mervRECURSIVE SELF-IMPROVEMENT
RESEARCH SYSTEM / MERV

Ideas become
more than ideas.

New directions. Parallel experiments.
The infrastructure to keep going.

FOLLOW ONE LOOP
01Ideas02Workloads03Infrastructure
ILLUSTRATIVE LOOP / NOT LIVE TELEMETRY
01
IDEATION

Find the next what if.

Invent a direction.
Combine what already works.

EXISTING RESEARCHpapers / methods / evidenceNEW HYPOTHESESquestions worth testingPRIOR LOOPretained findingsYOUR OBJECTIVEmodel / data / constraintsreplayLoRAroutingdistillationCOMBINE · CHALLENGE · PROPOSEreplay × low-rank adaptationsparse expert routingsuccessor distillationEXISTING RESEARCHmethods + evidenceNEW HYPOTHESESquestions + intuitionCOMBINEPRIOR LOOPYOUR GOALCHALLENGE → PROPOSEReplay × LoRASparse routingDistillation
PROPRIETARY RESEARCH + OPEN-ENDED EXPLORATIONDIRECTIONS IN MOTION
HYPOTHESES → EXECUTION GRAPH
02
SCHEMA → ACTION

One loop. Many things moving.

Tasks prepare the ground.
Experiments explore in parallel.

LOOP / 0012 TASKS → 3 INDEPENDENT EXPERIMENTS PARALLEL BRANCHES
DATA + EVALUATIONS READYReview evidence → reflect
E01

Compare replay ratios and adapter ranks. Starts when T01 and T02 finish; it does not wait for another experiment.

H100 · 4 independent seeds
Experiments wait on tasks. Never on each other.Next loop → new hypotheses + successor experiments
PARALLEL WORKLOADS → PROVISIONED COMPUTE
03
INFRASTRUCTURE

Under every idea,
a lot of machines.

VMs spin up. GPUs get to work.
Agents stay with the research.

WORKLOAD DISPATCH
dataset/v4eval-suite/v2experiment speccheckpoint lineage
COMPUTE FABRIC / RUNNING
ILLUSTRATIVE TOPOLOGY · 24 WORKERS
8× H100 SXM · 80 GB× 04
VM-01train
replay-adapter.1468 steps
VM-02train
replay-adapter.2483 steps
VM-03train
replay-adapter.3491 steps
VM-04train
replay-adapter.4492 steps
stage → run → capture → release↳ retained artifacts
p4d.24xlarge · 8× A100× 04
VM-05train
routing-ablation.1453 steps
VM-06train
routing-ablation.2467 steps
VM-07train
routing-ablation.3474 steps
VM-08train
routing-ablation.4474 steps
stage → run → capture → release↳ retained artifacts
a3-highgpu-1g · 1× H100× 04
VM-09train
routing-seed.1439 steps
VM-10train
routing-seed.2452 steps
VM-11train
routing-seed.3457 steps
VM-12train
routing-seed.4455 steps
stage → run → capture → release↳ retained artifacts
NC24ads A100 v4 · 1× A100× 04
VM-13train
held-out-eval.1425 steps
VM-14train
held-out-eval.2436 steps
VM-15train
held-out-eval.3440 steps
VM-16train
held-out-eval.4436 steps
stage → run → capture → release↳ retained artifacts
8× H100 SXM · 80 GB× 04
POD-17train
distill-sweep.1411 steps
POD-18train
distill-sweep.2421 steps
POD-19train
distill-sweep.3423 steps
POD-20train
distill-sweep.4418 steps
stage → run → capture → release↳ retained artifacts
4× H100 · GPU sandbox× 04
BOX-21train
successor-trial.1397 steps
BOX-22train
successor-trial.2405 steps
BOX-23train
successor-trial.3406 steps
BOX-24train
successor-trial.4399 steps
stage → run → capture → release↳ retained artifacts
OUTPUT STREAMmodel.safetensorseval-results.jsontraining.logdataset.manifestcheckpoint/step-1200
DATASETS / CHECKPOINTS / EVIDENCEKept after the machine is gone.SSH certificates · durable jobs · snapshots · object storage
THE PROVIDER LAYER

One research workflow.
An entire compute ecosystem.

20 cloud provider adapters in Merv Sandboxes.
Lambda
AWS
Google Cloud
Azure
RunPod
Modal
Crusoe
Cloudflare
DigitalOcean
Vast.ai
Fluidstack
Hyperstack
Paperspace
Vultr
Verda
TensorDock
Thunder Compute
Voltage Park
Akamai / Linode
GiveMeANode
TRAINING APIs / PLANNED

Another path from agent to training run.

Tinker is part of the integration roadmap.

SDK → TRAIN → SAMPLE → EVALUATE
Integration coverage

Provider adapters are implemented. Live research workflows have been validated on Lambda A10 and Cloudflare CPU; DigitalOcean and GiveMeANode have CPU transfer validation. Other providers and hardware configurations require live validation. The fleet above illustrates possible workloads, not a running deployment or a performance claim.

Evidence returns. New ideas emerge.NEXT LOOP
BUILD YOUR RESEARCH LOOP

What are you
trying to discover?

Let’s put it in motion