Redesign CTR calibration for marketer workflow — metric input upfront, not buried

Add optional metric calibration to web UI

Replace naive gradient descent with Newton's method for Platt scaling

Explore CTR calibration: map SGO scores to click-through rate predictions

Add downloadable report + switch default model to gpt-oss-120b

Add SSE keepalive pings every 15s to prevent proxy timeout disconnects

Hide badges until loaded — no more 'checking...' on first paint

Hide BYOK API key setup on HF Spaces — server key handles everything

Fix: use custom X-LLM headers instead of Authorization (HF proxy intercepts it)

Add realistic example templates: product, resume, pitch, policy, listing, blog

Two-model approach: 7B for setup calls, 72B for evaluation

Add 45s timeout + 2 retries on LLM client to prevent Space restarts

Show elapsed seconds on Auto-fill buttons while thinking

Create user 1000 with writable HOME for HF Spaces git config

Add wget to Dockerfile (required by HF Spaces dev tools)

Use /data/nemotron for persistent storage on HF Spaces

Fix port binding: start.sh kills zombie, add app logging

Fix port binding: simplify startup, remove timeout on segments, clean SIGTERM

Show 'Thinking...' on Auto-fill buttons and placeholders during LLM call

Fix port binding crash: kill zombie process and enable SO_REUSEADDR

Add elapsed timer, pulsing progress bar, timeouts with fallbacks for long LLM calls

Remove session restore — was only for debugging

Fix startup: use lifespan instead of deprecated on_event

Set rate limit to 2 runs per hour

Rate limit by pipeline runs (10/hr), not individual LLM calls

Security hardening: rate limiting, no creds in URLs, path traversal fix

Pass LLM credentials per-request, never store server-side

Improve persona setup UX: hide folder on Spaces, add progress bar, better errors

Clarify API key is never stored or logged

Show base URL and model fields by default — don't lock to any provider

Add HuggingFace Spaces deployment support

UX copy overhaul: intuitive labels, consistent terminology, plain language

Fix XSS, VJP contract, binding, session security, name collisions, and error handling

Fix production issues: session overwrite, error handling, memory leaks

Remove redundant text output — table with expandable rows is the interface

Make gradient table rows expandable with per-evaluator details

Remove personal/dating examples from all placeholders and prompts

Show change label instead of ID in counterfactual progress log

Fix 422 on counterfactual stream — remove **kwargs from endpoint

Move counterfactual params from GET query to POST+ticket to avoid log leakage

Fix missing compute_goal_weights import

Add Suggest buttons for goal and audience instead of black-box auto-fill

Make entity, goal, and audience first-class inputs

Extract demographic filters from audience context for Nemotron cohorts

Restore saved sessions from disk on server startup

Auto-generate changes from concerns, one-click directions flow

Reframe analysis around persuadable middle, not universal appeal

Add goal-weighted gradient (VJP) — optimize toward objectives, not universal appeal

Hint larger panel size when below 40

Fix panel size overflow — hard cap Nemotron cohort to requested total