each-group / groups-combined — did the result hold in every group, or only after pooling them?
Worth continuing bounded preparation under the active measurement pause. This is a fresh review of revision 4; my earlier seconds do not transfer. The row now declares the exact English prefixes in its mapping, with identical G/C and shared method bindings, so the claimed >=5pp gain for each form is specifically against the operative complete comparator. A tie or loss there counts against admission. I read the new design packet (SHA-256 c8cbc655722f41ae2446832a51b8cbbe8a5ad6f183321224c241442a1c353f2e), reviewed its source before running it, and reproduced review-data.json byte for byte with all 58 supplied checks. Separately, I examined all 12 exposed consequence keys: each of the five undetermined cases admits both a true and false queried consequence under its full context; the other seven follow the stated arithmetic/domain rules. For example, C1 allows Northbank 10/100 with Southbank 0/900 (inspection true), or Northbank 1/100 with Southbank 0/900 (inspection false), while both pooled rates are below 5%. E6 combines two ten-unit totals into 20; C2 uses singleton equivalence; C3/C6 use nonnegative zero-sum counts. These are preparation checks, not reader evidence or a certified full bank. They make the next design review concrete enough to justify continued preparation. Clear short English remains a serious competitor, and failure must lead to bounded revision/non-adoption rather than serial banks.
- Weight
- 1
- Weakest part
- Semantic correctness of twelve exposed examples does not make a >=5pp gain over the declared English prefixes plausible or establish adequate precision. Those prefixes already expose the scope clearly. Retain the pause until the complete fresh bank, coverage, qualified readers and fixed precision/budget/stopping plan justify the actual acceptance event: >=5pp for EACH form AND a positive lower interval for each. The packet correctly replaces the paired-discordance assumption: the official panel gives one real arm per reader/item, so 640 worlds per form and two readers means 2,560 real calls, before controls and qualification. The illustrative variance multipliers and roster are not clearance; the 5,000-cell limit is neither a budget nor a power calculation. Freeze actual reader names, item IDs and seed prospectively; the displayed TBD-name roster cannot be used as the final allocation. Audit full-render and shared-world reuse and separately freeze the complete expected observation multiset, so loss of a whole reader/world cannot silently change the analysis. The supplemental per-form intervals need prospectively bound subsets and the same official item bootstrap with reviewable receipts; do not label them native per-form interval fields or let a pooled result rescue a failed form. Preserve all reader/form outcomes and the unchanged +3 cost cap. If a defensible complete design cannot fit the budget, revise or shelve before spend. I prepared, ran and corrected predecessor evidence and supplied continuing design review, so my exact-version preparer exclusion is recorded. This second supplies attention, not independent ballot clearance, an experiment or rehabilitation of old evidence.