OPEN-WEIGHT INFRASTRUCTURERUNTIME 0.1

The model stays frozen.The system keeps growing.

Backplane mounts tenant policy, memory, tools, adapters, sensors, and bounded activation cards onto one open-weight model—with signatures, isolation, batching, telemetry, and rollback.

HOSTFROZEN
CARDSVERSIONED
STATETENANT-SCOPED
FAILUREREVERSIBLE
HOST / FROZENRUNTIME LIVE
00010203040506070809101112131415161718192021222324252627
RESIDUAL
MEMORY
CONTROL
WORLD
VERIFY
01Support boundaryVERIFY / OUTPUT
02Support memoryMEMORY / RESIDUAL
BASE WEIGHTSUNCHANGEDSTACK DIGESTSUPPORT-A17C
ONE BASE MODELMANY GOVERNED SYSTEMS

01 / THE MISSING LAYER

Models are shipped like files.Products need systems.

Today, capability usually means another prompt, another retrieval path, another fine-tune, or another model copy. Those pieces work. What is missing is one enforceable runtime contract for composing them after deployment.

TOKEN MIDDLEWARE

Everything through the front door

Policy, memory, sensor state, and tool results become prompt content. Every request pays the context cost.

MONOLITHIC ADAPTATION

Every update touches shared behavior

Sequential tuning entangles tenants and skills. Rollback becomes checkpoint management.

BACKPLANE

Mount capability as governed state

Route an ordered card stack per tenant. Enforce what it may read, write, retain, and spend.

HOST

02 / THE PRODUCT

A control plane insidethe model lifecycle.

Backplane does not claim to invent hooks, LoRA, memory, or guardrails. It gives them one governed execution and evidence boundary.

01

Runtime

Resolve the tenant, verify its stack, batch only compatible requests, execute against the declared backend, emit a receipt.

  • OpenAI-compatible API
  • HF + KV cache path
  • vLLM / LoRA routing
  • Scoped state + receipts
02

Card SDK

Package capability as a signed manifest and bounded JSON configuration—not executable code hiding behind a permission label.

"writes": ["VERIFY", "OUTPUT"]"max_latency_ms": 50"signature": "Ed25519"
03

Workbench

Inspect stacks, model slots, events, limits, digests, and causal interventions before a capability reaches production.

04

Fleet control

Roll out one card version, watch its effects, compare receipts, then detach or roll back without replacing the base model.

03 / RUN IT

One runtime.Three incompatible tenants.

Switch tenants and watch the ordered card stack, response, event trace, and stack digest change. These are fictional pilot scenarios, not customer claims.

LIVE CONTROL-PLANE SANDBOX
ACTIVE TENANTNorthstar Support

Private policy + approved support memory

Support boundaryVERIFY / OUTPUT
Support memoryMEMORY / RESIDUAL

This public endpoint runs the real isolation, policy, scheduling, and receipt path against a deterministic evidence backend—not a hidden GPU model.

EXECUTION RECEIPTREADY
Mount a tenant stack and run one request.

The response arrives with its ordered cards, stack digest, events, timing, and backend limitation notes.

04 / EVIDENCE, NOT VIBES

Keep the gain.Keep the negative result.

These are indexed receipts from the separate research workspace. They motivate the product; they are not presented as fresh production benchmarks.

010.0

null-card max logit delta

Qwen3-1.7B research receipt

0220/20 + 20/20

conflicting skills retained

merged control forgot 19/20 → 0/20

0318/18

control decisions replayed

same program across Qwen + Ministral

0423/24

wire-fed reasoning probes

external value never appeared in prompt

051/20

end-to-end latent link

negative result retained

0619/20

directly supervised link

same channel, honest training boundary

PROVENANCE BOUNDARY

Current repository tests validate isolation, packaging, scheduling, API behavior, and the public sandbox. Model capability receipts remain separately labeled until re-run.

Read the receipts ↗

05 / SEVEN DILIGENCE GATES

What ships.What still has to earn it.

01

Tenant isolation

Hashed keys, namespaced state, scoped stacks and receipts.

SHIPPED
02

Batch + KV metrics

Compatible microbatches and KV-cached HF generation; scale tests remain.

PARTIAL
03

vLLM path

Governed HTTP + per-request LoRA. Residual worker plugin is next.

PARTIAL
04

Signed cards

Ed25519 packages, strict contents, trusted publisher sets.

SHIPPED
05

Larger workloads

Cross-model receipts retained; 7B/14B production run remains.

PARTIAL
06

Direct comparison

Control-plane benchmark ships; matched RAG/LoRA study remains.

PARTIAL
07

Design partner

Three-tenant pilot is deployable; external criterion remains.

PILOT READY

06 / THE WEDGE

FIRST BUYER

One private model.Every tenant gets its own system.

Backplane starts with AI infrastructure teams serving open-weight models on-premises or in private clouds. Mount a different policy, memory, skill, adapter, and tool stack per customer without copying the host.

FROZEN HOSTPOLICYMEMORYSENSORSKILL
BACKPLANE // CYBERELF LABS

Stop rebuilding the model.Build the system around it.

Clone Backplane justin@cyberelf.link