Carbon Tracking Log — PUMA Experiments

PUMA tracks gCO₂eq for every experimental condition using CodeCarbon.

Overview

This log is the first systematic PM+LLM carbon measurement dataset in the literature.

---

Why Carbon Matters in PUMA

PUMA is the first PM+LLM benchmark to include carbon measurement as a primary experimental variable — not an afterthought. This follows Strubell et al. (2019) methodology and responds to the identified gap that no existing PM+LLM paper measures environmental cost.

The hypothesis: local quantized models (4-bit, CPU) have dramatically lower carbon footprint than cloud API models, making them environmentally preferable when quality is sufficient.

---

CodeCarbon Setup

from codecarbon import EmissionsTracker
 
# Per-experiment tracking (not global)
tracker = EmissionsTracker(
    project_name=f"puma-{model}-{strategy}-{task}",
    output_dir="./results/carbon/",
    output_file=f"emissions_{model}_{strategy}_{task}.csv",
    log_level="error",
    country_iso_code="ESP",  # Spain — EU grid mix
    save_to_file=True
)
tracker.start()
# ... inference calls ...
emissions_kg = tracker.stop()  # returns kg CO2 eq
emissions_gco2 = emissions_kg * 1000  # convert to gCO2 eq

---

Results Log (to fill during F2–F4)

Experiment	Model	Strategy	Task	gCO₂eq	Duration (s)	kWh	Date
Pending F2	llama3.2:8b	zero-shot	triage	TBD	TBD	TBD	—
Pending F2	llama3.2:8b	few-shot-3	triage	TBD	TBD	TBD	—
Pending F2	llama3.2:8b	few-shot-6	triage	TBD	TBD	TBD	—
Pending F2	llama3.2:8b	cot	triage	TBD	TBD	TBD	—
Pending F2	mistral:7b	zero-shot	triage	TBD	TBD	TBD	—
Pending F2	mistral:7b	few-shot-3	triage	TBD	TBD	TBD	—
Pending F2	mistral:7b	few-shot-6	triage	TBD	TBD	TBD	—
Pending F2	mistral:7b	cot	triage	TBD	TBD	TBD	—
Pending F3	llama3.2:8b	zero-shot	estimation	TBD			—
[…continue…]

---

Reference Values (Literature)

System	Approx. CO₂	Source
GPT-3 training	552,000 kg	Strubell et al. (2019)
Single GPT-4 API call	~0.002–0.02 kg	Estimates
PUMA full experiment (est.)	< 0.050 kg total	Target

---

Analysis Plan (Chapter 4)

• Carbon per condition (bar chart, sorted by gCO₂eq)
• Carbon vs F1 tradeoff frontier (scatter plot)
• Carbon vs MAE tradeoff frontier
• Total PUMA experiment carbon vs estimated GPT-4 equivalent
• Recommendation: “minimum carbon condition achieving MVP threshold”

---

🔗 Related Notes

PN-KeyConcepts-Agents-Reproducibility-RedTeam (Reproducibility section) | LN-Tools-Dev-Stack PR-PUMA-Ch4-Results | SP-Architecture (§CodeCarbon layer) PN-IssueTriage-StoryPoints — tasks tracked | EX-Hypotheses-H1-H2 — conditions tracked MOC-PUMA-Master | MOC-Methods-Frameworks

---

---

id: PRISMA-Log title: “PRISMA Screening Log” type: log tags: [slr, prisma, screening, log, literature] created: 2026-03-01

PRISMA Screening Log

Transparent record of all screening decisions for the PUMA SLR. Every AI-assisted decision is documented (PRISMA-trAIce compliance).

---

Identification Log

Date	Database	Search String (abbreviated)	N returned	N forwarded
TBD	arXiv	LLM benchmark PM estimation	TBD	TBD
TBD	IEEE Xplore	LLM issue triage 2022-2026	TBD	TBD
TBD	ACM DL	LLM story point effort	TBD	TBD
TBD	Semantic Scholar (API)	LLM SE benchmark reproducible	TBD	TBD
TBD	Google Scholar	Local LLM PM benchmark	TBD	TBD

Total identified: TBD After deduplication: TBD

---

Screening Decisions (Title/Abstract)

Zotero Key	Title (abbreviated)	AI suggest	Human decision	Reason (if excluded)
Tawosi2024	CoGEE: Story Point Estimation	Include	Include	—
Angermeir2025	Reproducibility LLM Studies SE	Include	Include	—
Tawosi2022	TAWOS Dataset	Include	Include	—
Ortu2015	Jira Social Repository	Include	Include	—
Manzoor2025	AI in PM 2019-2024	Include	Include	—
[add all screened papers]

---

AI-Assisted Screening (PRISMA-trAIce)

Date	Tool	Task	N processed	N uncertain	Validation done
TBD	Elicit	Abstract screening batch 1	TBD	TBD	Human reviewed all uncertain + 20% sample

---

Eligibility Decisions (Full Text)

Zotero Key	Decision	Reason
Tawosi2024	Include	Full paper verified, TAWOS dataset, GPT-4, MAE reported
Angermeir2025	Include	Full paper verified, 85 papers meta-study, ICSE 2026
[add all]

---

Final PRISMA Counts

Stage	N
Identified	TBD
After deduplication	TBD
Screened (title/abstract)	TBD
Excluded at screening	TBD
Full-text reviewed	TBD
Excluded at full-text	TBD
Included	TBD (target ≥40)