GitLab Transcend · Path B · Assisted

Kinwise Carbon Budget

Every scam check has a carbon cost. Now it is measured, and cut, one approved change at a time.
Kinwise screens the messages an older parent receives for scams. A GitLab Duo flow proposes greener changes, a person approves each step, and GitLab CI proves that no scam slips through.
GitLab Duo Agent Platform flows GitLab CI carbon and accuracy gate SCI · ISO/IEC 21031:2024
Carbon gate: PASS
SCI per scam check, v132.4 mg CO2e
After 3 approved changes13.9 mg CO2e
Change-57%
Scams caught26 / 26 → 26 / 26
False alarms0 / 26 → 0 / 26
Measured in GitLab CI pipelines #2982428704 and #2984892865, 52 labelled messages each
The problem

Scammers paraphrase, so a model has to read every message. Nobody sees what that costs.

$7.7B
lost to internet crime by Americans 60 and older in 2025, about 60% more than in 2024
FBI IC3 2025 report, via AARP, Apr 16 2026
8 of 26
scams in our labelled set caught by the original rule-based detectors alone
eval/messages.jsonl, all fictional
32 mg
CO2e per scam check once every message went to the model (v1), on every text, email and voicemail
SCI, GSF method, central estimate
Who it is for: families who look after an older parent from a distance, and the team that runs the screen, who now see the carbon and the recall of every change before it ships.
How it works on GitLab

Agents propose and prepare. People decide every step.

1 · Measure
Carbon gate on every MR
Replays 52 labelled messages with real model calls, computes SCI per check, checks recall, posts a table, fails if worse.
2 · Propose
green-reviewer flow
Reads the measured numbers and the code, opens up to three proposals with the arithmetic, then stops.
A person asks
3 · Implement
green-implementer flow
Checks the approver allowlist, edits only the named files, opens a merge request, then stops.
A person approves one
4 · Verify
The MR proves its effect
Lint, types, tests, build, SAST, secret detection, policy guard and the gate: main vs this MR.
A person merges
5 · Release
Deploy, then the ledger
Vercel deploy, smoke test, and the SCI of the whole change, including the agents' own tokens.
A person presses Play
SCI = (E x I + M) per R. Measured tokens, calls, bytes and seconds; sourced coefficients with low, central and high values; recomputed with the Green Software Foundation's Impact Framework on every run.
Measured impact

57% less carbon per scam check, and still every scam caught.

Per scam checkv1now
SCI, mg CO2e (central)32.413.9
Bounding range4.4 to 4711.5 to 238
Model calls1.000.71
Prompt tokens891510
Completion tokens20559
Scams caught26 / 2626 / 26
Greener but worse fails. Dropping the model would cut carbon by 100%, but only 8 of 26 scams were caught. The gate refused it (MR !5).
The agents count too. All 24 pipelines, 17 evals and 4 flow sessions cost 82.9 g CO2e. That is repaid after 4,472 checks: under 4 hours for 1,000 people.
Policy as code. Budget, recall floor and approver allowlist live in the repo; agent-written MRs cannot change them.
Pipelines #2982428704 (v1) and #2984892865 (now) · method and sources: docs/SCI.md · gitlab.com/gitlab-ai-hackathon/transcend-october-2026/43536883/hack-gl-kinwise