Actuator

The end-to-end control layer for model transformation

By monitoring and adjusting training dynamics to course-correct when quality drifts, Actuator preserves what makes your models smart, before the damage is done.

Get early access →

How it works

Actuator sits underneath every stage of post-training — fine-tune, align, distill, compress — and runs a closed-loop four-step cycle on each: monitor protected metrics, adjust pressure when they drift, guardrail against regressions, and certify the run with a full audit trail.

Fine-tune
↑↓
Align
↑↓
Distill
↑↓
Compress
↑↓
Actuator Control Layer
Monitor→ Adjust→ Guardrail→ Certify
↔ continuous closed-loop feedback
Multiple patents pending

Better control, better models

Today's post-training stack is a fragmented, open-loop affair. Teams pull together multiple tools, set knobs, run blind, eval after, and repeat. Actuator replaces that process with continuous live monitoring, automatic training-time adjustments, and guardrails to keep your model transformations on track. Quality in, quality out.

Safety & Alignment
+56.9pp
Worst-group accuracy @ 90% sparsity
Actuator stays unbiased while post-hoc methods collapse under aggressive compression.
0.0 0.2 0.4 0.6 50% 70% 90% Sparsity Worst-group accuracy
Actuator
Post-hoc CDA
Post-hoc ERM
CivilComments Fairness — DistilBERT
Robustness @ Compression
+13.1pp
Math reasoning vs. Wanda @ 50% sparsity
Co-scheduled compression recovers more quality across six benchmarks (GSM8K, MMLU shown).
GSM8K (Math Reasoning)
Actuator
39.0%
Wanda
25.9%
Post-hoc LoRA
12.8%
MMLU
Actuator
51.7%
Wanda
48.5%
Post-hoc LoRA
46.6%
Pruning — Llama-3.2-3B
Serving Speed & Latency
+7.8%
Acceptance rate vs. pure distillation
Actuator-discovered teacher guidances outperform pure distillation and even hand-crafted rules.
Draft Model Acceptance Rate
Actuator (aligned)
25.9%
Actuator (curated)
24.3%
Baseline
24.0%
Distillation to 0.9B — DeepSeek-Coder-6.7B
Reasoning & Capability
+15.8pp
Alignment preservation vs. DPO alone
Closed-loop stability prevents 2.3× more capability drift while maintaining reinforcement learning.
GSM8K (Capability Preserved)
Actuator
43.3%
DPO
27.5%
TruthfulQA (Output Quality)
Actuator
56.9%
DPO
56.4%
UltraFeedback RL — Qwen2.5-3B

Plug and play

Actuator makes post-training easy. It drops right in to your existing stack and provides the unified end-to-end software layer you need to ship better models while skipping the pain.

import actuator
 
run = actuator.Controller(
    protect = {"safety": 0.95, "math": 0.70},
    on_drift = "adjust", # auto adapt pressure
    on_breach = "halt", # stop and rollback
    certify = True, # produce audit trail
)
 
run.start(model) # any framework
Illustrative integration

Get early access

Actuator is currently in closed testing with our early design partners. If your team is running serious post-training and want to do it a better way, please reach out! We're excited to hear about what your team is working on and open to potential pilots or partnerships.

[email protected]