skip to content
// engineering

Low, balanced, max

A profile is how hard a Midkernel Scan run tries. The workflow says how to audit. The profile says which models to use and how long each stage of the run may spend.

The same workflow on low and on max reads the same instructions. It reaches different depths. Choosing the profile is the main decision before a run. This note is the manual for that choice.

How credits work lives on Pricing. This page is the profiles: models, depth, time caps, and when to pick each one.

What a profile sets

Profiles set the default model, time budget and price.

The models. low uses a fast model (the kind that answers in seconds and is good at the obvious). balanced and max use stronger models. The exact model and version are recorded on the run.

The depth. More time and a stronger model allow a deeper review; they do not guarantee more findings. On the goal-security-review workflow, candidate findings must survive two independent judging passes at every profile; the profile decides how much hunting happens before the judges see anything.

The budget. A run has several stages, and each profile sets a default cap per stage: 30 minutes on low, an hour on balanced, two hours on max. You can adjust the timeout before starting. Reaching a hard cap can fail a stage; it does not guarantee a partial report.

same workflow text

profile

  • low
  • balanced
  • max

The workflow text does not change when you pick a profile. The profile changes the models, the time budget, and the price.

profilemodelsdefault time per stagecredits
lowa fast modelup to 30 min10
balanceda stronger modelup to 1 h25
maxthe strongest modelup to 2 h50

When to use which

low is a lower-cost starting point for triage and first looks. Its shorter default time budget limits how deeply it can investigate. No profile guarantees that it will find a particular vulnerability.

balanced is the default for the default workflow: a stronger model and more time per stage, and the right depth for a repository review you will read carefully. If you run one workflow on a repository once, run it here.

max is for the runs you sign your name to. A release audit. A repository you have already run on balanced and want to squeeze. It provides the largest default time budget for investigation. The price is on screen before you start; credit math is on Pricing.

Tweaking

Profiles are starting points, not tiers. The model and per-stage timeout can be overridden before a run, and the selected settings are recorded with the run. Finer-grained scope and pass controls are planned.

  1. 01start from a profile
  2. 02tweak it
  3. 03the run record stores the tweak

The record is the reproduction. A colleague reads the run, not a memory of which preset you clicked.

// related
September 28, 2026// research
September 28th Research Briefing

A weekly publication of the top three papers from arXiv last week. Papers with arXiv v1 published 2026-09-21 through 2026-09-28 UTC.

September 21, 2026// research
September 21st Research Briefing

A weekly publication of the top three papers from arXiv last week. Papers with arXiv v1 published 2026-09-14 through 2026-09-21 UTC.

September 14, 2026// research
September 14th Research Briefing

A weekly publication of the top three papers from arXiv last week. Papers with arXiv v1 published 2026-09-07 through 2026-09-13 UTC.

Research notes, at most monthly. No spam.
← All research