docs: document verified Nemotron reviewer
Browse files
README.md
CHANGED
|
@@ -35,6 +35,10 @@ The system deliberately separates three responsibilities:
|
|
| 35 |
The models can read and route an investigation. They cannot silently invent a
|
| 36 |
nutrition value or override the evidence standard.
|
| 37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
## What the investigation agent does
|
| 39 |
|
| 40 |
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
|
|
| 62 |
measured result is recorded in the model card and its suggestions remain
|
| 63 |
bounded by deterministic policy fallbacks.
|
| 64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
## Persuasion Gap
|
| 66 |
|
| 67 |
Claim verification alone was not enough. A `HIGH PROTEIN` claim can be
|
|
@@ -79,6 +92,7 @@ the decision with the user.
|
|
| 79 |
- `10` transparent investigation traces are exported.
|
| 80 |
- The vision model has `1.30B` parameters.
|
| 81 |
- The fine-tuned evidence router has approximately `4.4M` parameters.
|
|
|
|
| 82 |
- The complete product interface is responsive and built on Gradio.
|
| 83 |
|
| 84 |
## What PacketCourt refuses to claim
|
|
|
|
| 35 |
The models can read and route an investigation. They cannot silently invent a
|
| 36 |
nutrition value or override the evidence standard.
|
| 37 |
|
| 38 |
+
NVIDIA `Nemotron-Mini-4B-Instruct` performs a second, independent review after
|
| 39 |
+
the investigation plan completes. It can identify the highest-priority missing
|
| 40 |
+
evidence, but it cannot alter PacketCourt's deterministic verdict.
|
| 41 |
+
|
| 42 |
## What the investigation agent does
|
| 43 |
|
| 44 |
Each packet creates a claim-dependent investigation plan. A `NO ADDED SUGAR`
|
|
|
|
| 66 |
measured result is recorded in the model card and its suggestions remain
|
| 67 |
bounded by deterministic policy fallbacks.
|
| 68 |
|
| 69 |
+
## A failed Nemotron deployment
|
| 70 |
+
|
| 71 |
+
The first NVIDIA reviewer used `NVIDIA-Nemotron-3-Nano-4B-BF16`. A real
|
| 72 |
+
ZeroGPU probe failed because the hybrid Mamba runtime required a specialized
|
| 73 |
+
CUDA build unavailable in the standard Gradio image. Rather than claim a model
|
| 74 |
+
that did not run, PacketCourt switched to `Nemotron-Mini-4B-Instruct`. The
|
| 75 |
+
replacement completed a real ZeroGPU evidence-gap review before it was
|
| 76 |
+
connected to the product.
|
| 77 |
+
|
| 78 |
## Persuasion Gap
|
| 79 |
|
| 80 |
Claim verification alone was not enough. A `HIGH PROTEIN` claim can be
|
|
|
|
| 92 |
- `10` transparent investigation traces are exported.
|
| 93 |
- The vision model has `1.30B` parameters.
|
| 94 |
- The fine-tuned evidence router has approximately `4.4M` parameters.
|
| 95 |
+
- The independent NVIDIA Nemotron reviewer has approximately `4B` parameters.
|
| 96 |
- The complete product interface is responsive and built on Gradio.
|
| 97 |
|
| 98 |
## What PacketCourt refuses to claim
|