feat: connect NVIDIA Nemotron evidence reviewer

Built with OpenAI Codex: connect the verified private Nemotron Mini 4B ZeroGPU reviewer and publish the GitHub link.

FIELD_NOTES.md

@@ -35,6 +35,10 @@ The system deliberately separates three responsibilities:
 The models can read and route an investigation. They cannot silently invent a
 nutrition value or override the evidence standard.
 
+NVIDIA `Nemotron-Mini-4B-Instruct` performs a second, independent review after
+the investigation plan completes. It can identify the highest-priority missing
+evidence, but it cannot alter PacketCourt's deterministic verdict.
+
 ## What the investigation agent does
 
 Each packet creates a claim-dependent investigation plan. A `NO ADDED SUGAR`
@@ -62,6 +66,15 @@ a stratified held-out split. PacketCourt only enables the router after its
 measured result is recorded in the model card and its suggestions remain
 bounded by deterministic policy fallbacks.
 
+## A failed Nemotron deployment
+
+The first NVIDIA reviewer used `NVIDIA-Nemotron-3-Nano-4B-BF16`. A real
+ZeroGPU probe failed because the hybrid Mamba runtime required a specialized
+CUDA build unavailable in the standard Gradio image. Rather than claim a model
+that did not run, PacketCourt switched to `Nemotron-Mini-4B-Instruct`. The
+replacement completed a real ZeroGPU evidence-gap review before it was
+connected to the product.
+
 ## Persuasion Gap
 
 Claim verification alone was not enough. A `HIGH PROTEIN` claim can be
@@ -79,6 +92,7 @@ the decision with the user.
 - `10` transparent investigation traces are exported.
 - The vision model has `1.30B` parameters.
 - The fine-tuned evidence router has approximately `4.4M` parameters.
+- The independent NVIDIA Nemotron reviewer has approximately `4B` parameters.
 - The complete product interface is responsive and built on Gradio.
 
 ## What PacketCourt refuses to claim

README.md

@@ -12,6 +12,7 @@ tags:
   - track:backyard
   - sponsor:openbmb
   - sponsor:openai
+  - sponsor:nvidia
   - achievement:offbrand
   - achievement:tiny
   - achievement:sharing
@@ -55,6 +56,8 @@ flowchart LR
     M --> A["Investigation agent"]
     A --> FR["Fine-tuned evidence router<br/>4.4M parameters"]
     FR --> A
+    A --> NV["NVIDIA Nemotron Mini 4B<br/>Independent evidence-gap reviewer"]
+    NV --> A
     A --> P["Deterministic evidence parser<br/>CPU"]
     P --> C["Claim-to-evidence audit"]
     P --> N["Whole-packet nutrition math"]
@@ -75,10 +78,11 @@ flowchart LR
 
 Photo transcription uses the 1.30B-parameter OpenBMB `MiniCPM-V-4.6` through
 a private ZeroGPU companion. A fine-tuned 4.4M-parameter evidence router
-selects the investigation tools required by each claim. The main CPU Space
-performs deterministic evidence auditing, whole-packet calculations,
-persuasion-gap analysis, and refusals. ZeroGPU is requested only while reading
-photos.
+selects the investigation tools required by each claim. NVIDIA
+`Nemotron-Mini-4B-Instruct` independently reviews the completed investigation
+for evidence gaps. The main CPU Space performs deterministic evidence auditing,
+whole-packet calculations, persuasion-gap analysis, and refusals. ZeroGPU is
+requested only while running the vision witness or Nemotron reviewer.
 
 ## What It Audits
 
@@ -150,11 +154,13 @@ Current deterministic evaluation result:
 
 - Main private product: https://huggingface.co/spaces/build-small-hackathon/packetcourt
 - Private OpenBMB ZeroGPU vision companion: https://huggingface.co/spaces/build-small-hackathon/packetcourt-vision
+- Private NVIDIA Nemotron reviewer: https://huggingface.co/spaces/build-small-hackathon/packetcourt-nemotron
 - Private golden evaluation dataset: https://huggingface.co/datasets/build-small-hackathon/packetcourt-golden-cases
 - Public transparent agent traces: https://huggingface.co/datasets/build-small-hackathon/packetcourt-traces
 - Fine-tuned evidence router: https://huggingface.co/build-small-hackathon/packetcourt-evidence-router
 - Public router training set: https://huggingface.co/datasets/build-small-hackathon/packetcourt-router-training
 - Public Field Notes report: https://huggingface.co/datasets/build-small-hackathon/packetcourt-field-notes
+- Public Codex-attributed GitHub repository: https://github.com/N-45div/PacketCourt
 
 ## Safety Boundary
 

app.py

@@ -17,6 +17,8 @@ sys.path.insert(0, str(ROOT / "src"))
 from packetcourt import audit_packet
 from packetcourt.ocr import extract_text
 from packetcourt.remote_vision import extract_remote, is_configured
+from packetcourt.remote_nemotron import is_configured as nemotron_is_configured
+from packetcourt.remote_nemotron import review as nemotron_review
 from packetcourt.samples import SAMPLES
 from packetcourt.vlm import model_status
 
@@ -26,6 +28,29 @@ class AuditRequest(BaseModel):
     back_text: str
 
 
+def run_audit(front_text: str, back_text: str):
+    result = audit_packet(front_text, back_text)
+    if not nemotron_is_configured():
+        return result
+    snapshot = {
+        "claims": [claim.model_dump(mode="json") for claim in result.claims],
+        "investigation": result.investigation.model_dump(),
+        "nutrition": result.nutrition.model_dump(),
+        "ingredients_found": bool(result.ingredients),
+        "expiry": result.expiry.model_dump(),
+        "limitations": result.limitations,
+    }
+    try:
+        result.agent_review = nemotron_review(snapshot)
+    except Exception as exc:
+        result.agent_review = {
+            "status": "UNAVAILABLE",
+            "rationale": f"Nemotron review unavailable: {type(exc).__name__}",
+            "model": "nvidia/Nemotron-Mini-4B-Instruct",
+        }
+    return result
+
+
 def build_gradio_engine() -> gr.Blocks:
     with gr.Blocks(title="PacketCourt Engine") as engine:
         gr.Markdown(
@@ -37,7 +62,7 @@ def build_gradio_engine() -> gr.Blocks:
         back = gr.Textbox(label="Back-label evidence")
         output = gr.JSON(label="Evidence case")
         gr.Button("Audit").click(
-            lambda front_text, back_text: audit_packet(front_text, back_text).model_dump(mode="json"),
+            lambda front_text, back_text: run_audit(front_text, back_text).model_dump(mode="json"),
             [front, back],
             output,
         )
@@ -68,6 +93,11 @@ def model() -> dict:
         if os.getenv("PACKETCOURT_ROUTER", "0") == "1"
         else "deterministic fallback"
     )
+    status["nemotron_reviewer"] = (
+        "nvidia/Nemotron-Mini-4B-Instruct"
+        if nemotron_is_configured()
+        else "not configured"
+    )
     if is_configured():
         status.update(
             enabled=True,
@@ -79,7 +109,7 @@ def model() -> dict:
 
 @app.post("/api/audit")
 def audit(request: AuditRequest) -> dict:
-    return audit_packet(request.front_text, request.back_text).model_dump(mode="json")
+    return run_audit(request.front_text, request.back_text).model_dump(mode="json")
 
 
 @app.post("/api/ocr")

frontend/app.js

@@ -59,6 +59,12 @@ function render(data) {
   $("#missing-evidence").textContent = data.investigation.missing_evidence.length
     ? data.investigation.missing_evidence.join(" · ")
     : "None. The required evidence path completed.";
+  const review = data.agent_review;
+  $("#nemotron-review").innerHTML = review.status === "NOT_REQUESTED"
+    ? ""
+    : `<b>NVIDIA NEMOTRON INDEPENDENT REVIEW · ${escapeHtml(review.status)}</b>
+       <strong>${escapeHtml(review.priority || "No additional action requested.")}</strong>
+       <span>${escapeHtml(review.evidence_request || review.rationale)}</span>`;
   $("#claim-grid").innerHTML = data.claims.length
     ? data.claims.map((claim) => `
       <article class="claim-card ${verdictClass[claim.verdict]}">

frontend/index.html

@@ -104,6 +104,7 @@
           <p><b>STOP REASON</b><span id="stop-reason"></span></p>
           <p><b>MISSING EVIDENCE</b><span id="missing-evidence"></span></p>
         </div>
+        <div class="nemotron-review" id="nemotron-review"></div>
       </section>
       <section class="gap-section">
         <div class="gap-heading"><p class="kicker">PERSUASION GAP</p><h3>Material context the front leaves quiet.</h3></div>

frontend/styles.css

@@ -15,6 +15,7 @@ main{max-width:1320px;margin:auto;padding:0 4vw}.hero{min-height:670px;display:g
 .sample-card{padding:24px;background:var(--cream);border:1px solid var(--line);border-radius:16px;cursor:pointer;text-align:left;transition:.2s}.sample-card:hover{border-color:var(--red);transform:translateY(-3px)}.sample-card b{display:block;font-size:19px;margin-bottom:8px}.sample-card span{font-size:13px;color:var(--muted)}
 .results{border-top:1px solid var(--line)}.hidden{display:none}.case-score{width:120px;height:120px;border:1px solid var(--ink);border-radius:50%;display:flex;align-items:center;justify-content:center;gap:8px}.case-score span{font:800 45px "Playfair Display"}.case-score small{font:500 8px/1.4 "DM Mono"}
 .agent-section{margin-bottom:28px;padding:26px;border:1px solid var(--line);background:var(--cream);border-radius:18px}.agent-heading{display:flex;justify-content:space-between;align-items:end;gap:20px;margin-bottom:18px}.agent-heading h3{font:700 clamp(27px,4vw,46px)/1 Georgia,serif;margin:0}.agent-heading>span{font:500 9px ui-monospace,SFMono-Regular,Menlo,monospace;padding:8px 11px;border:1px solid var(--line);border-radius:99px;color:var(--green)}.agent-steps{display:grid;grid-template-columns:repeat(2,1fr);gap:9px}.agent-steps article{display:grid;grid-template-columns:auto 1fr;gap:10px;padding:15px;border:1px solid var(--line);border-radius:11px;background:#f8f3e9}.agent-steps article>span{font:500 10px ui-monospace,SFMono-Regular,Menlo,monospace;color:var(--red)}.agent-steps b{font-size:13px;text-transform:capitalize}.agent-steps p{font-size:11px;line-height:1.4;color:var(--muted);margin:5px 0}.agent-steps small{grid-column:2;font:500 8px ui-monospace,SFMono-Regular,Menlo,monospace;text-transform:uppercase;color:var(--green)}.agent-stop{display:grid;grid-template-columns:1fr 1fr;gap:9px;margin-top:9px}.agent-stop p{margin:0;padding:14px;border-top:1px solid var(--line);font-size:11px;line-height:1.5}.agent-stop b,.agent-stop span{display:block}.agent-stop b{font:500 8px ui-monospace,SFMono-Regular,Menlo,monospace;color:var(--muted);margin-bottom:5px}
+.nemotron-review{display:grid;gap:6px;margin-top:12px;padding:16px;border:1px solid #9fc2a9;border-radius:11px;background:#edf5eb}.nemotron-review:empty{display:none}.nemotron-review b{font:600 8px ui-monospace,SFMono-Regular,Menlo,monospace;letter-spacing:.1em;color:var(--green)}.nemotron-review strong{font-size:14px}.nemotron-review span{font-size:11px;line-height:1.5;color:var(--muted)}
 .gap-section{margin-bottom:28px;padding:26px;border:1px solid var(--ink);background:#1b1b17;color:var(--cream);border-radius:18px}.gap-heading{display:flex;justify-content:space-between;gap:20px;align-items:end;margin-bottom:18px}.gap-heading .kicker{color:#bdb4a6}.gap-heading h3{font:700 clamp(27px,4vw,46px)/1 Georgia,serif;max-width:700px;margin:0}.gap-grid{display:grid;grid-template-columns:repeat(2,1fr);gap:12px}.gap-card,.gap-empty{padding:19px;border:1px solid #4a483f;border-radius:13px;background:#26251f}.gap-card.high{border-color:var(--red)}.gap-card.medium{border-color:var(--amber)}.gap-severity{font:500 8px ui-monospace,SFMono-Regular,Menlo,monospace;letter-spacing:.13em;text-transform:uppercase;color:#d9a65e}.gap-card h4{font-size:20px;margin:10px 0 16px}.gap-compare{display:grid;grid-template-columns:1fr 1fr;gap:10px}.gap-compare p{margin:0;padding:11px;background:#313029;border-radius:8px;font-size:12px;line-height:1.45}.gap-compare b{display:block;font:500 8px ui-monospace,SFMono-Regular,Menlo,monospace;color:#bdb4a6;margin-bottom:5px}.gap-card .evidence{border-top-color:#4a483f}.gap-empty{display:grid;gap:5px;color:#bdb4a6}
 .claim-grid{grid-template-columns:repeat(2,1fr)}.claim-card{background:var(--cream);border:1px solid var(--line);border-top:6px solid var(--muted);border-radius:16px;padding:23px}.claim-card.supported{border-top-color:var(--green)}.claim-card.contradicted{border-top-color:var(--red)}.claim-card.context{border-top-color:var(--amber)}.claim-top{display:flex;justify-content:space-between;gap:10px;align-items:start}.claim-name{font-size:21px;font-weight:800}.verdict{font:500 8px/1.3 ui-monospace,SFMono-Regular,Menlo,monospace;letter-spacing:.08em;border:1px solid var(--line);border-radius:99px;padding:7px 9px;text-align:right}.confidence{display:block;margin-top:8px;font:500 8px ui-monospace,SFMono-Regular,Menlo,monospace;letter-spacing:.1em;text-transform:uppercase;color:var(--muted)}.summary{min-height:45px;color:#534d43;line-height:1.5}.evidence{padding:11px 0;border-top:1px solid var(--line)}.evidence b,.evidence span{display:block}.evidence b{font:500 8px ui-monospace,SFMono-Regular,Menlo,monospace;letter-spacing:.12em;color:var(--muted);text-transform:uppercase}.evidence span{font-size:13px;margin-top:4px}.caveat{font-size:11px;color:var(--muted);margin-top:15px}
 .evidence-summary{margin-top:16px}.evidence-summary article{padding:25px;border:1px solid var(--line);border-radius:16px;background:var(--cream)}#nutrition-grid div{display:flex;justify-content:space-between;padding:10px 0;border-bottom:1px solid var(--line);font-size:13px}.date-card{background:var(--ink)!important;color:var(--cream)}.date-card .kicker{color:#cfc5b6}.date-card h3{font:700 28px/1.15 "Playfair Display"}details{margin-top:16px;border:1px solid var(--line);border-radius:14px;padding:17px;background:var(--cream)}summary{cursor:pointer;font-weight:700}pre{white-space:pre-wrap;font:11px/1.5 "DM Mono";overflow:auto}

src/packetcourt/models.py

@@ -80,6 +80,14 @@ class InvestigationPlan(BaseModel):
     router_model: str = "deterministic fallback"
 
 
+class AgentReview(BaseModel):
+    status: str = "NOT_REQUESTED"
+    priority: str = ""
+    evidence_request: str = ""
+    rationale: str = ""
+    model: str = ""
+
+
 class PacketAudit(BaseModel):
     claims: list[ClaimAudit]
     nutrition: NutritionFacts
@@ -91,3 +99,4 @@ class PacketAudit(BaseModel):
     back_text: str
     limitations: list[str]
     investigation: InvestigationPlan
+    agent_review: AgentReview = Field(default_factory=AgentReview)

src/packetcourt/remote_nemotron.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+import json
+import os
+from functools import lru_cache
+
+
+def is_configured() -> bool:
+    return bool(os.getenv("PACKETCOURT_NEMOTRON_SPACE"))
+
+
+@lru_cache(maxsize=1)
+def _client():
+    from gradio_client import Client
+
+    try:
+        return Client(os.environ["PACKETCOURT_NEMOTRON_SPACE"], hf_token=os.getenv("HF_TOKEN"))
+    except TypeError:
+        return Client(os.environ["PACKETCOURT_NEMOTRON_SPACE"], token=os.getenv("HF_TOKEN"))
+
+
+def review(snapshot: dict) -> dict:
+    result = _client().predict(json.dumps(snapshot), api_name="/predict")
+    return json.loads(result)

tests/test_audit.py

@@ -87,3 +87,4 @@ def test_investigation_requests_missing_evidence_and_stops_explicitly():
     assert any(step.tool == "inspect_nutrition" for step in result.investigation.steps)
     assert any("nutrition panel" in item.lower() for item in result.investigation.missing_evidence)
     assert "missing-evidence" in result.investigation.stop_reason
+    assert result.agent_review.status == "NOT_REQUESTED"