RPT
8-page-or-less report
Rubric-facing report: problem, external research, analysis, solution, implementation, outcomes, recommendations, and claim boundaries.
Primary written submissionMMC / Google DeepMind AI Research Foundations
ControlSift capstone hub
The complete project in one place: problem, external research, methods, experiment results, implementation, responsible-AI work, program alignment, report, slides, final presentation, reflection, weekly work, and submission package.
Plain-English project: ControlSift studies whether a small language model can distinguish strong cybersecurity control evidence from paperwork that's incomplete, irrelevant, or contradictory. The work includes a hardened synthetic benchmark, classical baselines, prompted Gemma runs, QLoRA, failure analysis, and research-governance artifacts.
Research-integrity noteClassical metrics use dataset v1.1.0; completed Gemma metrics use v1.0.0. The experiments are real and published, but cross-version scores are descriptive rather than a strict same-benchmark leaderboard.
Research artifactComplete
Published runs5 experiment rungs
Classical benchmarkv1.1.0
Gemma receiptsv1.0.0
Program alignmentMMC option 4 / UN SDG 10
Capstone packageComplete
Narrated capstone video
The completed eight-slide research presentation is hosted as the v1.0-capstone release asset.
What the grader needs
Every formal deliverable and administrative checkpoint is visible here.
SLD
8-slide presentation
Exactly eight slides aligned to the final research story and version disclosure.
Presentation artifactVID
Final presentation video
Completed narrated presentation, hosted with the final GitHub release.
CompleteSUB
Submission package
Final bundle map for report, slides, video, links, and supporting evidence.
CHK
GCLP checklist
Requirements crosswalk and GitHub structure.
NOW
Completion status
Final completion state and submission-ready artifact map.
Why the experiment exists and what it found
Start with the problem, then inspect sources, methods, full paper, results, and failure cases.
WHY
Problem and impact
Why relevant paperwork and proof of operation aren't the same thing.
SRC
Research sources
Annotated NIST, Google DeepMind, academic, and UN sources with claim boundaries.
MTH
Methods and evidence
Benchmark design, experiment paths, evidence handling, and integrity controls.
PPR
Research paper
Full technical narrative with final v1.0/v1.1 disclosure and references.
RES
Experiment ledger
Machine-readable metrics surfaced with the version boundary clearly labeled.
FAIL
Failure Lab
Hard boundaries and disagreement patterns hidden by aggregate scores.
How the work was executed and bounded
IMP
Implementation plan
Phases, compute, accounts, resources, risks, mitigations, and completed execution.
RAI
Responsible innovation
Privacy, human authority, synthetic-data boundaries, misuse risks, limitations, and non-claims.
AST
Assurance package
Protocol, Data Card, Model Card, AI risk register, intended use, limitations, and human-review requirements.
REP
Reproduce
Validation and rerun path for the public research artifact.
How ControlSift maps to the residency
SDG
UN SDG alignment
MMC option 4 wording reconciled to official UN SDG 10.
MAP
Curriculum map
Maps AI Research Foundations content to concrete ControlSift work.
REF
Reflection
What changed, what the model results taught, and future work.
LAB
Weekly labs
Residency-facing weekly outputs and work trail.
Published experiment snapshot
Majority / v1.10.067
TF-IDF / v1.10.533
Gemma zero-shot / v1.00.080
Gemma few-shot / v1.00.137
Gemma QLoRA / v1.00.083
Defensible interpretationWithin Gemma v1.0, few-shot is strongest and QLoRA doesn't beat it. The hardened v1.1 classical baseline is about 0.53. A definitive classical-vs-Gemma ranking would require a matching-version experiment and is outside this capstone's scope.