The platform

Five layers. One stack.

CASTLE · Edge · Algorithm · Knox · Supercharge. Each pillar deployable independently or as the full stack.

Platform livepillars|5edge_params|1.34Batlas_tables|55supercharge_components|6knox_signing|ML-DSA-87api_spend|$02026-05-12T17:35Z
5
Platform pillars
CASTLE · Edge · Algorithm · Knox · Supercharge
1.34B
Edge parameters (live)
release-v1 · best_val 3.3518
6
Supercharge components
meta · remember · verify · loom · speculate · oracle
ML-DSA-87
Knox PQ signing
FIPS 204 Level 5
55
Atlas tables
Canadian data substrate
$0
External API spend
local inference only
The stack · top to bottom

One platform. Five layers.

PILLAR 05
SUPERCHARGE
6-component cognitive scaffold · wraps Edge-11B
PILLAR 03
AXE ALGORITHM
CSR recall + ranking · embeddings at write-time
PILLAR 04
KNOX
ML-DSA-87 post-quantum signing · audit chain · fail-closed
PILLAR 01
CASTLE
Autonomous orchestration · fleet routing
PILLAR 02
EDGE
Foundation model · 1.34B parameters · BPE-32K
Pillar 01 · CASTLE

Autonomous orchestration.

The autonomous sprint engine.

CASTLE reads markdown task lists, classifies each item by type (code · docs · analysis · testing · data-gen · security), routes tasks to the best available fleet machine + model, saves results to ~/.axe/castle-results/, and reports cycle summaries to axe.observer. Runs every 30 min via LaunchAgent on local hardware — zero paid API spend.

Routing
best-fit machine + model via Ollama / llama-server HTTP
Cadence
30-minute cycle · self-suggests new tasks every 3rd run
Atlas table
agent_traces (heartbeat) · castle_interactions (future API)
Health
GREEN — cycle #178+ as of 2026-05-12
Pillar 02 · Edge

The foundation model.

Canada's native LLM, trained from scratch.

Edge-2 Large is a 1.34B-parameter language model trained from scratch on Apple Silicon via MLX. RoPE positional encoding, SwiGLU activations, grouped-query attention, BPE-32K tokenizer. Not a fine-tune of a foreign base — the weights are ours, the architecture is ours, the corpus is Canadian. Q3 2027 target: Edge-3 at 11B parameters as the household-name Canadian model.

Current size
1.34B parameters (release-v1)
Architecture
RoPE · SwiGLU · GQA · BPE-32K
Hardware
JL1 Mac Studio · 64GB · MLX
Best val_loss
3.3518 at step 56,000
Roadmap
Edge-3 11B by Q3-Q4 2027
Pillar 03 · AXE Algorithm

Recall and ranking.

ByteDance Monolith, but Canadian — and with embeddings at write-time.

Core primitive: CSR (Cuckoo-Sharded Recall) — a collision-free hashmap + HNSW index serving three consumers: recommendation recall · oracle external knowledge retrieval · per-customer cross-conversation memory. Semantic embeddings are indexed at WRITE-TIME, which legacy Monolith deployments can't retrofit due to technical debt. Moat: the user-interest graph compounds richer across the Castle WireGuard mesh.

Primitive
CSR · cuckoo-sharded recall + HNSW
Consumers
3 — recall · oracle · per-customer memory
Differentiator
embeddings at write-time
Backend
Qdrant on 10.10.0.1:6333 (Castle WG mesh)
Schemas
csr_schemas.py — 53 invariant tests · pass at runtime
Pillar 04 · Knox

Procurement-grade security.

Post-quantum cryptography. Audit chain. Fail-closed.

Knox is the security pillar — the file-level provenance trail that makes AXE deployments survive a government procurement audit. Post-quantum signing via ML-DSA-87 (FIPS 204 Level 5, CNSA 2.0). Continuous watchdogs verify audit-chain integrity, revocation- bundle freshness, key-rotation policy compliance, and EMA (Edge Model Attestation) bundle expiry across the fleet. Every certificate that goes stale is a CRITICAL alert. Every quarantine is logged in the same chain.

Signing
ML-DSA-87 (FIPS 204 Level 5 · CNSA 2.0)
Watchdogs
4 — audit_chain · revocation · key_rotation · attestation
Cadence
local 5-15min · external cross-node 5min · boot preemptive
Audit chain
~/.axe/knox/audit/chain.jsonl (HMAC-SHA256 envelope)
CI
GitHub Actions on memjar/axe-knox · 104 tests gate every merge
Pillar 05 · AXE Supercharge

The cognitive scaffold.

Six components wrap Edge for 70B-class quality on narrow tasks.

Supercharge is the post-training cognitive scaffold that wraps Edge-11B-Instruct at inference time. Six components: axe-meta · axe-remember · axe-verify · axe-loom · axe-speculate · axe-oracle. Goal: 70B-class performance on regulated-domain narrow tasks, air-gapped on Apple Silicon, ~$0.0001/query. Direct sales only — this layer is private commercial — never publicly licensed.

Components
6 — meta · remember · verify · loom · speculate · oracle
Currently live
axe-remember (c-data JSONL capture pipeline)
Phase
F (Q1 2028+)
Cost target
~$0.0001/query
Distribution
direct sales only · private repo · no public weights
Cross-pillar capabilities

What ships with every deployment.

Atlas data layer
Atlas · 55 tables · Canadian Postgres on DO droplet
Real-time messaging
AXP envelope protocol · HMAC-SHA256 · 14-agent bus
Knowledge ingestion
axe-obsidian Crown · vault + Apple Notes + Atlas rows + SFT pairs
Continuous training
forge-capture · live JSONL capture · nightly SFT pair extraction
Auditable inference
Knox EMA attestation on every model serve · chain-replayable
Cross-node mesh
Castle WireGuard mesh · zero-trust between nodes
Observer telemetry
axe.observer · triple-surface broadcast (bus + observer + Atlas)

The architecture isn't marketing.
It's the procurement story.

Deploy on your iron.

Procurement timelines we've hit: 30 days from first contact to signed PO. Deployment timelines: hours to days, not months.

Start the conversation