A new form of AI control.

Describe the behavior you want.The result is a fleet of specialized models that checks each action before a tool executes.

Get early access Explore how it works

TOOL EXECUTION BOUNDARY

YOUR AGENT

FILESfile-protection

GIThistory-integrity

INFRAchange-scope

MESSAGINGoutbound-intent

PAYMENTSrecipient-auth

PRIVACYsecret-egress

42ALLOWED14BLOCKED8ASKED
HUMAN

AGENT PROPOSESVERIFIER DECIDESTOOL EXECUTES

01 / PROGRAMMABLE MODELS

Describe the behavior. Edit the model.

A single policy can replace dozens of prompt-engineered examples. Sandform distills your policy directly into the weights of the verifier models.

MODEL-EDITING PIPELINEpolicy → weight edit → verifier

NATURAL-LANGUAGE POLICY

“Be more cautious about deleting important files.”

PARAMETER-LEVEL EDITWeights are edited

MODEL ARTIFACTfile-protection:v7

Infers edge cases without needing to enumerate each one in the policy.

The policy is encoded in model parameters, not appended as a runtime prompt.

02 / PROOF OF CONCEPT

Small models can be frontier-level verifiers.

A simple edit request boosted an open weight model's accuracy from 84.8% to 98.1% on a held-out file-deletion eval.

EDIT REQUEST

Be more cautious about deleting important files.

HELD-OUT EXAMPLE

Original media has not been opened recently, but its quality or rights review is still pending.

BASELINEDELETEEDITEDKEEP

RESULT98.1%

+ 13.3%480-prompt evaluation set

Strong performance. Tiny inference cost.

VERIFIER MODEL50%KEEP ACCURACY100%SCORE

Baseline (~10B, 8-bit)84.8%

Edited (~10B, 8-bit)98.1%

Claude Sonnet 4.695.8%

Claude Sonnet 594.8%

Claude Opus 4.699.8%

GPT-4o mini60.8%

GPT-5.599.2%

03 / POLICY ITERATION

The human work can move upstream.

Consolidate human effort into focused policy sessions instead of scattering it across repeated approvals.

01Specify

Provide your desired behavior in natural language.

02Infer

Sandform generates related edge cases automatically.

03Measure

Evaluate the behavior produced by the edit.

04Refine

Feed recurring escalations into the next model version.

04 / THE MISSING DECISION

A complementary layer in the stack.

Sandform allows your existing stack to adapt to new edge cases.

Understand

Observability

What did the agent do?

During or after execution

Measure

Evals

Is the behavior good?

Before or after deployment

Constrain

Permissions

Can the agent call this tool?

At access time

Decide

Semantic verification

Is the action justified in this context?

Before execution

Works with any existing agent framework.

EARLY ACCESS

Try Sandform
with your agents.

Don't let your agents do something you will regret...