try to fix gemma messing with other AIs

README.md

@@ -107,52 +107,55 @@ Ollama client sends `model: VISION_MODEL`, `think: false`, a JSON Schema in `for
 and the existing cancellation/timeouts. Its stream is consumed privately; only the
 final selected chat model's answer and opted-in reasoning reach the user.
 
-The bounded, validated vision record is:
+Gemma is the visual perception stage. The user's question guides which visible
+details it inspects; it does not answer that question. Qwen combines the report
+with the question, conversation, its own knowledge, and optional Web Search to
+reason and produce the final answer. The bounded perception record is:
 
 ```json
 {
-  "summary": "Concise factual overview",
-  "answer": "Direct visual conclusion relevant to the user's question",
-  "confidence": 0.9,
-  "evidence": ["Visible details supporting the conclusion"],
+  "scene": "Detailed factual overview of what is visible",
+  "subjects": [{"type": "bird", "description": "Fluffy gray down", "location": "center"}],
+  "visualDetails": ["Black head with large white facial patches"],
   "visibleText": ["Clearly readable text"],
-  "observations": [],
-  "alternatives": [],
+  "spatialRelationships": ["The bird faces the camera"],
+  "colorsAndAppearance": ["Gray body and black head"],
+  "recognizedElements": ["Juvenile penguin"],
+  "notableFeatures": ["Large white facial patches"],
   "uncertainties": []
 }
 ```
 
-Gemma adapts its emphasis to identification, UI debugging, OCR, counting, comparison,
-or scene understanding. It should give the most likely supported answer, qualified
-when appropriate, rather than refusing because absolute certainty is impossible.
-`answer` can be null for genuinely insufficient evidence. `confidence` is a number
-from 0 to 1 (or null), not a calibrated probability. Alternatives require actual
-support; empty arrays are preferred to speculative lists or boilerplate limitations.
-Each field has a length/item limit, with at most eight evidence items and three
-alternatives. The helper retains its 1,600-token output budget.
+Gemma reports direct observations and visually recognizable elements, while avoiding
+semantic conclusions such as a species inferred from markings or an unseen backend
+cause. Each field has a length/item limit that allows detailed screenshots. The
+generation cap is 2,400 tokens; simple images can use much less.
 
-Malformed JSON falls back to a bounded summary plus `rawAnalysis` (up to 6,000
-characters), with no invented answer or confidence. Valid JSON is parsed before
+Malformed JSON falls back to a bounded scene plus `rawAnalysis` (up to 6,000
+characters) as unstructured visual evidence. Valid JSON is parsed before
 individual fields are bounded, so truncation cannot corrupt its structure.
 An empty response or unavailable Gemma fails the turn explicitly, without asking the
 chat model to invent image details. The cache key is attachment, configured vision
-model, `VISION_PROMPT_VERSION` (currently 2), and trimmed user question; case and
+model, `VISION_PROMPT_VERSION` (currently 3), and trimmed user question; case and
 internal whitespace are preserved. Records include the analysis and timestamp.
-Only an exact matching key is reused. Legacy model-only snapshots are retained for
-compatibility but never used as cache hits. Bump the prompt version when changing
-the analysis contract. Cache records cascade away when the attachment is deleted.
+Only an exact matching key is reused. Version 2 answer-oriented reports remain in
+storage but cannot be cache hits for version 3. Legacy model-only snapshots are
+retained for compatibility but never used as cache hits. Bump the prompt version
+when changing the perception contract. Cache records cascade away when the
+attachment is deleted.
 
-Recent-history image context remains limited to the configured image count. A new
-question, including a follow-up without another upload, analyzes retained images
-for that question. Repeated questions and regeneration reuse matching results.
-Different question results coexist, including across shared branches. A new question
-can therefore require extra Gemma calls, bounded by the existing image count.
+Images on the current user message are analyzed individually and labeled in their
+attachment order. Earlier images are selected only for explicit historical references;
+they do not compete with an unrelated current image. Regeneration uses the pinned
+original prompt and its attachments. Repeated questions and regeneration reuse
+matching results. Different question results coexist, including across shared
+branches, so a new question can require extra Gemma calls.
 
 The final model receives intact structured evidence in a separate untrusted user
 context message, not a truncated JSON string in a system prompt. Trusted guidance
-tells it to answer directly using the answer, confidence and evidence, retain real
-ambiguity, avoid inventing alternatives, and distinguish visible UI state from unseen
-backend causes. It should not mention internal tools/models unless explicitly asked.
+tells Qwen to weigh observed features and visual recognition, perform the final
+interpretation itself, retain real ambiguity, and distinguish visible UI state from
+unseen backend causes. It should not mention internal tools/models unless explicitly asked.
 OCR and raw fallback text are data, never instructions. Analysis remains separate
 from visible messages; the final model never receives image bytes or filesystem paths.
 These are model instructions, not a deterministic guarantee of identification accuracy
@@ -160,8 +163,10 @@ or final-answer wording. Automated tests cover the contract and data flow; seman
 quality requires evaluation with real images and the deployed models.
 
 Web Search and Reasoning remain independent. For requested searches, concise visual
-evidence helps form the query without overriding the user's search decision. Search
-finishes before final generation; source ordering/rendering is unchanged. The final
+evidence from recognized elements, visible text, notable features, and scene helps
+form the query without overriding the user's search decision. Current images take
+priority over historical ones. Search finishes before final generation;
+source ordering/rendering is unchanged. The final
 request still uses `stream: true` and the existing capability-checked reasoning toggle.
 Stop/disconnect cancels the current vision or final generation request. Logging includes
 IDs, sizes, types and cache events only, never images/Base64/private analysis.