Everything through the front door
Policy, memory, sensor state, and tool results become prompt content. Every request pays the context cost.
OPEN-WEIGHT INFRASTRUCTURERUNTIME 0.1
Backplane mounts tenant policy, memory, tools, adapters, sensors, and bounded activation cards onto one open-weight model—with signatures, isolation, batching, telemetry, and rollback.
01 / THE MISSING LAYER
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.
Policy, memory, sensor state, and tool results become prompt content. Every request pays the context cost.
Sequential tuning entangles tenants and skills. Rollback becomes checkpoint management.
Route an ordered card stack per tenant. Enforce what it may read, write, retain, and spend.
02 / THE PRODUCT
Backplane does not claim to invent hooks, LoRA, memory, or guardrails. It gives them one governed execution and evidence boundary.
Resolve the tenant, verify its stack, batch only compatible requests, execute against the declared backend, emit a receipt.
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"Inspect stacks, model slots, events, limits, digests, and causal interventions before a capability reaches production.
Roll out one card version, watch its effects, compare receipts, then detach or roll back without replacing the base model.
03 / RUN IT
Switch tenants and watch the ordered card stack, response, event trace, and stack digest change. These are fictional pilot scenarios, not customer claims.
Private policy + approved support memory
This public endpoint runs the real isolation, policy, scheduling, and receipt path against a deterministic evidence backend—not a hidden GPU model.
The response arrives with its ordered cards, stack digest, events, timing, and backend limitation notes.
04 / EVIDENCE, NOT VIBES
These are indexed receipts from the separate research workspace. They motivate the product; they are not presented as fresh production benchmarks.
Qwen3-1.7B research receipt
merged control forgot 19/20 → 0/20
same program across Qwen + Ministral
external value never appeared in prompt
negative result retained
same channel, honest training 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
Hashed keys, namespaced state, scoped stacks and receipts.
SHIPPEDCompatible microbatches and KV-cached HF generation; scale tests remain.
PARTIALGoverned HTTP + per-request LoRA. Residual worker plugin is next.
PARTIALEd25519 packages, strict contents, trusted publisher sets.
SHIPPEDCross-model receipts retained; 7B/14B production run remains.
PARTIALControl-plane benchmark ships; matched RAG/LoRA study remains.
PARTIALThree-tenant pilot is deployable; external criterion remains.
PILOT READY06 / THE WEDGE
FIRST BUYER
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.