Know what can be separated.
Read tensor names, shapes and model configuration. Keep embeddings, attention, shared experts and the output head. Dense models and unknown packed formats stop here.
CPU metadata + safe tensor I/OA PLAIN LANGUAGE INTRODUCTION
Turn one giant AI model into a team of focused specialists.
A Mixture of Experts model already contains many smaller expert networks. Expert Foundry finds the useful groups, packages them into smaller working models, and tests whether they keep the parent model’s ability.
MetaHarness chooses the right specialist for each request. LatentMesh can pass verified task state between them. The original full model remains the fallback whenever a specialist is uncertain or fails its quality target.
- Starts with
- One compatible MoE model
- Creates
- Smaller specialist models
- Decides with
- Measured quality and cost
FRAME 01 · THE RAW MATERIAL
A large model already contains specialists.
MoE models use different expert networks for different tokens. Today they usually travel as one large package.
EF 01
This animation explains the intended system. Real model quality and GPU performance still require the qualification test described below.
TRY IT / SYNTHETIC MODEL
See expert separation work.
Choose how many experts to keep. The browser measures which groups are most useful, builds three specialists, and compares them with the full model and random selection.
02
SYNTHETIC FIXTUREExpert partition
SHARED BACKBONEPreserved in every specialist
ROUTED EXPERTS24 available · top 2 active
Selected for domainRemoved from variant
Ready. Run a fresh calibration and compare against the parent and random pruning.
Run receipt
Teacher agreement—Selected / synthetic holdout
Experts per variant—Same top 2 active per token
Calibration + selection—Measured on your browser CPU
Production promotionBlockedReal model evidence required
| Candidate | Teacher agreement | Output MSE | Evaluation wall time | Evidence |
|---|
The fixture has planted domain locality. Agreement is not task accuracy. Timings include JavaScript overhead and are not LLM throughput. No joule, RF, or GPU measurement is implied.
HOW IT WORKS / SCROLL TO EXPLORE
One model.
Many accountable mixtures.
A request router chooses a specialist.
A token router chooses its experts.
Neither replaces the other.
01 / INSPECTCONTROL ≠ COMPUTE
Parent MoEFrozen revision + tokenizer
MetaHarness adapterQuality floor · cost budget · fallback
01CodeBackbone + layer masks
02ReasoningBackbone + layer masks
03ToolsBackbone + layer masks
LatentMesh boundaryTyped deltas + hash-bound state
Inspect the checkpoint before any tensor is changed.
Test every routed expert.
Collect per-layer activation evidence on a calibration split. Score groups, not imaginary universal “coding neurons.” Compare selected masks with random masks on untouched tasks.
GPU useful for real-model forward passesA mixture of expert mixtures.
Export smaller compatible checkpoints with remapped routers. An outer router chooses the cheapest validated specialist that meets the quality floor. Otherwise, return to the parent.
Outer request route / inner token routeShare state, not assumptions.
Exchange typed state changes with base hashes, sequence checks and session identity. Raw latent and cross-specialist KV transfer stay disabled. M3 and M5X remain negative evidence.
No claim of RF or latent-channel gainSmaller is a hypothesis.
Measure quality, peak device memory, latency and mixed-task robustness under the same budget. Fewer resident experts do not guarantee faster tokens. Human review owns promotion.
Parent and random controls requiredMETAHARNESS / REQUEST ROUTER
Choose the right mixture.
Runs the pinned MetaHarness router locally. These example quality scores and costs are synthetic, not model benchmarks.
Raise the floor above 0.94 to test parent fallback. Routing never authorizes model execution or new tools.
{
"router": "@metaharness/router@0.4.0",
"level": "request",
"fallback": "parent",
"productionAuthorized": false
}LATENTMESH / STATE LAB
Send only what changed.
A working state delta with integrity checks. The wire bridge is a separate integration gate.
Identity binds parent, expert mask, tokenizer, precision, tenant and session.
{
"mode": "typed-state-only",
"rawLatentsEnabled": false,
"transport": "in-process",
"integrity": "SHA-256 (not authentication)"
}EVIDENCE / NO SILENT PROMOTION
What has actually passed?
Safe separation pipeline
CPU checkpoint inspection, per-layer masks, floating safetensors export, router remapping and artifact hashes. Supported schemas are checked explicitly.
Tests + fixture benchmarks
Run the local receipt and fixture controls. Exact executed results are included with the pipeline.
CPU fixture benchmark ↗Real model performance
No production checkpoint, GPU benchmark, ruvLLM loader roundtrip, remote latent transport, or wireless utility claim has passed this build.
≤ 2 ppTask quality loss
≥ 30%Peak memory reduction
≤ 10%P95 latency regression
1,000 × 5Holdout tasks × seeds
Proposed acceptance gates, not achieved results. Also require a random control, mixed-domain validation and verified artifact hashes. Evaluation is separate from promotion.
ARCHITECTURE DECISION RECORDS
The decisions behind the build.
ADR 001Separate groups with the shared backbone Accepted
An expert feed-forward block is not a standalone language model. Each export preserves shared tensors and renumbers both experts and router rows. Shared serving may reduce fleet duplication later; it is not a prerequisite.
ADR 002Keep two levels of routing Accepted
MetaHarness handles request-level specialist selection. The model runtime handles token-level top-K routing and residency. A quality miss uses the parent. Route changes never grant new tool permissions.
ADR 003Typed state first; raw latents gated Accepted
LatentMesh integration starts with explicit semantic state. Every cache is bound to the parent revision, masks, tokenizer, precision, position encoding, tenant and session. Hashes detect corruption; authenticated transport is still required.
ADR 004Streaming CPU export, GPU evaluation Accepted
Separating existing tensors is an I/O problem and should not consume GPU memory. Calibration and heldout generation are compute tasks. The hosted workbench does not execute production weights or rent a GPU.
ADR 005Evidence before optimization or release Accepted
CPU synthetic timings qualify implementation behavior only. Real quality, memory and latency must be measured on the same sealed task set before promotion. Ruflo coordinates the work; promotion remains human governed.