The Celeborn Model Mixer

Disclosure · Version 2026-07-11 · celeborncode.ai

1. What it is

The Model Mixer routes each planning session and each card to whichever AI model gives the best result for the least cost, then escalates to a stronger model only when the work needs it. Because a cheaper model handles the work a cheaper model can handle, the blended tier costs less than paying frontier prices for everything. That lower price is the routing discount — we do not charge you extra for auto-routing, we charge you less.

We show you the model that actually ran each card, on the board and in the CLI. Nothing about which model did your work is hidden from you.

2. The providers your work may run on

3. What we log — features, never content

To route well and to keep the blend honest, Celeborn records one row per routing episode. Every row is features only: salted hashes and numbers that describe the shape and outcome of the work, never the work itself. We can show you a row — there is nothing in it to leak.

  1. Salted identifiersA per-install salted hash of the project and the card id. Not the names — the hashes are not reversible and not cross-linkable between installs.
  2. Card shapeClass (P/E/C/S), title/notes LENGTHS (not the text), tag list, spine membership, blocked-by/blocks counts, engine floor, whether a Stop condition is set.
  3. LanguagesDetected from touched-file extensions only (e.g. “python”, “tsx”). The file paths themselves are discarded and never stored.
  4. Routing decisionWhich model/rung ran the card, which policy version and rule decided it. This is what makes every routing choice auditable.
  5. Execution meteringTurns, tool calls, token counts, wall-clock, and computed cost — the numbers that let us price the blend.
  6. OutcomeWhether the Stop condition was reached, whether tests passed, retries, whether it escalated, and whether it was reopened within 7 days (the honesty checks).

What a row never contains:

4. Retention

Rows are stored under your own account and are visible only to you — access is gated per project by the same ownership rules as the rest of your hosted board. We retain them to operate and improve the routing algorithm. Because the rows are features-only and salted per install, they cannot be tied back to your source, your prompts, or another install of Celeborn. Delete your account and the associated rows are removed with it.

5. Data residency & the frontier-only tier

The cheap leg uses Zhipu / Z.ai’s GLM on a metered commercial API. If your organization cannot route any work to that provider, choose the frontier-only tier: routing degenerates to the frontier rung everywhere, so nothing runs on the cheap leg. You keep every capability; you pay the frontier price instead of the blended one. This is a single setting, not a different product — celeborn mixer tier frontier-only (and celeborn mixer tier blended to switch back). Unlike opting out of telemetry, the tier setting leaves your outcome loop intact: you keep contributing rows, you just never route to the cheap leg.

6. The routing algorithm is proprietary

We disclose the providers, exactly what we log, and how we price the blend. We do not disclose the routing weights — which signals push a card up or down the ladder, and by how much. That model is Celeborn’s, learned from the outcome rows above, and it is what the blended tier’s margin is built on.

7. Consent & opt-out

Telemetry sync is on by default and funds the blended tier. You can opt out at any time by running celeborn consent and disabling the Model-Mixer telemetry item. Opting out stops the upload at the source — no row ever leaves your machine — and flattens you to frontier-only routing (the discount tier is funded by the outcome loop, so it turns off with it). You lose nothing but the discount tier. If you want frontier-only routing but wish to keep contributing telemetry, use celeborn mixer tier frontier-only instead (§5). See also the User Agreement.