Clarify AI sidebar and fix empty Jev evidence reviews

.impeccable/config.json

@@ -2,16 +2,6 @@
   "detector": {
     "ignoreRules": [],
     "ignoreFiles": [],
-    "ignoreValues": [
-      {
-        "rule": "side-tab",
-        "value": "*",
-        "files": [
-          "viewer/assets/radsysx-viewer.css"
-        ],
-        "createdAt": "2026-09-23T02:43:10.069Z",
-        "reason": "Assistant design review: the pre-existing 2px muted #5c7c91 rule groups nested Jev claim judgments and execution receipts; it is not a thick decorative card accent. Verified in the rendered 280px sidebar and git blame."
-      }
-    ]
+    "ignoreValues": []
   }
 }

AGENTS.md

@@ -285,7 +285,7 @@ Last updated: 2026-09-22
 - Favor rigorous, beautiful, professional solutions with high signal-to-noise.
 - Prefer Linux-native commands and paths.
 - Record durable behavior changes in this file or the nearest relevant child `AGENTS.md`.
-- `.impeccable/config.json` records reasoned, file-scoped design-detector exceptions. The muted 2 px Jev judgment separator is intentional grouping, not a decorative card accent; its `side-tab` exception is limited to `viewer/assets/radsysx-viewer.css`.
+- Keep sidebar information separated into Chat, Research and Jev review views. Use a quiet reading-room palette and progressive disclosure; technical receipts and full abstracts stay collapsed by default. Jev judgment groups use subtle horizontal separators; the old side-accent exception is removed.
 
 ## Child DOX Index
 
@@ -339,7 +339,7 @@ The standalone [evidence-review runbook](backend/evidence_review/README.md) docu
 - See `roadmap/ai-backend/NIM_IMPLEMENTATION.md` for configuration, commands and dated acceptance. Keep provider keys separate and never silently fall back when the chosen model fails.
 
 - The sidebar **Settings → Research models** saves Gemini/NVIDIA NIM provider and exact model per signed account. Environment supplies the default until a saved choice exists. The dropdown includes the entire hosted NVIDIA catalog; catalog inclusion does not verify tool support or entitlement. Saving closes that account's sessions/jobs and requires reconnection; live voice selection and offline evidence CLI model flags remain separate.
-- Keep the AI sidebar compact: one header row, inline data confirmation/connect, and media controls visible only during an active connection. Data attestation, transmission disclosure and separate microphone/image consent remain required.
+- Keep the AI sidebar compact: show the text/research model near the header and separate Chat, Research and Jev review views. Optional voice setup is collapsed within Chat; data attestation and disclosure live by the composer. Keep media state visible while connected and preserve separate microphone/image consent.
 
 - Sidebar Jev reviews run through backend-owned `/api/ai/sidebar/*/evidence-reviews` contracts in research/pilot only, using backend-only `RADSYSX_TYPESAFE_AI_API_KEY`. They retain original answers, require independent text confirmation, and persist POSIX-private artifacts in `.ai-evidence/` beside the actual database (or absolute `RADSYSX_AI_EVIDENCE_DIR`). Ending voice preserves review; explicit cancel/account changes/logout stop it; source-history deletion cancels jobs and removes their artifacts. Clinical use and the qualified human evidence-quality study remain unapproved/pending.
 

CLAUDE.md

@@ -86,3 +86,5 @@ NVIDIA NIM is available for explicit evidence evaluation and opt-in PubMed resea
 Typed **Send** and explicit **Research** work without Gemini Live/OpenAI Realtime. Confirm synthetic/deidentified content and choose the standard model in **Settings → Text & research models**. Only the question, bounded text history (chat only) and neutral viewer metadata are sent; image pixels require separate live sharing. **Connect voice** starts a separate voice conversation. See `roadmap/ai-backend/DESKTOP_AI_ACTIVATION.md`.
 
 ChatGPT/Codex subscription sign-in is available under desktop AI Settings for typed chat and public PubMed research. It uses isolated backend-owned Codex App Server and the OS keyring, not OpenAI API credentials or Realtime entitlement. Read `roadmap/ai-backend/CODEX_SUBSCRIPTION.md`; never copy the user's existing Codex auth.
+
+The AI sidebar separates Chat, Research and Jev review. Optional voice setup stays in Chat; the review workspace hides the composer and collapses abstracts/technical receipts. Sidebar evidence uses versioned intact cited passages and explicitly reports empty previews as unavailable. See `roadmap/ai-backend/JEV_SIDEBAR_IMPLEMENTATION.md` for the corrected workflow and live acceptance.

README.md

@@ -285,15 +285,15 @@ The OHIF sidebar offers `gemini-3.8-live-extended-thinking` and `gpt-realtime-2.
 
 1. Run `npm run desktop` and open **Settings → API keys** in the **RadSysX AI** panel. Enter and save your own Gemini key, OpenAI key, or both. Saving refreshes provider availability without restarting; the first connection verifies provider access. Keys are sent only to the local backend for encrypted storage, and are never displayed again or saved in browser storage/history. Replacing or removing a key ends your active assistant sessions and tasks.
 2. Open a synthetic/deidentified study. The key settings distinguish **your saved key**, an **app-configured key**, and **not configured**. Removing a saved key explicitly returns that provider to an app-configured key when one exists.
-3. Select the provider, confirm the displayed content is synthetic or deidentified, then connect. If the initial catalog loaded before local sign-in, use **Retry assistant setup** first. This records your declaration, not an automated deidentification result. Changing provider ends the old session and requires fresh confirmation.
+3. In **Chat → Voice**, select the provider, confirm the displayed content is synthetic or deidentified, then connect. If the initial catalog loaded before local sign-in, use **Retry setup** first. This records your declaration, not an automated deidentification result. Changing provider ends the old session and requires fresh confirmation.
 4. After **Connected**, enable the microphone or send text to the voice provider. Independent text chat and research are described below. To discuss what is visible, click **Share active image** and wait for **image sharing on · image sent**. The panel names the active image; capture covers that selected viewport, at up to one JPEG frame per second. The whole app, sidebar and other windows are outside this capture.
 5. Ask for a viewer action, image explanation, or public research. Review a proposed durable report save or deletion before applying it. Local-only DICOM must be imported/associated through the worklist before saving a report.
 
 **Text without voice:** confirm synthetic/deidentified data, type your question, then choose **Send** for a discussion or **Research** for an explicit literature search. No Gemini Live or OpenAI Realtime connection is required. **Settings → Text & research models** selects the standard Gemini/NVIDIA or signed-in ChatGPT/Codex model. The request includes your question and neutral case/series metadata, not image pixels or patient records. Describe relevant findings in your question; this path does not inspect scans. **Connect voice** separately starts a new voice conversation. Jobs show progress, cancellation and saved results.
 
 **ChatGPT subscription:** in **Settings**, choose **Sign in with ChatGPT**, continue in the official browser flow, then select **ChatGPT / Codex subscription** and an account-available model under Text & research models. Sign-in preserves your existing model until you save. Uses your plan’s Codex allowance for text/public PubMed research; Realtime voice remains separately API-key billed. Credentials stay in an isolated OS-keyring account. See the [subscription runbook](roadmap/ai-backend/CODEX_SUBSCRIPTION.md) for setup, limits and validation.
 
-**Jev evidence review** is visible above the conversation: use **Review latest evidence with Jev** for the latest completed PubMed result, or **Saved research** to reopen a conversation. Preview the exact claims and abstracts, select claims, confirm public/synthetic text, then explicitly start the review. Jev shows abstract-support judgments and saved execution receipts; the original answer stays unchanged. It does not analyze image pixels.
+The sidebar separates **Chat**, **Research**, and **Jev review**. Chat discusses the case; Research shows literature answers and collapsed source lists. Optional voice setup is inside Chat. From a completed PubMed result, choose **Review evidence with Jev** to enter a dedicated review view. Select the unchanged cited passages, inspect their abstracts, confirm public/synthetic text, then explicitly start Jev. Results show abstract-support judgments with execution receipts under details. Empty previews explain why Jev has not run and offer **Prepare review again**. It does not analyze image pixels.
 
 **Literature research** cards show the recorded provider/model and worker steps, including waiting for the model, searching PubMed and preparing the answer. Completed, timed-out and cancelled jobs stay visible in history. Model configuration alone does not mean a job is running. [Desktop activation evidence](roadmap/ai-backend/DESKTOP_AI_ACTIVATION.md) separates local checks from hosted-provider results.
 

WARP.md

@@ -97,3 +97,5 @@ NVIDIA NIM is available for explicit evidence evaluation and opt-in PubMed resea
 Typed **Send** and explicit **Research** work without Gemini Live/OpenAI Realtime. Confirm synthetic/deidentified content and choose the standard model in **Settings → Text & research models**. Only the question, bounded text history (chat only) and neutral viewer metadata are sent; image pixels require separate live sharing. **Connect voice** starts a separate voice conversation. See `roadmap/ai-backend/DESKTOP_AI_ACTIVATION.md`.
 
 ChatGPT/Codex subscription sign-in is available under desktop AI Settings for typed chat and public PubMed research through isolated, pinned Codex App Server. Subscription credentials stay in the OS keyring; Realtime remains API-key billed. Read `roadmap/ai-backend/CODEX_SUBSCRIPTION.md`.
+
+The AI sidebar separates Chat, Research and Jev review. Optional voice setup stays in Chat; the review workspace hides the composer and collapses abstracts/technical receipts. Sidebar evidence uses versioned intact cited passages and explicitly reports empty previews as unavailable. See `roadmap/ai-backend/JEV_SIDEBAR_IMPLEMENTATION.md` for the corrected workflow and live acceptance.

backend/clinical/AGENTS.md

@@ -69,6 +69,8 @@
 
 ## Explicit Jev evidence reviews
 
+- New sidebar preparations use the versioned intact-passage extractor (`cited-passages-v2`) for trailing paragraph citations and comma-grouped source IDs. Preserve original offsets/text and rebuild the declared plan during evaluation. A preview with no executable pairs is unavailable with `no_reviewable_claims`; legacy ready/zero-pair summaries are projected the same way without rewriting artifacts. Explicit Prepare review again creates a fresh preview and requires renewed confirmation.
+
 - Own `ai_evidence_contracts.py`, `ai_evidence_repository.py`, `ai_evidence_artifacts.py`, `ai_evidence_review.py` and their additive `ai_jev_reviews`/`ai_jev_operations` persistence. These focused contracts stay separate from live conversation messages and primary-model tool results.
 - Only signed unexpired `ai.run` actors in enabled research/pilot may prepare an owned completed PubMed research result. Freeze its original answer/source/context identity and recorded generation, retrieve original abstracts by fixed PMID endpoints, then require exact-preview public/synthetic confirmation and backend-issued claim selection before pinned `jev-1.13.0` inference. Saved-history review does not require a live connection. The preview is not patient-text approval.
 - One active review per actor, two globally; preparation is bounded to 20 seconds, evaluation/cleanup to 70 (the runner itself remains 60 plus five). Recheck authority/source identity before every external call. Account stop invalidates queued starts, even if they were waiting for the owner lock. Operation identity and starting state commit together; duplicate requests never schedule duplicate inference.

backend/clinical/ai_evidence_review.py

@@ -80,9 +80,10 @@ def artifacts(self,row):
     def _summary(self,row):
         values={k:v for k,v in row.progress_json.items() if k in {
             'generation','totalPairs','completedPairs','settledPairs','submittedAttempts','unknownUsageAttempts'}}
+        empty = row.status == 'ready' and not values.get('totalPairs')
         return EvidenceReviewSummary(review_id=row.id,session_id=row.session_id,tool_call_id=row.tool_id,
-            source_context_version=row.context_version,status=row.status,created_at=row.created_at,
-            updated_at=row.updated_at,reason=row.reason,**values)
+            source_context_version=row.context_version,status='unavailable' if empty else row.status,created_at=row.created_at,
+            updated_at=row.updated_at,reason='no_reviewable_claims' if empty else row.reason,**values)
 
     def list(self,actor,session_id):
         self.require(actor)
@@ -247,11 +248,12 @@ async def _work(self,job):
                         'data_class':'public_literature','generation':job.row.progress_json['generation'],
                         'result':result.model_dump(mode='json'),'evidence':[e.model_dump(mode='json') for e in evidence],
                         'capture_exclusions':[e.model_dump(mode='json') for e in exclusions]},limits=self.limits)
-                    plan=build_review_plan(snapshot,limits=self.limits)
+                    plan=build_review_plan(snapshot,limits=self.limits,builder_version='cited-passages-v2')
                     ref,digest=self.artifacts(job.row).create_preview(snapshot,plan,job.row.progress_json['generation'])
                     self.repository.update_if_current(job.row.id,job.row.generation,preview_ref=ref,preview_hash=digest,
                         progress_json={**job.row.progress_json,'totalPairs':len(plan.pairs)})
-                    status='ready'
+                    status='ready' if plan.pairs else 'unavailable'
+                    reason=None if plan.pairs else 'no_reviewable_claims'
                 else:
                     with self.artifacts(job.row).open_run() as store:
                         job.store=store

backend/evidence_review/AGENTS.md

@@ -25,6 +25,7 @@ None.
 ## Evaluation inputs and providers
 
 - `sentence-citations-v1` keeps exact Unicode spans, compound sentences and separate cited abstracts. Ambiguous paragraph citation attachment is excluded; curated annotations must validate against original spans and sources.
+- Sidebar preparation explicitly uses `cited-passages-v2`: retain a whole paragraph/list item when only its final sentence is cited, rather than guessing individual sentence attribution. Recognize exact single and comma-grouped `[s1, s2]` citation spans; each cited abstract still receives an independent judgment. The runner validates the declared supported builder and reconstructs the full immutable plan. Legacy CLI/default and saved v1 plans retain their original sentence rules; never silently reinterpret an existing preview.
 - `settings.py` reads only explicitly requested deployment settings. Clinical/unknown modes reject network evaluation; Jev does not require Gemini credentials. The normal application provider catalog is unchanged.
 - `typesafe.py` pins Jev 1.13.0 and validates all five-way probabilities, model identity and usage. Adapters issue one attempt; the runner owns retries/deadlines. Transport discards error bodies, rejects compression and bounds success bodies.
 

backend/evidence_review/runner.py

@@ -46,7 +46,11 @@ async def evaluate_snapshot(snapshot: Snapshot, plan: ReviewPlan, *, adapter: Ev
     snapshot = load_snapshot(canonical_json(snapshot.model_dump(mode="json")), limits=limits)
     annotations = tuple(SpanAnnotation(start=u.start,end=u.end,citation_spans=u.citation_spans)
                         for u in plan.units if u.origin == "curated")
-    expected = build_review_plan(snapshot,limits=limits,annotations=annotations)
+    versions = {u.builder_version for u in plan.units}
+    if len(versions) > 1:
+        raise ValueError("evaluation_input_mismatch")
+    expected = build_review_plan(snapshot,limits=limits,annotations=annotations,
+        **({'builder_version':next(iter(versions))} if versions else {}))
     expected = select_review_plan(expected, selected_unit_ids=selected_unit_ids)
     selection = ([u.unit_id for u in expected.units if u.unit_id in selected_unit_ids]
                  if selected_unit_ids is not None else None)