Eval: xlsx evals use soft markers that could match incidentally

[olaservo]

Problem

The xlsx eval tasks check for generic terms as expected output markers:

Task	Marker	Risk
xlsx-openpyxl	`"openpyxl"`	Common Python library name
xlsx-formulas	`"formula"`	Generic Excel term
xlsx-financial	`"blue"`	Common color word
xlsx-verify	`"recalc.py"`	Slightly more specific but still a tool name

Compare with synthetic skills that use unambiguous markers:

• `SKILLJACK_GREETING_SUCCESS`
• `SKILLJACK_CODE_FORMATTED`
• `SKILLJACK_TEMPLATE_LOADED`

A model could mention "openpyxl", "formula", or "blue" in a response about Excel without having actually followed the xlsx skill instructions. The word "blue" in particular is highly likely to appear incidentally.

Suggestion

Since the xlsx skill is a real production skill (not synthetic), injecting artificial markers isn't ideal. Options:

1. Use more specific compound markers — e.g., check for `"blue text"` AND `"openpyxl"` together
2. Use regex patterns — the EvalConfig already supports RegExp
3. Accept the trade-off and document that xlsx evals have weaker assertion guarantees than synthetic skill evals

Files

• `evals/tasks/xlsx-*.json` (4 files)
• `evals/lib/eval-checker.ts` (already supports RegExp in expectedOutput)