Upload folder using huggingface_hub

app/ai/interpreter.py

@@ -1,5 +1,6 @@
 import asyncio
 import json
+import hashlib
 import logging
 import re
 from typing import AsyncGenerator
@@ -90,6 +91,21 @@ def _friendly_error(error: BaseException) -> str:
     return f"AI interpretation failed: {text}"
 
 
+def _numeric_grounding(text: str, context: dict) -> dict:
+    """Check numeric presence only; this does not establish biological truth."""
+    from app.services.evidence_policy import numbers
+    evidence = json.dumps(context, sort_keys=True, separators=(",", ":"), default=str)
+    unsupported = set(numbers(text)) - set(numbers(evidence))
+    if unsupported:
+        raise ValueError("AI explanation contained numeric claims absent from the recorded evidence.")
+    return {
+        "status": "numeric_fidelity_passed",
+        "evidence_sha256": hashlib.sha256(evidence.encode()).hexdigest(),
+        "evidence_class": "AI-generated interpretation",
+        "limitations": ["Numeric presence does not verify units, attribution, citations, causality or biological correctness."],
+    }
+
+
 async def interpret_stream(pipeline_type: str, context: dict) -> AsyncGenerator[str, None]:
     acompletion = _get_acompletion()
     if acompletion is None:
@@ -118,11 +134,18 @@ async def interpret_stream(pipeline_type: str, context: dict) -> AsyncGenerator[
                     timeout=25,
                     api_key=candidate["api_key"],
                 )
+                # Do not expose an incomplete provider reply or mix it with a
+                # backup reply. Admit the complete explanation only after its gate.
+                parts = []
                 async for chunk in response:
                     if chunk.choices and chunk.choices[0].delta.content:
-                        yield _chunk_event(chunk.choices[0].delta.content)
-
-                yield _done_event({"model": candidate["model"], "pipeline_type": pipeline_type})
+                        parts.append(chunk.choices[0].delta.content)
+                text = "".join(parts)
+                if not text.strip():
+                    raise ValueError("AI provider returned an empty explanation.")
+                grounding = _numeric_grounding(text, context)
+                yield _chunk_event(text)
+                yield _done_event({"model": candidate["model"], "pipeline_type": pipeline_type, "grounding": grounding})
                 return
             except Exception as e:
                 last_error = e
@@ -173,7 +196,7 @@ async def interpret_text(pipeline_type: str, context: dict) -> dict:
                 )
                 text = response.choices[0].message.content if response.choices else ""
                 if text:
-                    return {"interpretation": text}
+                    return {"interpretation": text, "grounding": _numeric_grounding(text, context)}
             except Exception as e:
                 last_error = e
                 logger.warning("LLM provider %s attempt %d failed: %s", candidate["name"], attempt + 1, e)

app/ai/tool_interpreter.py

@@ -61,10 +61,12 @@ _TOOL_PROMPTS: dict[str, str] = {
         "concerns about data quality. 2-4 sentences max."
     ),
     "md": (
-        "You are a computational biophysicist. A molecular dynamics simulation "
-        "just completed. Summarize the simulation parameters, energy minimization "
-        "result, key structural observations, and whether the simulation converged. "
-        "2-3 sentences max."
+        "You are a computational biophysicist. Explain the recorded engine, "
+        "simulation duration and computed observables. State convergence only "
+        "when an explicit convergence diagnostic is present; a short or stable "
+        "RMSD trace alone does not establish convergence. A simulation does not "
+        "establish experimental binding, efficacy or a biological mechanism. "
+        "Use only reported results. 2-3 sentences max."
     ),
     "phylo": (
         "You are an evolutionary biologist. A phylogenetic analysis just completed. "
@@ -78,9 +80,11 @@ _TOOL_PROMPTS: dict[str, str] = {
         "2-3 sentences max."
     ),
     "admet": (
-        "You are a medicinal chemist. An ADMET property prediction just completed. "
-        "Summarize absorption, distribution, metabolism, excretion, and toxicity "
-        "profiles. Flag any Lipinski violations or red flags for drug-likeness. "
+        "You are a medicinal chemist. Separate recorded physicochemical descriptors "
+        "and rule-based alerts from predictions made by actually executed ADMET "
+        "models. Discuss absorption, metabolism or toxicity only for endpoints "
+        "reported by a named model; preserve applicability-domain limitations. "
+        "Descriptors and alerts do not establish measured ADMET or clinical safety. "
         "2-3 sentences max."
     ),
     "primers": (
@@ -206,6 +210,12 @@ async def interpret_tool_result(tool_name: str, result: dict) -> dict | None:
     except Exception:
         return None
 
+    system_prompt += (" Use only the recorded method identity and evidence class. "
+                      "Missing values mean not reported, never zero. Do not infer "
+                      "experimental validation, causality or clinical suitability "
+                      "from computation. Qualitative biological statements must "
+                      "remain hypotheses unless supported by recorded evidence. "
+                      "Do not invent citations or links.")
     result_text = _summarize_result(result)
 
     user_prompt = (