Added
- Deck generation verifies rendered slides before reporting completion. It
checks thumbnails, rendered titles and highlighted data claims, then reports
each slide's result. Artifacts with missing evidence remain drafts.
Addresses #1219. - Cost tiers can now pair a model or capability with reasoning effort in
providers.json, such as{"capability":"default","reasoning":"low"}.
Existing string targets still work. Tier effort applies only when that tier
selects the model; explicit model routes and effort overrides keep precedence.
Fixed
- Research verification no longer fails a synthesis for referring to its own
numbered points. A synthesis line such as "See point 2." or "Corrected in
point 2: ..." next to a workspace citation failed withnumber_mismatch,
because the cited file did not contain a standalone 2. A whole number after
point,points,itemoritems, or continuing such a list with a comma,
"and" or "or", is now skipped when no other word follows it. Numbers with a
unit, including formatted units such as5 **MiB**, and section numbers are
still checked. Compact lists such aspoints 2,3are also recognized, while
decimals and thousands-formatted numeric claims remain checked. Scores and
counts introduced by a colon, both endpoints of hyphen and Unicode
dash ranges, and quantities with either Unicode micro-unit prefix also remain checked.
Nested emphasis and whitespace before a percent sign cannot hide quantities.
Literal destination backticks and emphasis before attached units also retain
unsupported quantities for evidence checks.