Add demo video and link the blog and X posts in the README

Make the demo export image-only for faster transcription

Square off the rounded corners

Switch tab icons to the lighter skeuomorphic variant

Replace tab icons with transparent skeuomorphic set

Remove the "100% local" sidebar badge

Rewrite the synthetic-demo line; drop the judges reference

Optimize README for judges: tags, model links, Modal/Codex

Enlarge nav icons, icon the board, align top controls

Soften the demo privacy blur

Add BLUR_SENSITIVE_INFO_DEMO privacy blur toggle

Preload mascot art and hard-cache image assets

Slim the demo export and rewrite its chat around real data

Accent the user-facing next-action labels

Make the Q&A system prompt accent-correct PT-BR

Teach the Q&A model the agency quote/order/booking vocabulary

Straighten the conversa phone mockup

Update vision GGUF variant test to v1 fine-tune defaults

Update QA default-model test to v4 fine-tune

Move Clear button to topbar and kill nav-switch flicker

Add accents and cedillas to the PT-BR UI strings

Add Limpar/Clear button to start over

Fix identity demo holder name (matched transcription)

Log Stage-2 extraction + fix demo name leakage + demo toggle

Default back to fine-tunes; preload the transformers checkpoints

Log which model serves each inference, not just at load

Default vision + text to base models

Make QA GGUF env-overridable + fix stale preload list

Isolate uploads per browser session (no shared/persisted data)

Default to published v1 vision + v4 text models

Move text-gen files to finetune/text/; drop finetune tests

Merge branch 'vivamais-vision-v1' into vision-into-main

Consolidate vision finetune files under finetune/vision/

Force train GPU via decorator env, not with_options

Add Modal dataset publisher + refresh dataset card

Correct v1 eval with fixed metric + write v1 model card

Add v5 train config (restraint data, 2 epochs, no in-train eval)

Generate the synthetic dataset on Modal, not the local box

Teach restraint: drop optional fields from synthetic docs

Split Stage-2 eval into recall-on-populated vs over-fill

Add Stage-2 field diagnostic: split recall vs over-fill

Record v1 synthetic-only held-out eval: base vs v1

Make sharded eval robust to HEIC and per-shard failures

Reject weak Rio v4 travel teacher rows

Clear deterministic generation sampling flags

Record v3 dataset volume restore

Pin Rio v4 as standalone Modal module

Use A100 for Rio v4 text run

Add Rio v4 text safety recipe

Add v3 text dataset publisher