Custos builds a model bake-off into the review grid

You can now re-run any cell of a document review on a different AI model and keep both answers side by side.

workflowanalytics

Custos's fork uses a spreadsheet-style grid to pull answers out of documents - one row per document, one column per question. The new work gives every cell its own version history. Each time the AI rewrites a cell, the previous answer is archived rather than discarded, along with the exact model and instructions that produced it.

A new history window lets you browse those archived answers and line up "before" and "after" in the same view. The headline addition: a toolbar to re-run a single cell on a different model - say, one vendor's model versus another's - without losing the earlier result. You can also re-run a hand-picked set of rows instead of rebuilding the whole review. In effect, it's a hands-on tool for judging which model gives the better answer, baked straight into the review screen.

So what Anyone choosing or defending which AI model does their document review now has a built-in way to compare them on real cells, not in the abstract.

View this fork on GitHub →

Spotted something wrong? Or know the PR text has fresher detail than the writeup above?

Commits in this thread

3 commits from Custos/legalos, oldest first. Source extracted verbatim from the harvested git log.

SHA	Subject	Author	Date
`ce6e9c27`	Tabular: run on selected rows + version history with model & prompts	Custos	2026-05-04	↗ GitHub
commit body - New tabular_cell_versions table archives prior cell state (content, model, system_prompt, column_prompt) on every overwrite. - tabular_cells gains model/system_prompt/column_prompt/updated_at columns capturing what produced the current content. - POST /tabular-review/:id/generate accepts optional document_ids; when provided, those rows force-rerun and prior values are archived. - POST /tabular-review/:id/regenerate-cell archives prior content first. - New GET /tabular-review/:id/cells/:cellId/versions returns history. - Frontend: 'Run selected (N)' button appears in the header when rows are checked, posts document_ids to the streaming endpoint. - .env.example: add XAI_API_KEY and R2_REGION.
`c2471472`	Tabular: cell history modal - view archived versions, models, prompts	Custos	2026-05-04	↗ GitHub
`1f3d4cb9`	Multi-model comparison from cell history modal	Custos	2026-05-04	↗ GitHub
commit body - /regenerate-cell now accepts an optional model param that overrides the user's tabular_model for that one run; the new content is saved with that model id, prior content archived as usual. - regenerateTabularCell(reviewId, docId, colIdx, model?) on the client. - Cell history modal: a 'Run on ...' row of model pills in the header lets you re-run the same cell against any other supported model (Claude, Gemini, Grok). Each run appears as a new history entry, so you can click between versions to compare. The current model is disabled.

Capture this thread into my fork

Download a single Markdown prompt that tells Claude how to port every commit above into your working tree — adapting paths and structure to match your repo. Run it via claude -p < capture-thread-47.md from inside the repo you want the changes in.

⬇ Download capture-thread-47.md