README.md

---
license: cc-by-4.0
tags:
  - build-small-hackathon
  - field-notes
  - packetcourt
---

# Field Notes: Building PacketCourt

## The packet takes the stand

PacketCourt began with a narrow household problem: a food packet's front is
designed to persuade, while the evidence needed to interpret that persuasion
is scattered across the back. A shopper should not need to understand serving
bases, ingredient ordering, date arithmetic, or regulatory language while
standing in a grocery aisle.

The first idea was a nutrition scanner. That was too broad and too easy to turn
into an unexplained health score. PacketCourt instead asks one auditable
question:

> Does the evidence printed on this packet support the impression created by
> its front?

## Small models as witnesses, not judges

The system deliberately separates three responsibilities:

1. OpenBMB MiniCPM-V-4.6 transcribes visible front and back label evidence.
2. A fine-tuned 4.4M-parameter PacketCourt router selects the evidence tools
   required by each detected claim.
3. Deterministic code performs calculations and produces final verdicts.

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`
claim sends the investigation toward ingredients. `HIGH PROTEIN` requires a
nutrition panel and its measurement basis. `FSSAI APPROVED` requires licensing
evidence and a warning that registration is not a health endorsement.

The agent stops in one of two explicit states:

- all evidence tools required by the detected claims completed; or
- required evidence is missing, so the audit returns a concrete request rather
  than guessing.

Every plan, tool decision, evidence extraction, calculation, verdict, and
limitation is exported as a trace.

## Real packets broke the two-photo assumption

A real packet quickly showed that front-and-back capture was too neat an
abstraction. Claims, dates, directions, ingredients, and nutrition tables can
wrap across several panels. PacketCourt now accepts additive multi-angle phone
capture, labels each transcription by photo number, merges unique evidence,
and skips exact duplicate transcriptions.

The same test exposed table-style OCR such as `Protein (g) 12` and
`Sodium | mg | 410`. The deterministic parser was expanded to recover those
rows while explicitly explaining when OCR found a nutrition basis and packet
size but omitted the nutrient quantities needed for arithmetic.

## A failed first fine-tune

The first evidence-router training run reached only `0.40` held-out accuracy.
The dataset was too small and its random split did not preserve every routing
class. That model was published privately but was not enabled in the product.

The corrected run balanced claim variants across five routing classes and used
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
technically supportable while a full packet also contains substantial sugar or
sodium. PacketCourt therefore calculates a **Persuasion Gap**: material
back-label context that competes with the impression emphasized on the front.

This is not a health score. The output cites the exact calculation and leaves
the decision with the user.

## Community learning without silent self-training

Users can confirm an audit or submit an evidence-backed correction through the
Community Review Agent. Each review preserves the original evidence,
investigation path, and Nemotron review in a public queue.

Feedback is not automatically trusted. New records remain in
`pending_human_review` with `training_eligible: false` until the supplied packet
evidence is checked. Approved cases can enter a versioned router-training
release and must pass the golden-case suite before deployment.

## Current evidence

- `20` unit and end-to-end integration tests pass.
- `35/35` golden-case checks pass across `10` packet cases.
- `10` transparent investigation traces are exported.
- Multi-angle packet capture and exact-duplicate removal run publicly.
- A public correction-driven learning queue contains real review records.
- 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

PacketCourt does not declare food healthy, safe, illegal, or fraudulent. It
does not treat OCR as ground truth. It does not use an LLM to perform arithmetic
that deterministic code can perform exactly. When supplied evidence is
insufficient, the correct result is `CANNOT VERIFY`.

That refusal is not a missing feature. It is the product's standard of proof.