A silver-haired cyberelf operating luminous evaluation instruments inside a dark AI systems lab

Independent AI systems lab 2026

Build strange systems. Measure them honestly.

Cyberelf Labs builds four independent authorities for bounded model change: SalienceLean proves the contract, Whetstone measures behavior, Foundry binds the decision, and Backplane governs execution.

Run the full loop Map the stack

Founded by Justin Garringer / Waterloo, Iowa

Descend
01

Product architecture

One loop // Four independent authorities

A model change should not
approve itself.

Each product owns one narrow decision. Formal truth cannot erase a behavioral regression; a benchmark win cannot mint runtime authority; a deployment cannot silently become the next baseline. Foundry closes the transaction only when every required receipt agrees.

CYBERELF // GOVERNED IMPROVEMENT REFERENCE LOOP FAIL CLOSED
00 / INPUT Candidate bounded intervention
01 / FORMAL SalienceLean what must be true 02 / EMPIRICAL Whetstone what is actually true
03 / TRANSACTION Foundry all required receipts agree 04 / RUNTIME Backplane what may execute
05 / LEDGER Next baseline adopted or unchanged
FORMAL ∧ EMPIRICAL ∧ IDENTITY ∧ RUNTIME BOUNDED CHANGE — NOT AGI OR RSI
01 / PROVESalienceLeanTyped Lean obligations and sealed proof receipts.Open formal surface ↗ 02 / MEASUREWhetstonePaired adversarial evaluation with regression vetoes.Open empirical surface ↗ 03 / DECIDEFoundryIdentity-bound admission and baseline adoption.Run the control loop ↗ 04 / EXECUTEBackplaneSigned, scoped, reversible runtime capability.Enter the runtime ↗
02

Empirical authority

Whetstone // Deployment authority

Capability is not a claim.
It is a promotion decision.

Whetstone is the stack’s empirical authority: infrastructure for deciding whether a model, adapter, prompt, workflow, or agent has actually earned promotion—without training on the exam that judged it.

Open the live demo Published cycle evidence
◇ PROMOTION GATE / GPU CYCLE
COMMITTED
Candidate: 38 of 38. Baseline: 20 of 38. Eighteen gains, zero regressions, and twenty ties. Promotion passed.
Candidate38 / 38
Baseline20 / 38
Gains18
Regressions00
38 PAIRED PROBESSTRICT ROUTED SYSTEMGPU MEASURED
01Private cohortExam stays sealed
02Verifier gradingAuthority is earned
03Paired evidenceGains vs. regressions
04Pass / Hold / BlockA decision, not a vibe
03Formal authority

SalienceLean // Prove the narrow contract

Machine-checked.
Meaningfully
bounded.

SalienceLean accepts typed obligations, renders an allowlisted Lean certificate, compiles it with a pinned toolchain, audits its axioms, and seals the result. It proves the declared statement—not that the surrounding model is universally safe.

Inspect the proof surface Source + receipts
SL // REFERENCE MANIFEST 5 / 5 PASS
SCHEMAPASS TEMPLATEPASS KERNELPASS AXIOMSPASS RECEIPTPASS
MANIFEST176fb99eeac390db…0 SORRIES
04Transaction authority

Foundry // Close the governed loop

A score is not
a deployment decision.

Foundry preserves the independent vetoes, binds every receipt to the same artifact, authorizes a signed Backplane deployment, and records adoption only after the deployed candidate becomes the explicit next baseline.

Run BLOCK → ADOPTED Control plane + evidence
FY // MEASURED GPU CLOSURE ADOPTED
01SalienceLeanformal obligationPASS
02Whetstone18 gains / 0 regressionsPASS
03Foundryidentity + admissionPASS
04Backplanesigned runtime packageDEPLOYED
05Ledgercandidate becomes baselineADOPTED
SEALED BUNDLE580ee395b914ba9c…REFERENCE CYCLE

Measured reference cycle on local hardware; not a customer deployment or a claim of AGI, RSI, or open-ended autonomy.

05

Runtime authority

Backplane // Governed runtime authority

The model stays frozen.
The system keeps growing.

Backplane turns one open-weight model into a governed platform. Mount signed policy, memory, tools, sensors, adapters, and bounded activation cards per tenant—without forking the base model or surrendering isolation, telemetry, and rollback.

Run the live sandbox Runtime + architecture
BP // TENANT-07 / STACK ACTIVE GOVERNED
GPU artifact4-bit
Runtime packageSigned
Isolation failures0 / 75
Control-plane bench82.55 r/s
DECLARATIVE PACKAGESED25519 SIGNEDRECEIPT EMITTED
01

One model, many products.

Each tenant gets an explicit card stack instead of a new fine-tuned fork.

02

Change stays reversible.

Signed packages, capability negotiation, scoped state, and receipts make every extension inspectable.

03

Research meets serving.

The measured closure deployed a real 4-bit model plus private adapter; the public sandbox exposes the same admission and execution contract without exposing the artifact.

06

Operating thesis

The lab doctrine

Measurement is not paperwork.
It is part of the machine.

A / EVIDENCE

Receipts over vibes.

Every ambitious system needs an honest instrument panel: paired outcomes, replayable runs, visible regressions, and a decision boundary that fails closed.

B / BOUNDARIES

Private by construction.

Secrets stay inside the grading boundary. Exposed items burn. Public interfaces reveal outcomes and provenance—not the exam that produced them.

C / EXPLORATION

Strangeness must survive contact.

Unusual ideas are welcome. They still have to run, recover, generalize, and produce evidence strong enough to outlive the demo.

07

Selected systems

Research that executes

Things that survived
contact with reality.

Cyberelf Labs sits between research and product: a local-GPU visual workstation, digital organisms, continual-learning rigs, explicit boundaries, and enough instrumentation to know when the magic is fake.

07A / VISUAL SIGNAL SYSTEMS

Entoptic

A browser-native visual signal workstation for patching typed device racks across color, depth, motion, masks, normals, confidence, and stereo. Real-time shader synthesis executes on the client GPU; frames stay in the browser unless exported.

Open the live workstation
  • WebGL
  • Client GPU
  • Typed surfaces
Morpho digital organism growing, receiving a line cut, and recovering

07B / DIGITAL BIOLOGY

Morpho

A CUDA petri dish where neural cellular automata grow tissue, metabolize, take cell-level injuries, and repair. Genomes mutate, lineages replay, and failure is visible—not averaged away.

  • PyTorch
  • Neural CA
  • Evolution archive

07C / CONTINUAL LEARNING

Substrate
Search

A search rig one layer below model architecture. It evolves memory, consolidation, context routing, trust gates, and contradiction repair through replayable lifetimes—then keeps a behavioral atlas instead of crowning one fake universal winner.

Related mechanism study
08

Founder

Justin Garringer // Founder, Cyberelf Labs

A small lab for
hard systems.

Justin founded Cyberelf Labs in Waterloo, Iowa to work where familiar tools produce misleading confidence: an eval that rewards leakage, a learner that forgets its past, or a digital organism that only looks alive.

The lab is intentionally lean. The work spans AI evaluation, agent reliability, continual learning, digital biology, and whatever comes next when the yardstick matters as much as the thing being measured.

GitHub

Have a hard system and a weak yardstick?

Let’s sharpen it.