Two things break a performance line: an **opaque formula** and an **unstable answer**. Ask the same question four times, get four numbers, and nobody dares to take any of them into a review conversation. Here is how it runs after handover — the model understands the question and phrases the answer; the engine does the arithmetic.
Three weighted scores become a composite. The composite is ranked, grade quotas are set from the period's business completion rate, grades are forced down the ranking, and departments adjust within their quota. Three manual steps, redone monthly — and a different person gets a different result.
The sheet is one row per person: the current month has three scores and a total, and each earlier month is a single grade column. Answering 'how did this person move from January to May' or 'who has two consecutive D's' means scanning rows, and with enough people something gets missed.
Let a model do the arithmetic inside its prose and the same question returns different figures on different runs. It reads a hire date as the raw spreadsheet serial and repeats it. Ask how many people got a D and it gives a percentage, not names. None of that can go into a review.
Scoring items, weights, the three-step distribution rule, the org hierarchy and the Excel column mapping are issued to the desktop app as company configuration. Until the weights are confirmed, the composite is taken from the sheet's own total column rather than re-weighted.
Review sheets rarely have an API, so the source is the Excel file uploaded in the conversation. It is captured as a snapshot with its column definitions. On a later question with no new upload, the most recent matching snapshot is claimed; if there is none, it asks you to upload rather than going looking for a substitute.
Per-person multi-month detail, a full roster for a group or department, threshold filters like 'at least two D's', and the list of people matching a given trend are four query types the engine computes under one set of definitions. The answer is identical no matter how many times you ask.
Consistently improving means every month is strictly better than the one before. Trending up means the end of the period is better than the start, with fluctuation allowed in between. The business treats these as different questions, so the engine returns a separate list for each.
A department head sees the department, a team lead the team, a group lead the group. People who have left or are not in the review are excluded automatically and never appear in a distribution or a list.
Items, weights and grading rules are configuration issued per company, not code. They are calibrated once at delivery; a later change to the weights is a configuration change, with no app update.
Because these numbers go into review conversations. Arithmetic done inside a model's prose returns different answers to the same question and cannot be audited. The engine computes under fixed definitions, the model only phrases the result, and every figure traces back to the sheet.
Most companies run the monthly review in Excel with no API, so this line works from the uploaded sheet. Tables in Feishu Base and Feishu Docs can be read as well.
No. Visibility is trimmed by role as a hard constraint, not a prompt. The company-wide view is available to HR and authorised management only.