tofu, keto, Rhubarb: The Subchunk Refinement Signals

01What are tofu, keto, and Rhubarb?

tofu, keto and Rhubarb are codenames for refinement signals. They fine-tune the quality estimate at a very granular level: individual subsections (subchunks) of a page. Their exact function is not fully documented in the available material. That makes them the weakest evidence in the entire matrix.

The names follow Google's food-based naming convention: tofu (soy), keto (dietary), Rhubarb (vegetable). Google uses this convention for internal systems that are not user-facing. The names are internal identifiers, not descriptive labels.

IS-Calibration (System 12) feeds tofu, keto and Rhubarb. Human rater judgments therefore calibrate them, the same ground truth that calibrates chard and OriginalContentScore. The connection to IS-Calibration is the strongest evidence for these systems: they are part of the quality calibration chain, even if their specific function is unclear.

02Subchunk-level refinement

The concept of 'subchunks' is architecturally significant. Google does not evaluate a page as a single unit. It divides the page into subsections and evaluates each one independently. A long article might have an introduction, three main sections and a conclusion. Each is a subchunk.

Subchunk-level evaluation means that a page with one excellent section and three mediocre sections is scored differently than a page with four good sections. The subchunk scores aggregate into the page-level score. That score feeds into NSR (System 7) for site-level aggregation.

The Subchunk-Deltas field measures the difference between subchunk scores. High variance between subchunks (one excellent, three poor) might signal a page that was partially generated or partially copied: a human wrote the excellent subchunk, and the poor ones were generated.

03tofu/keto in the ranking architecture

These signals are indirect for Expertise (Exp: indirect) and Trust (T: indirect). The indirect Expertise signal arises because subchunk-level refinement requires a model that can judge quality at section level. That is a more sophisticated evaluation than whole-page scoring.

The indirect Trust signal comes from the connection to IS-Calibration. If raters trust a page, the subchunk signals will reflect that trust at the section level. A section that raters rated as untrustworthy (e.g., containing unverified claims) will receive a lower subchunk score.

The weakest source in the matrix is here. NavBoost ([A] DOJ, [P] patent), Panda ([P] patent, [A] DOJ) and chard ([B] leak, [O] QRG) each have several sources. The tofu/keto/Rhubarb signals have only [B] (existence of codenames) and [C] (function inference). There is no patent, no DOJ testimony and no official communication.

04References