Sign Up Sign In

Docs

Proven feature packs for a subject model, each shipped row with an evidence report.

What you get

  • Shelf packs: catalogs we train and test on public checkpoints, when imported
  • Custom packs (quoted): the same lab on a checkpoint you grant us, any size
  • A proof report on each shipped feature
  • Optional: try a shipped feature on your own prompt
  • HTTP API for the catalog, try-on-prompt, and reports

Quickstart

  1. Create an account and open your workspace.
  2. Open Features. The proven pack appears when the lab export is imported; until then you may see UI fixtures.
  3. Open Reports for the tests on shipped features.
  4. Optionally open Steer to try a shipped feature on a prompt. That run is not the proof.

How we build packs

Features are trained and tested before they appear in the catalog. We use sparse dictionaries (SAEs, transcoders, and crosscoders) on a subject checkpoint. Only features that pass causal tests ship.

API

Authenticate with a session cookie or an API key from your profile (Authorization: Api-Key <key>). Team-scoped under /a/<team_slug>/api/:

  • GET /a/<team>/api/pack/ — catalog (no direction vectors)
  • POST /a/<team>/api/steer/ — try a feature: {"feature_id", "alpha", "prompt"}
  • GET /a/<team>/api/reports/<id>/ — evidence report

OpenAPI: /api/schema/swagger-ui/. Plan limits: see Pricing. 

curl -s -H "Authorization: Api-Key YOUR_KEY" \
  https://zytheris.ca/a/YOUR_TEAM/api/pack/

What counts as proof

A label is not proof. A feature ships as proven only if intervening on it caused the hypothesized effect under the tests we ran. Trying it on a new prompt in the app does not replace that report.

Custom packs

Weights do not go through the website. Request a custom pack with a public Hugging Face ID, a gated repo, or private access under an agreement. Model size is not capped; we provision GPUs. We pull the checkpoint in the lab, train and test offline, then import the export into your team workspace. Do not email weight files.

We do not guarantee that any feature will pass. If none do, the workspace still gets the evidence report (a true negative). The quote includes one re-run on an agreed scope change. Thresholds are not lowered after the fact to invent a catalog.

Limits

  • You cannot upload model weights through the site.
  • Features are tested before they appear in the catalog.
  • Optional prompt trials run on private GPUs.

More

Packs · Request a custom pack · Pricing · Security · Contact · Get started