SCI MODEL EXPLORER

Make the assumptions visible.

Compare retaining an unused legacy process with retiring it on the same fixed-size infrastructure. Explore the allocation model, inspect the real local observations, and see what the evidence supports.

MODELED REFERENCE · NO SUPPORTED SAVINGS

VM capacity stays allocated.

Both scenarios share the same demand, infrastructure and completed migration history. Stopping one process does not release the VM’s allocated memory or disk. Runtime allocations are equal by assumption; the incremental proof-verifier is added separately. This model does not prove that actual energy use is identical.

Loading the published model and its sources…

Scenario assumptions

These are adjustable model inputs. They are not measurements of the hosted deployment.

CPU utilization, both scenarios 50%

CCF utilization sensitivity. The same value is applied to both conditions.

Effective controller vCPUs

e2-medium exposes two vCPUs with one sustained fractional vCPU. This is an allocation choice.

Cloud Run active fraction 100%

Default: one continuously active instance-hour. Actual serving time is not measured. Zero active time does not establish zero infrastructure footprint.

Completed deliveries per hour

Reference demand is identical on both sides. It is not the observed production throughput.

Shared migration amortization

The prior migration’s modeled CI and inference costs are allocated to both scenarios.

Include retirement-proof verifier

The actual proof-verifier job is an incremental modeled cost after retirement. Migration CI remains shared.

Inference calls per migration 40

Explicit scenario count, not measured Duo calls. Tool actions are not inference calls.

Published inference-energy proxy

Default: Google’s full-serving-stack Gemini median, not actual Duo energy. Other literature proxies have different boundaries, including GPU-only energy. Missing serving CPU, memory, PUE and embodied/training costs remain in K. No second PUE is applied to the Google proxy.

The paired comparison

Calculated in your browser · read-only

LEGACY RETAINED

— + K

conditional reference range · gCO₂e per 100 completed deliveries

LEGACY PROCESS RETIRED

— + K

conditional reference range · gCO₂e per 100 completed deliveries

—

Modeled change after retirement · gCO₂e / 100 deliveries

No supported savings. Runtime allocations and prior migration costs are shared; the optional proof-verifier adds a conditional modeled cost.

Runtime capacity is identical. Incremental proof-verification cost is shown separately. The displayed ranges reflect CI regional scenarios, not statistical confidence bounds.

K is shared unquantified carbon, not zero. It includes unquantified shared network, model-training/embodied and serving-infrastructure residuals per 100 deliveries. It cancels only in this specific post-migration comparison. Additional unquantified after-only verification costs are not included in the displayed increment. These numeric references are not complete measured absolute SCI scores.

What the numeric reference includes

Hourly allocation, before and after. Operational energy uses the published power coefficients; embodied values use an explicitly representative platform model.

Modeled component	Reference gCO₂e / hour

Why CI and Duo appear on both sides

The baseline already has a migrated successor with an unused legacy process retained. The proposed condition retires that process. Duo, MR and migration CI have therefore happened in both conditions. Charging those costs only to one side would change the question and invalidate the pairing. This comparison does not establish the net benefit of adopting ZeroScream instead of having no tool.

SEPARATE LOCAL OBSERVATIONS

Measured evidence stays separate.

Resident memory

—

Lower mean RSS in the measured local Node processes, across 1,000 verified deliveries. Freed RSS is not released cloud VM capacity.

CPU-package energy

—

These Windows observations are not used as Iowa cloud energy measurements. The package study includes other host activity and does not measure DRAM, display, storage, remote CI or Duo. The confidence interval crosses zero.

A conditional model, with a narrow conclusion

SCI = (operational carbon + allocated embodied carbon) / 100 completed deliveries
Operational carbon = modeled kWh × location-based grid intensity
Runtime allocation difference = 0 under the same-capacity assumptions
After − before = optional modeled verifier allocation
Absolute reference = numeric modeled components + shared residual K

The Google 2025 annual Iowa grid factor is a proxy for runtime and inference scenarios, not an October 2026 hourly reading. CI uses the published us-east5 to us-east1 regional range; its actual location is unknown. Carbon-free-energy percentages are not applied as a second discount. Cloud Run and runner hardware embodied values are representative proxies, not provider-specific LCAs. No external carbon certification or environmental-prize eligibility is asserted.

What would support a stronger conclusion?

A comparable application boundary with the same useful work in both conditions.
Measured or justified modeled energy for every significant component, including idle provisioned hardware and remote services.
Appropriate electricity factors, hardware lifecycle allocation and an explicit amortization period.
A supported reduction that remains credible under uncertainty. A lower sample mean alone is insufficient.

Published inputs and evidence

The adjacent model JSON records the chosen coefficients and links. Values are assumptions unless explicitly identified as local observations.

Inspect the model JSON · Read the measurement study