Crisis turn: card-styled sidebar entries + scannable chat markdown

Typo-aware safety pattern matching for crisis / imminent-risk regex

Crisis typo-tolerance + harm-frame echo verifier

Add substance_use + privacy_confidentiality routes, affirm-with-question detection

Session-state audit fix: thread session_id end-to-end, advance substantive-only, use monotonic seq counter

Robust natural-language affirm/negate/unsure detection + advance-after-PERMISSION

Remove 'Inspect -> Support card' UI leak from OFFER chat prose

Vary OFFER vs LISTEN openers on general_student_support; smooth consent question embed

Polish consent_acknowledged + PERMISSION hint phrasing

Fix conversational planner: advisor over-trigger, yes-after-offer flow, crisis + misconduct pattern gaps

Diversity probe sweep + V1->V4 narrative README + HF Spaces entry + DV/incomplete fixes

Response quality: greeting/goodbye/meta handlers, slang routing, lexical variety, last-verified badge, markdown norm, crisis variants

V4.4 Tier 1: regex hardening, clarify-stage enum, stat caveats, PDF unicode, retry/backoff, reproducibility doc

V4.3: prompt-injection audit, input length cap, per-layer ablation, unguarded baseline, limitations restored

V4.2 part 3: sycophancy guard, F-1 flag decay, incomplete-message handler

V4.2 part 2: authority-misconduct route + conversation history in rephraser

V4.2 part 1: fix broken ISSS URL, ship URL audit + Karthik data brief

V4.1: fix repetition/scroll/topbar, add PDF export, counselor-assist note

V4: streaming, controlled paraphrasing, support plan, voice, sweeps

Add LLM plan-and-rephrase layer with Groq + Anthropic providers

Add stage-aware listening planner and F-1 student awareness

Improve conversational MVP routing

Polish MVP support experience

Add Eval B safety supplement

Ingest Core dataset and harden router policy

Prepare Core dataset intake and resource registry

Polish peer helper and scope handling

Add Core safety metadata and eval summaries

Implement EmpathRAG Core hybrid router

Add Karthik eval harness and safety patches

Start V2.5 support navigator hardening

Checkpoint V2 curated support navigator

Add curated corpus integration scaffold

Start v2 safety hardening

Add sliding window conversation memory - 3 turns, n_ctx 4096

Tighten SYSTEM_PROMPT with few-shot example, reduce max_tokens to 200, fix paragraph post-processor

Improve SYSTEM_PROMPT for conversational peer support - validate, reflect, one question

fix: guardrail dual-import path, bertscore key names, ragas reuse pipeline.llm (Day 14)

feat: real DeBERTa guardrail wired, skip_ig flag, smoke test updated

feat: wire real DeBERTa guardrail, fix smoke test for crisis intercepts

fix: wire real guardrail, anchor gitignore, fix dtype warning

fix: guardrail_ig.py — IG in embedding space, token_type_ids, correct dtype

fix: scope models/ gitignore to root only, add src/models package, remove .claude

Add pipeline orchestrator + smoke test — 4/5 emotion predictions correct (Day 12)

Day 2: NLI pairs built, fix gitignore for large files

Day 1: data pipeline, session tracker, query router, adversarial probes, Colab training notebooks