feat: dynamic audits and community evidence review agent

Codex-authored dynamic claim handling, ambiguous date evidence, and approval-gated feedback learning queue.

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 N DIVIJ
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -32,6 +32,12 @@ verdicts instead of an unexplained health score.
 
 **Open PacketCourt:** https://build-small-hackathon-packetcourt.hf.space/
 
+PacketCourt also includes a correction-driven Community Review Agent. User
+feedback is bundled with the original evidence, investigation path, and
+Nemotron review in a public queue. To prevent feedback poisoning, corrections
+must be evidence-reviewed before they become eligible for the next router
+fine-tune.
+
 ## Why PacketCourt
 
 A packet may lead with `HIGH PROTEIN`, `MULTIGRAIN`, or `100% NATURAL` while
@@ -162,6 +168,7 @@ Current deterministic evaluation result:
 - 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 community feedback queue: https://huggingface.co/datasets/build-small-hackathon/packetcourt-community-feedback
 - 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
 

app.py

@@ -2,8 +2,11 @@ from __future__ import annotations
 
 import os
 import sys
+import json
+from datetime import datetime, timezone
 from pathlib import Path
 from tempfile import NamedTemporaryFile
+from uuid import uuid4
 
 import gradio as gr
 import uvicorn
@@ -28,6 +31,12 @@ class AuditRequest(BaseModel):
     back_text: str
 
 
+class FeedbackRequest(BaseModel):
+    verdict: str
+    correction: str = ""
+    audit: dict
+
+
 def run_audit(front_text: str, back_text: str):
     result = audit_packet(front_text, back_text)
     if not nemotron_is_configured():
@@ -112,6 +121,67 @@ def audit(request: AuditRequest) -> dict:
     return run_audit(request.front_text, request.back_text).model_dump(mode="json")
 
 
+@app.post("/api/feedback")
+def feedback(request: FeedbackRequest) -> dict:
+    if request.verdict not in {"accurate", "needs_correction"}:
+        return {"status": "REJECTED", "message": "Choose accurate or needs correction."}
+    if request.verdict == "needs_correction" and len(request.correction.strip()) < 8:
+        return {"status": "REJECTED", "message": "Explain the correction so it can be reviewed."}
+
+    record_id = str(uuid4())
+    record = {
+        "id": record_id,
+        "created_at": datetime.now(timezone.utc).isoformat(),
+        "verdict": request.verdict,
+        "correction": request.correction.strip()[:1200],
+        "front_text": str(request.audit.get("front_text", ""))[:3000],
+        "back_text": str(request.audit.get("back_text", ""))[:9000],
+        "claims": request.audit.get("claims", []),
+        "investigation": request.audit.get("investigation", {}),
+        "nemotron_review": request.audit.get("agent_review", {}),
+        "proposed_router_examples": [
+            {
+                "text": claim.get("claim", ""),
+                "candidate_tools": [
+                    step.get("tool", "")
+                    for step in request.audit.get("investigation", {}).get("steps", [])
+                ],
+            }
+            for claim in request.audit.get("claims", [])
+        ],
+        "review_status": "pending_human_review",
+        "training_eligible": False,
+        "learning_policy": "Only approved corrections enter the next evidence-router fine-tune.",
+    }
+    dataset_id = os.getenv("PACKETCOURT_FEEDBACK_DATASET")
+    if not dataset_id:
+        return {
+            "status": "UNAVAILABLE",
+            "message": "The community learning queue is not configured on this deployment.",
+        }
+    try:
+        from huggingface_hub import HfApi
+
+        HfApi().upload_file(
+            path_or_fileobj=json.dumps(record, indent=2).encode(),
+            path_in_repo=f"feedback/{record_id}.json",
+            repo_id=dataset_id,
+            repo_type="dataset",
+            commit_message=f"feedback: queue PacketCourt review {record_id[:8]}",
+        )
+    except Exception as exc:
+        return {
+            "status": "UNAVAILABLE",
+            "message": f"Feedback could not be persisted: {type(exc).__name__}",
+        }
+    return {
+        "status": "QUEUED",
+        "id": record_id,
+        "message": "Review queued. It will become training data only after evidence review.",
+        "dataset": f"https://huggingface.co/datasets/{dataset_id}",
+    }
+
+
 @app.post("/api/ocr")
 async def ocr(front: UploadFile | None = File(default=None), back: UploadFile | None = File(default=None)) -> dict:
     result: dict[str, dict[str, str]] = {}

data/community_feedback_README.md

@@ -0,0 +1,41 @@
+---
+license: mit
+tags:
+- build-small-hackathon
+- packetcourt
+- feedback
+- traces
+- human-in-the-loop
+---
+
+# PacketCourt Community Feedback
+
+This dataset is PacketCourt's public correction-driven learning queue.
+
+Each submitted review preserves:
+
+- the original front and back label evidence;
+- PacketCourt's claim verdicts and investigation path;
+- the independent Nemotron evidence-gap review;
+- the user's correction or confirmation;
+- candidate evidence-router examples.
+
+## Learning policy
+
+New records begin with:
+
+```json
+{
+  "review_status": "pending_human_review",
+  "training_eligible": false
+}
+```
+
+Public feedback never fine-tunes a production model immediately. That would
+allow accidental or malicious feedback to poison later audits. A correction
+must first be checked against the supplied packet evidence. Approved records
+can then be promoted into a versioned router-training release and evaluated
+against PacketCourt's golden cases before deployment.
+
+Nemotron reviews investigations and missing evidence. It is not silently
+self-modified by community feedback.

docs/COMMUNITY_LEARNING.md

@@ -0,0 +1,23 @@
+# Community Learning Loop
+
+```mermaid
+flowchart LR
+    A["Packet audit"] --> R["User review"]
+    R --> Q["Public feedback queue"]
+    Q --> H["Evidence review"]
+    H -->|reject| X["Retain as rejected trace"]
+    H -->|approve| T["Versioned router training set"]
+    T --> F["Fine-tune tiny evidence router"]
+    F --> E["Golden-case regression evaluation"]
+    E -->|pass| D["Deploy reviewed checkpoint"]
+    E -->|fail| X
+```
+
+The loop is deliberately approval-gated. User feedback is valuable evidence,
+but it is not automatically true. Every queued correction includes the audit,
+investigation trace, and Nemotron review so a reviewer can decide whether it
+should become training data.
+
+PacketCourt's deterministic verdict engine and safety boundaries are never
+rewritten by public feedback. Nemotron remains an independent reviewer rather
+than a model that silently trains on its own outputs.

frontend/app.js

@@ -7,6 +7,8 @@ const verdictClass = {
   "TECHNICALLY TRUE, CONTEXT MISSING": "context",
   "CANNOT VERIFY": "unknown",
 };
+let currentAudit = null;
+let feedbackVerdict = "";
 
 function setMode(mode) {
   $$(".mode-switch button").forEach((button) => button.classList.toggle("active", button.dataset.mode === mode));
@@ -46,6 +48,7 @@ function escapeHtml(value = "") {
 }
 
 function render(data) {
+  currentAudit = data;
   $("#claim-count").textContent = data.claims.length;
   $("#router-model").textContent = data.investigation.router_model;
   $("#agent-steps").innerHTML = data.investigation.steps.map((step, index) => `
@@ -103,6 +106,30 @@ function render(data) {
   $("#results").scrollIntoView({ behavior: "smooth" });
 }
 
+$$("[data-feedback]").forEach((button) => button.addEventListener("click", () => {
+  feedbackVerdict = button.dataset.feedback;
+  $$("[data-feedback]").forEach((choice) => choice.classList.toggle("active", choice === button));
+}));
+
+$("#submit-feedback").addEventListener("click", async () => {
+  if (!currentAudit || !feedbackVerdict) {
+    $("#feedback-status").textContent = "Run an audit and choose a review outcome first.";
+    return;
+  }
+  $("#feedback-status").textContent = "Packaging evidence review...";
+  const response = await fetch("api/feedback", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({
+      verdict: feedbackVerdict,
+      correction: $("#feedback-correction").value,
+      audit: currentAudit,
+    }),
+  });
+  const result = await response.json();
+  $("#feedback-status").textContent = result.message;
+});
+
 $("#audit-text").addEventListener("click", async () => {
   $("#audit-text").textContent = "Examining evidence...";
   try { await runAudit($("#front-text").value, $("#back-text").value); }

frontend/index.html

@@ -116,6 +116,22 @@
         <article class="date-card"><p class="kicker">DATE EVIDENCE</p><h3 id="expiry-status"></h3><p id="opening-status"></p><p>Expiry interpretation is evidence, not a food-safety guarantee.</p></article>
       </div>
       <details><summary>View machine-readable evidence case</summary><pre id="raw-json"></pre></details>
+      <section class="feedback-agent">
+        <div>
+          <p class="kicker">COMMUNITY REVIEW AGENT</p>
+          <h3>Help the next audit get sharper.</h3>
+          <p>Corrections enter a public review queue with this audit's evidence, router path, and Nemotron review. They never retrain production models without approval.</p>
+        </div>
+        <div class="feedback-controls">
+          <div class="feedback-choice">
+            <button class="button quiet" data-feedback="accurate">This audit is accurate</button>
+            <button class="button quiet" data-feedback="needs_correction">This needs correction</button>
+          </div>
+          <textarea id="feedback-correction" placeholder="What did PacketCourt miss or misread? Cite the packet evidence."></textarea>
+          <button class="button dark" id="submit-feedback">Queue evidence review <span>→</span></button>
+          <span id="feedback-status"></span>
+        </div>
+      </section>
     </section>
 
     <section class="method">

frontend/styles.css

@@ -20,6 +20,7 @@ main{max-width:1320px;margin:auto;padding:0 4vw}.hero{min-height:670px;display:g
 .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}
 .method{border-top:1px solid var(--line)}.method-grid{grid-template-columns:repeat(4,1fr);margin-top:40px}.method-grid div{padding:20px;border-top:2px solid var(--ink)}.method-grid span{font:500 10px "DM Mono";color:var(--red)}.method-grid p{font-size:13px;line-height:1.5;color:var(--muted)}
+.feedback-agent{display:grid;grid-template-columns:.8fr 1.2fr;gap:28px;margin-top:18px;padding:26px;border:1px solid var(--line);border-radius:18px;background:var(--cream)}.feedback-agent h3{font:700 clamp(25px,4vw,42px)/1 Georgia,serif;margin:0}.feedback-agent p{font-size:12px;line-height:1.55;color:var(--muted)}.feedback-controls{display:grid;gap:10px}.feedback-choice{display:flex;gap:8px;flex-wrap:wrap}.feedback-choice button.active{background:var(--red);border-color:var(--red);color:white}.feedback-controls textarea{min-height:95px;padding:13px;border:1px solid var(--line);border-radius:11px;background:#f8f3e9;resize:vertical}.feedback-controls>span{font:500 9px/1.5 ui-monospace,SFMono-Regular,Menlo,monospace;color:var(--green)}
 footer{display:flex;justify-content:space-between;gap:20px;padding:25px 5vw;border-top:1px solid var(--line);font:500 10px "DM Mono";color:var(--muted)}
-@media(max-width:900px){.hero{grid-template-columns:1fr;min-height:auto;padding:80px 0}.hero-visual{height:430px}.trust-strip{grid-template-columns:1fr}.trust-strip div{border-right:0}.section-heading,.case-header,.agent-heading{align-items:start;flex-direction:column}.upload-grid,.text-grid,.claim-grid,.evidence-summary,.gap-grid,.agent-steps,.agent-stop{grid-template-columns:1fr}.method-grid{grid-template-columns:repeat(2,1fr)}}
+@media(max-width:900px){.hero{grid-template-columns:1fr;min-height:auto;padding:80px 0}.hero-visual{height:430px}.trust-strip{grid-template-columns:1fr}.trust-strip div{border-right:0}.section-heading,.case-header,.agent-heading{align-items:start;flex-direction:column}.upload-grid,.text-grid,.claim-grid,.evidence-summary,.gap-grid,.agent-steps,.agent-stop,.feedback-agent{grid-template-columns:1fr}.method-grid{grid-template-columns:repeat(2,1fr)}}
 @media(max-width:560px){.top-status,.engine-link{display:none}.hero h1{font-size:58px}.hero-visual{transform:scale(.8);transform-origin:left top;height:350px;width:125%}.workspace,.results,.method{padding:65px 0}.mode-switch{overflow:auto}.mode-switch button{white-space:nowrap;padding:13px 10px}.sample-grid,.method-grid{grid-template-columns:1fr}.case-score{width:90px;height:90px}.claim-top{display:block}.verdict{display:inline-block;margin-top:8px}footer{display:block}footer span{display:block;margin:5px 0}}

nemotron_space/README.md

@@ -0,0 +1,28 @@
+---
+title: PacketCourt Nemotron Reviewer
+emoji: 🟩
+colorFrom: green
+colorTo: gray
+sdk: gradio
+sdk_version: 5.49.1
+app_file: app.py
+pinned: false
+license: other
+tags:
+  - build-small-hackathon
+  - sponsor:nvidia
+  - nemotron
+  - agent
+  - zerogpu
+models:
+  - nvidia/Nemotron-Mini-4B-Instruct
+---
+
+# PacketCourt Nemotron Reviewer
+
+Private ZeroGPU companion that uses NVIDIA Nemotron Mini 4B as an independent
+evidence-gap reviewer for PacketCourt investigations.
+
+Nemotron may request missing packet evidence or confirm that the bounded
+investigation plan is complete. It never issues or overrides PacketCourt's
+deterministic verdicts.

nemotron_space/app.py

@@ -0,0 +1,86 @@
+from __future__ import annotations
+
+import json
+import traceback
+from functools import lru_cache
+
+import gradio as gr
+import spaces
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+MODEL_ID = "nvidia/Nemotron-Mini-4B-Instruct"
+
+SYSTEM = """You are PacketCourt's evidence-gap reviewer.
+Review the supplied packet investigation plan. Do not judge whether food is
+healthy, safe, legal, or fraudulent. Do not change claim verdicts. Return only
+compact JSON with these keys:
+- status: COMPLETE or NEEDS_EVIDENCE
+- priority: one short sentence naming the most important next action
+- evidence_request: one short sentence, or an empty string
+- rationale: one short sentence grounded only in the supplied investigation
+"""
+
+
+@lru_cache(maxsize=1)
+def load_model():
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(
+        MODEL_ID,
+        trust_remote_code=True,
+        torch_dtype=torch.bfloat16,
+        device_map="auto",
+    )
+    model.eval()
+    return tokenizer, model
+
+
+@spaces.GPU(duration=180)
+def review_investigation(snapshot: str) -> str:
+    try:
+        tokenizer, model = load_model()
+        messages = [
+            {"role": "system", "content": SYSTEM},
+            {"role": "user", "content": snapshot[:12000]},
+        ]
+        prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+        generated = model.generate(**inputs, max_new_tokens=180, do_sample=False)
+        text = tokenizer.decode(generated[0][inputs.input_ids.shape[1] :], skip_special_tokens=True).strip()
+        start, end = text.find("{"), text.rfind("}")
+        if start == -1 or end == -1:
+            raise ValueError("Nemotron did not return JSON")
+        payload = json.loads(text[start : end + 1])
+        return json.dumps(
+            {
+                "status": "NEEDS_EVIDENCE" if payload.get("status") == "NEEDS_EVIDENCE" else "COMPLETE",
+                "priority": str(payload.get("priority", ""))[:240],
+                "evidence_request": str(payload.get("evidence_request", ""))[:240],
+                "rationale": str(payload.get("rationale", ""))[:320],
+                "model": MODEL_ID,
+            }
+        )
+    except Exception as exc:
+        return json.dumps(
+            {
+                "status": "UNAVAILABLE",
+                "priority": "",
+                "evidence_request": "",
+                "rationale": f"{type(exc).__name__}: {exc}",
+                "model": MODEL_ID,
+                "diagnostic": traceback.format_exc(limit=3),
+            }
+        )
+
+
+demo = gr.Interface(
+    fn=review_investigation,
+    inputs=gr.Textbox(label="Structured PacketCourt investigation snapshot", lines=14),
+    outputs=gr.Textbox(label="Nemotron evidence-gap review"),
+    title="PacketCourt Nemotron Reviewer",
+    description="NVIDIA Nemotron Mini 4B reviews bounded PacketCourt investigations for missing evidence.",
+    flagging_mode="never",
+)
+
+if __name__ == "__main__":
+    demo.launch()

nemotron_space/requirements.txt

@@ -0,0 +1,5 @@
+gradio==5.49.1
+spaces>=0.42.0
+torch>=2.6.0
+transformers>=4.53.0
+accelerate>=1.8.0

requirements.txt

@@ -1,5 +1,6 @@
 gradio==5.49.1
 gradio_client==1.13.3
+huggingface_hub>=0.33.0
 pillow>=11.0.0
 pydantic>=2.10.0
 pytesseract>=0.3.13

router_dataset/README.md

@@ -0,0 +1,29 @@
+---
+license: cc-by-4.0
+task_categories:
+  - text-classification
+language:
+  - en
+tags:
+  - build-small-hackathon
+  - packetcourt
+  - claim-routing
+size_categories:
+  - n<1K
+---
+
+# PacketCourt Evidence Router Training Set
+
+Small, inspectable claim-routing dataset used to fine-tune
+[`packetcourt-evidence-router`](https://huggingface.co/build-small-hackathon/packetcourt-evidence-router).
+
+The five labels map packet text to the next bounded investigation tool:
+
+- `ingredients`
+- `nutrition`
+- `license`
+- `dates`
+- `refuse_absolute`
+
+The router only proposes a tool. PacketCourt's deterministic evidence engine
+remains responsible for final verdicts and calculations.

src/packetcourt/audit.py

@@ -23,6 +23,17 @@ ADDED_SUGAR_TERMS = {
     "sucrose",
 }
 
+SWEETENER_TERMS = {
+    "sucralose",
+    "aspartame",
+    "acesulfame",
+    "saccharin",
+    "stevia",
+    "maltitol",
+    "sorbitol",
+    "erythritol",
+}
+
 
 def _ingredient_evidence(ingredients: list[str], matches: list[str]) -> list[Evidence]:
     return [Evidence(source="ingredient list", text=item) for item in matches]
@@ -31,6 +42,31 @@ def _ingredient_evidence(ingredients: list[str], matches: list[str]) -> list[Evi
 def _audit_claim(claim: str, back_text: str, ingredients: list[str], nutrition) -> ClaimAudit:
     lowered_ingredients = [item.lower() for item in ingredients]
 
+    if claim == "Sugar Free":
+        sweeteners = [
+            original
+            for original, lowered in zip(ingredients, lowered_ingredients)
+            if any(re.search(rf"\b{re.escape(term)}\b", lowered) for term in SWEETENER_TERMS)
+        ]
+        if nutrition.total_sugar_g is not None:
+            return ClaimAudit(
+                claim=claim,
+                verdict=Verdict.CONTEXT_MISSING,
+                summary="A sugar quantity is visible, but sugar-free claim compliance depends on the declared basis and applicable product rules.",
+                evidence=[Evidence(source="nutrition panel", text=f"Total sugar {nutrition.total_sugar_g:g}g ({nutrition.basis})")]
+                + _ingredient_evidence(ingredients, sweeteners),
+                caveat="Sweeteners are surfaced as context; their presence does not itself contradict a sugar-free claim.",
+                confidence="medium",
+            )
+        return ClaimAudit(
+            claim=claim,
+            verdict=Verdict.CANNOT_VERIFY,
+            summary="The packet claims sugar free, but no readable total-sugar quantity and measurement basis were supplied.",
+            evidence=[Evidence(source="front claim", text=claim)] + _ingredient_evidence(ingredients, sweeteners),
+            caveat="A readable nutrition panel is required. Sweeteners are relevant context but are not sugar.",
+            confidence="low",
+        )
+
     if claim == "No Added Sugar":
         matches = [
             original
@@ -185,6 +221,24 @@ def _audit_claim(claim: str, back_text: str, ingredients: list[str], nutrition)
                 confidence="high",
             )
 
+    meaningful_terms = [
+        token.lower()
+        for token in re.findall(r"[A-Za-z]{3,}", claim)
+        if token.lower() not in {"with", "extra", "real", "enriched", "fortified", "source", "contains"}
+    ]
+    evidence_matches = [
+        item for item in ingredients if any(term in item.lower() for term in meaningful_terms)
+    ]
+    if evidence_matches:
+        return ClaimAudit(
+            claim=claim,
+            verdict=Verdict.CONTEXT_MISSING,
+            summary="Related ingredient evidence is visible, but the packet does not provide enough quantified evidence to verify the full front claim.",
+            evidence=_ingredient_evidence(ingredients, evidence_matches),
+            caveat="PacketCourt will not infer quantity, quality, or nutritional significance from an ingredient name alone.",
+            confidence="medium",
+        )
+
     return ClaimAudit(
         claim=claim,
         verdict=Verdict.CANNOT_VERIFY,
@@ -240,6 +294,28 @@ def _persuasion_gap(claims: list[ClaimAudit], ingredients: list[str], whole_pack
                 evidence=[Evidence(source="claim interpretation", text="FSSAI registration is not a health score.")],
             )
         )
+    if "Sugar Free" in claim_names:
+        sweeteners = [
+            item
+            for item in ingredients
+            if any(re.search(rf"\b{re.escape(term)}\b", item, re.IGNORECASE) for term in SWEETENER_TERMS)
+        ]
+        maltodextrin = next((item for item in ingredients if "maltodextrin" in item.lower()), None)
+        if sweeteners or maltodextrin:
+            context = []
+            if maltodextrin:
+                context.append(f"the ingredient list begins with or includes “{maltodextrin}”")
+            if sweeteners:
+                context.append(f"sweetener evidence includes “{', '.join(sweeteners)}”")
+            findings.append(
+                PersuasionFinding(
+                    headline="Sugar free does not mean ingredient-context free.",
+                    front_impression="The front emphasizes the absence of sugar.",
+                    quiet_context="; ".join(context).capitalize() + ".",
+                    severity="medium",
+                    evidence=_ingredient_evidence(ingredients, ([maltodextrin] if maltodextrin else []) + sweeteners),
+                )
+            )
     return findings
 
 

src/packetcourt/investigator.py

@@ -5,6 +5,7 @@ from .models import InvestigationPlan, InvestigationStep
 
 
 POLICY_TOOLS = {
+    "Sugar Free": "inspect_nutrition",
     "No Added Sugar": "inspect_ingredients",
     "Multigrain": "inspect_ingredients",
     "100% Natural": "apply_safety_boundary",
@@ -16,6 +17,16 @@ POLICY_TOOLS = {
     "High Protein": "inspect_nutrition",
 }
 
+def policy_tool_for(claim: str) -> str:
+    if claim in POLICY_TOOLS:
+        return POLICY_TOOLS[claim]
+    lowered = claim.lower()
+    if any(term in lowered for term in ("calcium", "dha", "protein", "sugar", "fat", "sodium")):
+        return "inspect_nutrition"
+    if any(term in lowered for term in ("real", "with", "contains", "ingredient", "grain", "natural")):
+        return "inspect_ingredients"
+    return "inspect_label_evidence"
+
 
 def build_investigation(
     claim_names: list[str],
@@ -31,7 +42,7 @@ def build_investigation(
     for claim in claim_names:
         routed_tool, source = route_claim(claim)
         router_model = source if source != "deterministic fallback" else router_model
-        tool = routed_tool or POLICY_TOOLS[claim]
+        tool = routed_tool or policy_tool_for(claim)
         if tool in seen:
             continue
         seen.add(tool)
@@ -44,14 +55,14 @@ def build_investigation(
             )
         )
 
-    if claim_names and not ingredients and any(POLICY_TOOLS[name] == "inspect_ingredients" for name in claim_names):
+    if claim_names and not ingredients and any(policy_tool_for(name) == "inspect_ingredients" for name in claim_names):
         missing.append("A readable ingredient list")
-    if claim_names and nutrition.basis == "unknown" and any(POLICY_TOOLS[name] == "inspect_nutrition" for name in claim_names):
+    if claim_names and nutrition.basis == "unknown" and any(policy_tool_for(name) == "inspect_nutrition" for name in claim_names):
         missing.append("A readable nutrition panel with its measurement basis")
     if expiry.instruction and not expiry.packed_on:
         missing.append("The packing or manufacturing date needed to resolve relative shelf life")
 
-    if expiry.best_before or expiry.instruction or expiry.after_opening_instruction:
+    if expiry.best_before or expiry.instruction or expiry.after_opening_instruction or expiry.visible_date_texts:
         steps.append(
             InvestigationStep(
                 tool="resolve_dates",
@@ -59,6 +70,8 @@ def build_investigation(
                 status="completed" if expiry.best_before or expiry.after_opening_instruction else "needs evidence",
             )
         )
+    if expiry.visible_date_texts and not expiry.best_before:
+        missing.append("Labels identifying the visible dates as packed, manufactured, best-before, or expiry")
 
     stop_reason = (
         "Stopped with explicit missing-evidence requests."

src/packetcourt/models.py

@@ -62,6 +62,7 @@ class ExpiryInfo(BaseModel):
     best_before: str | None = None
     instruction: str | None = None
     after_opening_instruction: str | None = None
+    visible_date_texts: list[str] = Field(default_factory=list)
     status: str = "Not enough label evidence"
 
 

src/packetcourt/parser.py

@@ -9,6 +9,7 @@ from .models import ExpiryInfo, NutritionFacts, WholePacketNutrition
 
 CLAIM_PATTERNS: list[tuple[str, str]] = [
     ("High Protein", r"\bhigh[\s-]*protein\b"),
+    ("Sugar Free", r"\bsugar[\s-]*free\b"),
     ("No Added Sugar", r"\bno[\s-]*added[\s-]*sugar\b"),
     ("Multigrain", r"\bmulti[\s-]*grain\b"),
     ("100% Natural", r"\b100\s*%\s*natural\b"),
@@ -19,6 +20,12 @@ CLAIM_PATTERNS: list[tuple[str, str]] = [
     ("Whole Grain", r"\b(?:made\s+with\s+)?whole[\s-]*grains?\b"),
 ]
 
+GENERIC_CLAIM_TERMS = re.compile(
+    r"\b(?:free|extra|real|enriched|fortified|rich|source|natural|healthy|safe|pure|with)\b",
+    re.IGNORECASE,
+)
+NON_CLAIM_LINES = re.compile(r"\b(?:nutraceutical|brand|flavour|flavor)\b", re.IGNORECASE)
+
 
 def normalize_space(text: str) -> str:
     return re.sub(r"\s+", " ", text or "").strip()
@@ -26,7 +33,19 @@ def normalize_space(text: str) -> str:
 
 def extract_claims(front_text: str) -> list[str]:
     text = normalize_space(front_text).lower()
-    return [name for name, pattern in CLAIM_PATTERNS if re.search(pattern, text)]
+    claims = [name for name, pattern in CLAIM_PATTERNS if re.search(pattern, text)]
+    matched_patterns = [pattern for _, pattern in CLAIM_PATTERNS]
+    for raw_line in (front_text or "").splitlines():
+        line = normalize_space(raw_line).strip(" |•*-")
+        if (
+            3 <= len(line) <= 100
+            and GENERIC_CLAIM_TERMS.search(line)
+            and not NON_CLAIM_LINES.search(line)
+            and not any(re.search(pattern, line, re.IGNORECASE) for pattern in matched_patterns)
+            and line.lower() not in {claim.lower() for claim in claims}
+        ):
+            claims.append(line)
+    return claims
 
 
 def _number_after(label: str, text: str, unit: str) -> float | None:
@@ -89,7 +108,8 @@ def calculate_whole_packet(nutrition: NutritionFacts) -> WholePacketNutrition:
 
 def extract_ingredients(back_text: str) -> list[str]:
     match = re.search(
-        r"\bingredients?\s*:\s*(.+?)(?=\b(?:nutrition|allergen|contains|best before|mfd|pkd|manufactured|packed)\b|$)",
+        r"\bingredients?\s*:\s*(.+?)(?=\b(?:nutrition|allergen|contains|net\s*(?:weight|wt)|storage|directions?"
+        r"|after[\s-]*opening|best before|mfd|pkd|manufactured|packed|fssai|dates?|unit sale price)\b|$)",
         normalize_space(back_text),
         re.IGNORECASE,
     )
@@ -135,6 +155,17 @@ def _add_months(value: date, months: int) -> date:
 
 def parse_expiry(back_text: str) -> ExpiryInfo:
     text = normalize_space(back_text)
+    visible_dates = list(
+        dict.fromkeys(
+            match.group(0).upper()
+            for match in re.finditer(
+                r"\b(?:\d{1,2}[\s/\-.]+)?(?:JAN|FEB|MAR|APR|MAY|JUN|JUL|AUG|SEP|OCT|NOV|DEC)"
+                r"(?:[A-Z]*)[\s/\-.]+\d{2,4}\b",
+                text,
+                re.IGNORECASE,
+            )
+        )
+    )
     packed_match = re.search(
         r"\b(?:pkd|packed(?:\s+on)?|mfd|manufactured(?:\s+on)?)\s*[:\-]?\s*"
         r"(\d{1,2}[\s/\-.]+(?:\d{1,2}|[A-Za-z]{3,9})[\s/\-.]+\d{2,4})",
@@ -164,6 +195,11 @@ def parse_expiry(back_text: str) -> ExpiryInfo:
         status = f"Best-before evidence resolves to {best_before.isoformat()}"
     elif instruction and not packed:
         status = "Relative shelf-life found, but the starting date is missing"
+    elif visible_dates:
+        status = (
+            f"Visible date evidence found ({', '.join(visible_dates)}), but the label does not identify "
+            "which date is packing, manufacture, or expiry."
+        )
     else:
         status = "No resolvable best-before date found"
 
@@ -177,5 +213,6 @@ def parse_expiry(back_text: str) -> ExpiryInfo:
         best_before=best_before.isoformat() if best_before else None,
         instruction=instruction,
         after_opening_instruction=after_opening_match.group(0) if after_opening_match else None,
+        visible_date_texts=visible_dates,
         status=status,
     )

tests/test_audit.py

@@ -88,3 +88,26 @@ def test_investigation_requests_missing_evidence_and_stops_explicitly():
     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"
+
+
+def test_sugar_free_packet_surfaces_sweetener_and_requests_nutrition_panel():
+    result = audit_packet(
+        "Sugar Free\nEnriched with Extra Calcium with DHA\nReal Badam",
+        (
+            "Ingredients: Maltodextrin (65%), Badam, Soya Protein Isolate, Sucralose, "
+            "Vitamins and Mineral Mix. Net Weight: 200g. Dates: FEB 2024 - JUL 2025."
+        ),
+    )
+    assert by_claim(result, "Sugar Free").verdict == Verdict.CANNOT_VERIFY
+    assert any("Sucralose" in evidence.text for evidence in by_claim(result, "Sugar Free").evidence)
+    assert any("sugar free" in finding.headline.lower() for finding in result.persuasion_gap)
+    assert result.ingredients[-1] == "Vitamins and Mineral Mix"
+    assert result.expiry.visible_date_texts == ["FEB 2024", "JUL 2025"]
+    assert any("visible dates" in item.lower() for item in result.investigation.missing_evidence)
+
+
+def test_dynamic_front_claim_is_audited_instead_of_dropped():
+    result = audit_packet("Real Badam", "Ingredients: Maltodextrin, Badam. Net Weight: 200g.")
+    claim = by_claim(result, "Real Badam")
+    assert claim.verdict == Verdict.CONTEXT_MISSING
+    assert any("Badam" in evidence.text for evidence in claim.evidence)

vision_space/README.md

@@ -24,4 +24,3 @@ transcribe only visible evidence from front and back food-package photos.
 
 The main PacketCourt product applies deterministic claim auditing, arithmetic,
 and refusal rules after transcription.
-

vision_space/requirements.txt

@@ -5,4 +5,3 @@ pillow>=11.0.0
 torch>=2.8.0
 torchvision>=0.23.0
 transformers>=5.7.0
-