Codicil

Financial Disclosure Auditor — Cross-Temporal Contradiction Detector

Codicil reads years of SEC 10-K filings and finds specific, cited instances where management's forward-looking statements in one filing directly contradict what they disclosed had actually occurred in the next. If it cannot quote verbatim from both filings, it outputs nothing.

Built for the lablab.ai × AMD Developer Hackathon 2026.

Live demo: nurusyda-codicil.hf.space

---

What it does

Most AI tools summarise documents. Codicil audits them.

For every consecutive pair of annual filings, the model is given the earlier filing's forward-looking language and the later filing's disclosure of actual outcomes. It is instructed to produce a finding only if it can extract a verbatim quote from each — the earlier projection and the later reality — and explain the contradiction. No inference. No sentiment scores. No hallucinations.

Verified batch run — 10 major US banks, 10 years of 10-K filings:

Bank	Finding	Documents
TFC	Contradiction between forward-looking guidance and actual results for FDIC DIF assessment premiums	Amendment No. 64 (2023) → No. 77 (2024)
TFC	Contradictory statements regarding the effective end date of the Stress Capital Buffer	Amendment No. 99 (2026) → No. 107 (2026)
TFC	Contradiction regarding the regulatory deadline for submitting the first resolution plan under the Tailoring Rules	Amendment No. 34 (2020) → No. 45 (2021)
BAC	Contradiction in stated regulatory submission frequency for resolution plans	Amendment No. 2 (2017) → No. 14 (2018)
GS	Contradiction in stated regulatory disclosure frequency for ESG risks under the CRR	Amendment No. 3 (2023) → No. 16 (2024)
MS	Contradiction in the stated expected compliance date for the final phase of initial margin rules	Amendment No. 3 (2020) → No. 15 (2021)

Total batch time: ~1,854 seconds. Total GPU compute cost: ~$0.65.

---

Why AMD MI300X

The MI300X's 192GB HBM3 memory is load-bearing, not decorative.

Qwen3.6-35B-A3B at a 131K token context window requires GPU memory that a standard H100 cannot hold at this utilisation level. The MI300X enables loading the full model with headroom for long-context batches, making multi-year cross-filing analysis possible in a single inference pass. No chunking. No retrieval errors. That is the product's core guarantee.

Batch across 10 tickers runs via vLLM's PagedAttention on ROCm, with efficient KV-cache reuse across concurrent requests.

---

Architecture

SEC EDGAR
    │
    ▼
scripts/ingest_edgar.py          ← downloads 10-K filings, extracts thematic sections
    │                               groups: capital_liquidity, risk_language,
    │                               covenant_legal, management_guidance
    ▼
scripts/codicil_inference.py     ← sends section chain to Qwen3.6-35B on MI300X
    │                               auditor prompt: cite-or-be-silent discipline
    │                               output: Earlier quote + Later quote + reasoning
    ▼
scripts/batch_runner.py          ← loops ingest + inference across multiple tickers
    │                               writes data/batch_results/batch_live_*.json
    ▼
scripts/sector_compare.py        ← aggregates batch results into sector_summary.json
    │                               sorted by finding count descending
    ▼
scripts/app.py                   ← Pure Python HTTP server UI
                                    Analyze tab: single-company live analysis
                                    Sector tab: 10-bank comparison dashboard
                                    Gap detection: instant, GPU-free
                                    Offline fallback: demo works without MI300X

---

Running locally

Requirements: Python 3.11, a running vLLM server pointed at Qwen3.6-35B-A3B

git clone https://github.com/nurusyda/codicil
cd codicil
python -m venv env && source env/bin/activate
pip install -r requirements.txt

Set the inference server URL:

export LLAMA_SERVER_URL=http://<your-mi300x-ip>:8000/v1

Run the app:

python scripts/app.py
# Open http://127.0.0.1:7860

GPU-free demo mode: The Sector tab loads from data/sector_summary.json without inference. The Analyze tab's gap detection also runs without GPU. Only the Analyze → Run inference step requires the MI300X.

---

Running the batch

python scripts/batch_runner.py \
  --tickers JPM BAC WFC C GS MS USB PNC TFC COF \
  --form-type 10-K \
  --limit 10

python scripts/sector_compare.py
# Output: data/sector_summary.json

Dry-run (no inference, verifies ingestion only):

python scripts/batch_runner.py --tickers JPM BAC --form-type 10-K --limit 5 --dry-run

---

Serving the model (AMD MI300X)

docker start rocm && docker exec -it rocm /bin/bash

vllm serve Qwen/Qwen3.6-35B-A3B \
  --dtype bfloat16 \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.92 \
  --enable-prefix-caching \
  --reasoning-parser qwen3 \
  --port 8000 \
  --host 0.0.0.0

Turn off the GPU immediately after batch completes. Verify "Offline" in AMD portal — not just poweroff from host.

---

Scripts

Script	Purpose
scripts/app.py	Main UI — FastAPI + WebSocket + all API endpoints
scripts/ingest_edgar.py	SEC EDGAR downloader and section extractor
scripts/codicil_inference.py	Auditor prompt + vLLM inference client
scripts/batch_runner.py	Multi-ticker batch orchestrator
scripts/sector_compare.py	Aggregates batch results into sector summary
scripts/risk_factor_diff.py	Risk factor disappearance detector (year-over-year diff)
cross_check.py	Cross-module import verifier — run before every commit (repo root, not scripts/)

---

Stack

• Model: Qwen3.6-35B-A3B (Alibaba Cloud)
• Inference: vLLM on ROCm (AMD MI300X via AMD Developer Cloud)
• Embeddings: BAAI/bge-m3 via sentence-transformers (local_files_only=True)
• Vector store: ChromaDB
• Data source: SEC EDGAR via sec-edgar-downloader
• Backend: Python / http.server (stdlib)
• Frontend: Vanilla JS served by http.server
• Hosting: Hugging Face Spaces

---

Disclaimer

Codicil surfaces candidates for attorney or analyst review. It does not replace legal judgment or investment advice. Every flagged finding requires human verification. Output is based on public SEC filings only.

---

Built May 2026. AMD Developer Hackathon submission.