Big-O Software Carbon Intensity (SCI) Disclosure

① Functional unit (R)

One diagnostic decision = the decision engine (ROC utility patent M683041) completes one rough-set feature reduction plus one multi-criteria decision (MCDM) evaluation. This is what ranks and recommends items in each website diagnosis on bigomorfi.tech.

② Boundary

③ Formula

SCI = ((E × I) + M) ÷ R, expanded as in the worksheet provided by the Taiwan Institute of Economic Research (TIER):

E

Software energy (kWh) = power W ÷ 1000 × run time TiR

I

Grid carbon intensity (kgCO₂e/kWh)

PUE

Data-centre power usage effectiveness, applied to operational carbon

M

Embodied carbon = TE × (TiR ÷ EL) × (RR ÷ ToR) × Quantity

R

Number of functional units (decisions measured)

CPU%

Share of CPU actually used by the software (idle subtracted)

④ Parameters and sources

⑤ Result

The same measurement is reported under the three boundaries in the TIER worksheet. Our service runs on a cloud virtual machine, so the vCPU boundary is the primary result.

Primary result split: operational 0.003168 mg + embodied 0.000464 mg

⑥ Measurement period and method

Measured on 2026-08-21 on the host (AWS EC2 t3.xlarge, Taipei ap-east-2) with TIER’s GreenCalculatorJavaNoUI 1.0.4: 200 warm-up runs, then 20 rounds of a 60-second idle baseline followed by 1,000 loaded decisions. The worksheet was finalised after TIER’s reply on 2026-10-02.

⑦ Uncertainty and limitations

• Round-to-round coefficient of variation is about 10.2% (raw measurement file); individual values vary within this range.
• Physical power cannot be metered on a cloud VM; vCPU power uses a public reference value.
• PUE is the AWS global average and the grid factor an annual average; Taipei facility values and hourly variation are not reflected.
• Embodied carbon TE comes from the tool / public datasets, not from a supplier declaration for this host.
• Only the decision engine is covered; a full website diagnosis also calls an external LLM once, which is not included.
• Results differ by about 19× across the three boundaries; check boundary and parameters before comparing with other vendors.

⑧ Improvement actions taken

Measures already live on the platform and traceable in our version history. They do not imply any guaranteed carbon or cost reduction.

• The decision step uses deterministic rough-set + MCDM computation (CPU-level) instead of per-request LLM inference for ranking.
• Monthly AI model lifecycle check (from 2026-10) compares official OpenAI / Anthropic / Google deprecation schedules; models being retired were switched to current versions on 2026-10-02.
• Image generation has quality tiers: high quality must be explicitly chosen (and uses more points); legacy model keys default to low/medium quality.
• Caching avoids recomputation: diagnosis results for the same URL are reused for 14 days; AI-generated images are cached by prompt hash.
• Batch AI jobs have a per-run cost cap and stop when it is reached.
• Main services run in the AWS Taipei region (ap-east-2).

⑨ Disclosure evidence (DSCS)

Every number on this page is contained in the disclosure JSON below, and its SHA-256 hash is shown. Anyone can copy the JSON and recompute the hash to confirm the figures have not been changed afterwards.

{"schema":"bigo-sci-disclosure/v1","publisher":"Big-O Information Intelligence Corp. (大鷗摩飛智能有限公司)","method":"ISO/IEC 21031:2024 Software Carbon Intensity (SCI) — self-disclosure, not third-party verified","formula":"SCI = ((E x Quantity x I x PUE) + TE x (TiR/EL) x (RR/ToR) x Quantity) / R x CPU%","software":"EcoGEO Audit decision engine (Rough Set QuickReduct + MCDM, ROC patent M683041), core-algo service","functionalUnit":"1 diagnostic decision = 1 rough-set feature reduction + 1 MCDM evaluation","host":"AWS EC2 t3.xlarge (4 vCPU), ap-east-2 Taipei; CPU Intel Xeon Platinum 8259CL","measurement":{"date":"2026-08-21","tool":"GreenCalculatorJavaNoUI 1.0.4 (provided by Taiwan Institute of Economic Research)","rounds":20,"decisionsPerRound":1000,"totalDecisions":20000,"warmupDecisions":200,"idleBaselineSecondsPerRound":60,"uniqueResultHashes":20000,"loadSecondsTotal":498.7058,"roundToRoundCv":0.1021},"parameters":{"I_kgCO2e_per_kWh":0.466,"PUE":1.15,"EL_hours":35040,"TiR_hours":0.13852939,"softwareCpuShare":0.1016,"R":20000},"blocks":{"vCPU":{"W":2.1,"Quantity":4,"TE_kg":277,"RR":1,"ToR":48,"sci_mg":0.003631,"operational_mg":0.003168,"embodied_mg":0.000464},"cpu":{"W":136.67,"Quantity":1,"TE_kg":277,"RR":4,"ToR":48,"sci_mg":0.052007},"wholeMachine":{"W":182.23,"Quantity":1,"TE_kg":48,"RR":1,"ToR":1,"sci_mg":0.069688}},"primaryBlock":"vCPU","primaryResult_mgCO2e_per_decision":0.003631,"calculationVersion":"2026-10-02","excluded":["third-party LLM inference (runs in provider data centres)","front-end UI, database writes, email/notifications, PDF rendering","network transfer and end-user devices"]}

This version has not yet been anchored with OpenTimestamps; it will be anchored and listed in the DSCS Trust Ledger once finalised.

Version history

The 2026-09-29 to 09-30 versions reported 0.078 mgCO₂e per decision using the whole-machine boundary without idle subtraction. On 2026-10-02 the figure was corrected per TIER’s reply (vCPU quantity applied once; CPU% = software share with idle subtracted, 10.16%) to the values on this page.

This tab is Big-O’s self-assessment and plan for green software practices on its own platform. It is not an ISO certification, has not been verified by a third party, and is not a carbon measurement. The SCI figure Big-O publicly discloses is the one in the “SCI disclosure (POC-verified scope)” tab.

① Green software maturity matrix (self-assessment)

Self-assessed on 4 dimensions × 5 levels, with reference to O’Reilly’s “Building Green Software” and Green Software Foundation materials. Assessment version: 2026-05-24; actual progress on each dimension will be updated at the next self-assessment.

② Practices

These describe what we do; they do not imply any guarantee of carbon or cost reduction.

• ✓ In placeMain services run in the AWS Taipei region (ap-east-2), close to most users.
• ✓ In placePages and static assets are served from the Vercel edge cache; cache hits do not reach the EC2 host.
• ✓ In placeNext.js code splitting and tree shaking reduce front-end download size.
• ✓ In placeAI models are tiered by task: simple tasks use smaller models rather than always using large ones.
• ✓ In placeVideos are generated only when a user requests them, never pre-generated.
• ✓ In placeTimestamp evidence uses OpenTimestamps, sharing Bitcoin block anchoring instead of running our own chain.
• ✓ In placePM2 fork mode with automatic graceful restarts keeps server memory use in check.
• ✓ In placePostgreSQL index tuning shortens database query time.
• → PlannedPlanned: caching for large language model responses to avoid repeated inference.
• → PlannedCDN edge nodes do not yet cover Southeast Asia; a Singapore node is being evaluated.
• … ExploringExploring carbon-aware scheduling: moving scheduled jobs to hours with lower grid carbon intensity.
• … ExploringEvaluating lower-carbon region options from cloud and CDN providers.

③ Roadmap

These are planned directions, not commitments; timing and scope may change.