v0.27.2: Fix report/committed output reference count mismatch, CRCS full form

Flowchart v5: cleaner layout, baseline→judge, correct threshold

Flowchart: fix KG rejection path, baseline arrow, Patnaik year

v0.27.0: Evidence quality overhaul from full audit findings

Cache-bust: add ?v=0.26.4 to app.js import to force reload

Expand stopwords for OpenAlex query cleaning

v0.26.4: Full audit tool, dedupe web sources by domain

v0.26.3: Fix m* label (was always m1), L5 candidate table, OpenAlex queries

v0.26.2: Fix m* label bug, L3A/L5P empty cards, refinement Changes column

v0.26.1: Fix L3A empty, L5R truncation, move formula to Refinement Summary

v0.26.0: Fix report, streamline UI, add tests

v0.25.2: Fix L5R empty card, convergence messaging, refinement count

v0.25.1: Rewrite flowchart, fix provenance depWeights, deep audit doc

v0.25.0: Align with HTML artifact — t=0 counting, Phase 2 tracking, post-stability L1/S1

v0.24.2: Dedupe references, fix Phase 2 score math, label breakdown terms

v0.24.1: Fix nil-evidence false positive stripping relevant citations

v0.24.0: Merge OpenAlex+Gemini evidence, score breakdown, references block

Auto-retry transient Gemini errors (503, 429, 500, 502, 504) with backoff

Structured citation document model + 76 new tests

Fix L3 relevance, evidence scoring honesty, citation finalisation, safety + baseline prompts

Harden citation finalizer and live audit coverage

Fix plain References citation bug and bump to v0.22.4

Add organized audit test suite and bump to v0.22.3

Harden citation finalization and bump to v0.22.2

v0.22.1: Upgrade Polish & Judge prompts so CRCS wins more often

v0.22.0: Fix phantom citations, remove baseline card, organize prompts

Fix: converged → phase2_converged (undefined name error after L5P)

Bump version to v0.21.2 for build verification

v0.21.1: Decouple pipeline from SSE stream to fix HF Spaces timeout

v0.21.1: Fix SSE timeout + add diagnostic logging

Fix heartbeat_wrap cancelling in-flight generator on timeout

Fix SSE network error on HF Spaces by adding heartbeat keepalive

v0.21.0: Architecture matches Companion Paper — single Phase 1 + iterative Phase 2

v0.20.0: Phase 2 lattice protocol fully implemented per Companion Paper

v0.19.0: Codex review fixes — XSS, reject gate, state reset, audit scores

v0.18.0: Judge A/B labels, citation renumbering, no review duplication

v0.17.0: prompts.py is single source of truth — all inline prompts removed

v0.16.0: Centralized prompts.py + Prompts tab

Make all prompts domain-generic (not medical-specific)

Spread citations throughout, baseline covers 4 judge dimensions

Fix E-score: use evidence_text for scoring, align summary table

Fix E-score calibration: start from 0.6, floor at 0.4

Fix version display to v0.15

v0.15.0: CRCS answer covers all 4 judge criteria, awesome judge

v0.14.0: E-score fix, blind judge, KG gate fix, flowchart tab

Fix color highlighting in judge answer panels

v0.13.0: Fix KG false rejection, balance answer lengths

v0.12.0: Full model names, judge answers, review fix, version in UI

Model pills, m* labels, judge collapsible, color-coded answers

Real references, better refinement table, judge with answer preview