runtime actually learns between calls. The bench (repeat_learning_torture.py) presents the same N tasks across multiple rounds; on failure, the system writes a "scroll" to its memory spine with the canonical answer, then retries the same task in the next round. CyberdyneLabs measured MBPP repeat-learning round 2 on 20 problems: MONSTER 13/20 vs same-weights PARROT 12/20. Round 1 PARROT was ahead — round 2 MONSTER overtook because PARROT cannot write its own scroll between rounds.
MBPP_REPEAT_LEARNING_V1.md.
S
Semantic-fee curve
An optional fee mechanism on PhysarumChain where the fee scales with the structural complexity of the transaction's Cl(4,1) multivector — specifically, the grade-2 norm. In plain words: more information encoded → slightly higher fee. The floor remains 0.0000001 MKB (the anti-spam floor); contract interactions pay ~10× the floor through the grain-priced gas system.
/physarum#fees.
Singularity Monolith
The 1.60 GB resident-VRAM weight pool that allowed DeepSeek-V4-Flash to run on a single 8 GB GPU. Packed as 430 594 648 uint32 blocks, indexed by 4 992 projection entries (singularity_index.json + singularity_scales.json). The shared backbone lives in the monolith; routed experts are streamed from disk via expert streaming.
V4_FLASH_TECH_BRIEF.md.
Substrate-explore-fallback
An ADAM mechanism that triggers when the beam-search frontier in ARC-AGI-3 saturates — too many no-op duplicates, no scoring lift. The runner hands control back to ADAM's substrate, which produces an exploration prior from lg_quantum_think activations rather than the deterministic scoring loop. New trajectories from the fallback are pushed back into the beam for the next expansion round.
Together with warmed world_model, this is one of the two mechanisms responsible for the 24 → 25 / 183 climb between 14 May and 18 May.
Reference: ADAM Phase 162 logs.
T
Tessaractic substrate · 4D hypercube grid
The simulation substrate of Hypercolony — a 16×16×8×5 hypercubic grid (10 240 cells total) with walls, food, light, and a 4-dimensional pheromone field. Agents move through the fourth axis as naturally as through the first three. Three competing clan strategies (Lexicons / Phero-Mystics / Solar Nomads) inhabit the same substrate; the asabiyyah cycle emerges from local rules.
/hypercolony-app.
TRUTH_LEDGER
The single source of truth for the CyberdyneLabs stack. The first TRUTH_LEDGER.md (2026-04-27) introduced the A/B/C/D categorisation of every claim — measured · surgery · scaffold · unsafe. Updates land first in CURRENT_TRUTH_LEDGER.md (2026-04-29 onward). When numbers change, the ledger tracks the delta and tags errata; previous statements are marked, not deleted.
CURRENT_TRUTH_LEDGER.md.
V
V4-Flash · DeepSeek-V4-Flash flagship demo
The flagship phase of the CyberdyneLabs research arc (2026-04-21 → 04-26). DeepSeek-V4-Flash, an open-weight Mixture-of-Experts model with 284 billion total parameters (13 billion active per token, ~148 GB on disk in FP4), was driven through end-to-end inference on a single RTX 3060 Ti, 8 GB VRAM, 13 GB RAM, 80 GB swap, WSL2. Decode best 7.5 s/tok, p50 9.6 s/tok, cold prefill 47.3 s. Roofline: 89 % wall = expert I/O.
The phase produced every piece of infrastructure the rest of the lab now runs on: the Singularity Monolith, the hot-expert cache, the singularity index, the .planck pack format, and the C-extensions for packed Q8 decode.
V4_FLASH_TECH_BRIEF.md.
W
Warmed world_model
A run-persistent state inside ADAM that accumulates frame deltas, causal biases, and grid signatures across ARC-AGI-3 attempts. On every subsequent run, the world_model gives the beam search a non-empty prior on the very first frame — actions that were useful in the past on a similar scene are scored higher than random.
Together with substrate-explore-fallback, this is what turned the 14 May score of 24/183 into the 18 May score of 25/183. The world_model is the carrier of "experience changes memory" in the ARC closed loop.
Reference: ADAM Phase 162 logs.
Wound organ · phys05_wound
The repair organ that runs after verifier-fail in the Frankenstellm rescue path. Together with critic_lite it forms the BD8 surgery target: when an organ output fails verification, critic produces a diagnosis and wound emits a patched output, and the verifier re-runs.
Honest state: rescue rate on ARIZ JSON repair is 0/n across BD8 V1–V5. The mechanism is wired; the repair organs need retraining on JSON failures, not the terminal-stderr distribution they were trained on. Wound v2 is retained in production for the in-chat rescue path (where it does work).
RUNTIME_ORGANISM_BENCH_V5.json.