Implementation Plan

Phases, resources, risks, and mitigations for the completed ControlSift capstone.

Phases

Phase 0 / Done

Problem framing, MMC option 4 / UN SDG 10 mapping, synthetic benchmark, assurance artifacts, and capstone hub.

Phase 1 / Done

Classical benchmark hardening and v1.1 majority / TF-IDF baselines.

Phase 2 / Done

Hugging Face access, free GPU path, Gemma v1.0 zero-shot and few-shot evaluation.

Phase 3 / Done

Gemma v1.0 QLoRA train/eval; few-shot remains the strongest Gemma rung.

Phase 4 / Done

Error analysis, Failure Lab, external research sources, report, paper, slides, and final version-boundary disclosure.

Phase 5 / Done

Final narrated presentation completed, published with the v1.0-capstone release, and submission package assembled.

Compute: local CPU for classical work; Kaggle / Colab free GPU path for Gemma experiments.
Models: TF-IDF + logistic regression; Gemma 3 1B; QLoRA with 4-bit NF4.
Platforms: GitHub, GitHub Pages, Hugging Face, Kaggle / Colab.
Data: synthetic evidence packets only; no employer, patient, or customer evidence.
People: solo project with assistive tooling for drafting and engineering support.

Risk	Impact	Mitigation / outcome
Lexical shortcut	Artificially easy benchmark	Hardened the generator before sealing v1.1 classical results.
Overclaiming AI performance	False success narrative	Preserved the QLoRA negative result and low parse-success rate.
Cross-version comparison	Misleading leaderboard	Classical v1.1 and Gemma v1.0 are disclosed separately; cross-version scores are descriptive only.
Synthetic-to-real overgeneralization	Unsafe production interpretation	Explicit non-production scope, Limitations, Intended Use, and human-review requirements.
Secret exposure	Credential compromise	Tokens remain in platform secret stores and are excluded from git.
Format miss	Capstone rejection	Rubric-facing report, exactly eight slides, and the final under-five-minute narrated presentation are complete.