model is flan-t5-small

model is flan-t5-small

model is flan-t5-small

model is flan-t5-small

move to gradio sdk

move to gradio sdk

move to gradio sdk

wip

wip

wip

wip

wip

wip

wip

wip

wip

wip

Enforce fixed sequence lengths to match warm-up shapes

wip

wip

Reduce padding sizes from 512x256 to 256x128

wip

Use fixed-size padding for kernel reuse across all requests

wip

Add error handling to per-request warm-up

Add per-request warm-up to eliminate first-sample kernel compilation

Promote warm-up and per-sample timing to INFO level logging

Add detailed timing logs to score_posterior_sample; warm up with multiple shapes

Filter /health requests from access logs

Fix cold-start latency: realistic warm-up shapes and vectorized sampling

Reduce log noise from health checks and keep-alive pings

Transform FRONTEND_URL to hf.space/health endpoint and use GET requests

Add info-level log for each keep-alive ping

Fix keep-alive to ping frontend root URL with HEAD request

Add frontend keep-alive pinging

Fix float32 underflow in consistency score; recalibrate quantile boundaries

compute consistency

Add consistency score to quantify source/summary alignment

Add warm-up forward pass to both scoring backends on init

Remove debug API_TOKEN log, add bearer token auth via X-API-Token header

Log API_TOKEN value at startup for debugging

Switch auth to X-API-Token header; HF Spaces strips Authorization

Temporarily disable auth to confirm service works end-to-end

Add auth diagnostic logging to debug 401 mismatch

Simplify bearer token auth: check Authorization header directly in endpoint

Fix bearer token auth: use closure so FastAPI resolves Security() correctly

Add optional bearer token auth to /score endpoint

Include ambiguity_boundaries in normalization block of /score response

Add ambiguity_score to API response with dedicated quantile normalizer

Load scoring service as background task so /is-ready returns false during init