Review & Fix
Parsing isn't the end — every inference is yours to check, trace, and undo.
Overview · Enter Smart Review
After parsing you land in Smart Review, with four things keeping you in control:
AI proactively flags doubtful spots, e.g. "Is FBS 10% standard or high?"
Left is the AI inference, right is the source passage
Experiment name / sub-experiment / step / parameter inline edits take effect instantly
A "Review changes" dialog lists every add / delete / edit you made
Every change is recorded in history; open Sub-experiment → History to see the full change chain and undo in one click.
Edit the parse on the review page + review your changes
After parsing, the review page is now fully editable — experiment name, sub-experiments, every step (temperature / volume / reagents / code, etc.), every parameter; you can also add / delete sub-experiments, steps, and parameters. No need to create the experiment first and fix things afterward.
- Before [Confirm Create], required fields / duplicate parameter names are caught and highlighted on the spot
- If you changed anything, a "Review your changes" dialog lists exactly what you changed vs. the AI parse (old → new, added, removed) before it lands in the database
8 finding types
The AI self-checks and flags each one, with these code identifiers:
missing_param— should fill but didn't (concentration / temperature / duration)provenance_low— source unreliable (inference rather than verbatim)unit_mismatch— inconsistent unitsvague_step— too vague ("some / appropriate")unusual_value— outlier values (too high / too short)missing_control— no control / baselinecontradicts_methods— contradicts Methods sectionspec_format— field format off (should be number not text)
4 provenance levels
EXPLICIT— From the source: the paper states it directlyPARTIAL— Partly in source: the paper mentions it but lacks detailINFERRED— AI inferred: filled in from context, with a citationMISSING— Missing from source: not in the paper, you must fill it in
On the review page, each step shows its source tag inline whenever it isn't "From the source" (Partly in source / AI inferred / Missing from source), so you can spot what needs checking at a glance.
The top of the page summarizes how many items need a look, and "Next item to check" jumps through them one by one. Steps that came straight from the source are folded by default; expand them when you need to. Each label type gets a short one-time hint the first time it appears. "Looks fine" is only a check-off on this page to help you keep your place, and it doesn't change any data.
INFERREDfield to see AI's cited paper section + "similar studies" (how others used it). You judge: accept / adjust / skip.