AI for Managers Final Project

Start with 01 Final Team Project Report Human Review.pdf. It uses the provided cover plus four body pages and four appendix pages. It is a review draft, not a completed two-rater submission.

REQUIRED HUMAN ACTIONS
1. Independently score all 45 voter-visible outputs (four criteria 0 to 2 and unsupported-claim count) in 02 Human Review.xlsx or the blank CSV template. Read 03 Pilot Responses Review.html and official PDF reference pages. Keep the condition key and AI judge scores aside until your scores are saved.
2. Export the completed Ratings sheet using the supplied export_review_ratings.py script. Run evaluate.py report with human_ratings.csv in results/colab_lora_v1. Report actual mismatches; do not call the current zero paired denominator zero disagreement.
3. Review factual failures, synthetic labels, candidate excerpt context and editorial topic labels. Confirm the proposed autonomy map, kill criterion and revise recommendation.
4. Fill all team NetIDs on the cover and approve the GenAI disclosure. The report currently lists abbreviated member names and NetIDs pending.
5. Update the report with human results and mismatch counts, inspect the final PDF, and submit that single PDF on Canvas by Thursday October 15 at 11:59 pm. The assignment maximum is nine pages with 11-point body text and one-inch margins. A persistent public application is not required.

PLATFORM ROLES
Local: development, PDF retrieval and untuned 4B CPU demo. Local-only links are not professor endpoints.
Colab: completed Tesla T4 fine-tuning and the 15-case, three-condition experiment. Runtime storage is temporary; the archived artifacts are portable.
Hugging Face: published trained 2B LoRA adapter and recorded evaluation. Paid inference is inactive. The reserved Space has no application file and is not a working endpoint.
Candidate quotes: standalone HTML viewer. Download and open it in a browser; no GPU, local model or Python dependency. Source links require internet access.

MEASURED RESULTS
Independent GenAI judge strict passes: naive untuned 2/15; source-only untuned 3/15; source-only fine-tuned 7/15. Human ratings remain 0/45. One visible unsupported claim was found in the fine-tuned condition. Do not describe the system as ready for public voter use.

CONTENTS
00 Start Here.pdf: links, status and review steps.
01 report PDF and editable text: course deliverable for review.
02 Human Review.xlsx: blank independent ratings and quotation review fields.
03 Pilot Responses Review.html: all frozen answers with reference links and evidence.
04 Candidate Quotes Demo.html: nine exact excerpts with source and date metadata.
05 Test Data and Results.zip: test cases, rubric, expected reference notes, all 45 raw and visible responses, AI judge scores, blank human template, split data and provenance.
06 Prototype and Colab.zip: portable project with source retrieval, app, training scripts, notebook and saved records. Extract with the ballot-guide-lab folder intact for the Colab upload cell. GPU inference still requires pinned dependencies and the base model download.
07 Trained Model Adapter.zip: adapter weights and tokenizer with provenance. Requires the pinned Qwen3.5-2B base; it is not a standalone model. No verified trained GGUF export exists.
08 Links.txt: Colab, Hugging Face model and official PDF URLs.
09 SHA256 Manifest.txt: file checksums.

No private meeting audio, passwords, tokens or course attachments are in this review package. The source articles retain their own rights. The candidate feature is outside the original 15-case AI evaluation.